A Robust Coinversion Model for Soil Moisture Retrieval From ...
5230
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014
A Robust Coinversion Model for Soil Moisture Retrieval From Multisensor Data Xianfeng Zhang, Jiepeng Zhao, and Jie Tian
Abstract—Optical remote sensing has been widely used to estimate soil moisture. However, modeling soil moisture dynamics across a large area based on remotely sensed optical data still poses a problem because of its spatial discontinuity due to cloud contamination. This study proposes a multisensor strategy for better mapping surface soil moisture on a daily basis at a regional scale. The basic idea is to decompose the surface soil moisture at any location into two terms, namely, baseline value in an observed period and daily variation, and to estimate for each term differently. For a certain day of interest, the corresponding 16-day composite of Moderate Resolution Imaging Spectroradiometer (MODIS) data is used to estimate the soil moisture baseline values across space, and the Advanced Microwave Scanning Radiometer for EOS (AMSR-E) data are employed to estimate the daily variations. The proposed model was applied to produce daily surface soil moisture maps at a 1-km resolution for the fairly large study area of Xinjiang, China, regardless of the local weather conditions. It was found that the integrated use of MODIS and AMSR-E data was able to achieve significantly higher accuracy in surface soil moisture estimation (with a root-mean-square error of 3.99% in May and 4.43% in August, 2009) than the approaches based on either data alone could. The proposed model is expected to perform well for mapping surface soil moisture in other arid areas after the required parameters are calibrated with the local field data. Index Terms—Advanced Microwave Scanning Radiometer for EOS (AMSR-E), coinversion, Moderate Resolution Imaging Spectroradiometer (MODIS), multisensor, soil moisture.
I. I NTRODUCTION
S
OIL moisture is a critical environmental variable for the global water and energy budgets that have a great impact on the climate change over land [1]–[3]; it is fundamental to vegetation growth and serves as an important parameter in monitoring agricultural drought and predicting crop yield [4]–[6]. Remote sensing has lately been used to map the spatial and temporal distribution of soil moisture at a large scale [3]. To the present, three types of remote-sensing-based approaches
Manuscript received March 14, 2012; revised October 6, 2012, March 6, 2013, July 4, 2013, and August 27, 2013; accepted September 30, 2013. This work was supported in part by the Chinese Ministry of Science and Technology under Grant 2012BAH27B03 and in part by the Natural Scientific Foundation of China under Grant 41071257. X. Zhang is with the Institute of Remote Sensing and Geographic Information System, Peking University, Beijing 100871, China (e-mail: xfzhang@pku. edu.cn; [email protected]). J. Zhao is with Beijing Institute of Surveying and Mapping, Beijing 100038, China. J. Tian is with the Department of International Development, Community, and Environment, Clark University, Worcester, MA 01610 USA. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2013.2287513
have been explored, which rely on thermal inertia (TI), spectral vegetation index, and microwave derivatives, respectively [7]–[11]. In comparison, the microwave remote sensing of soil moisture offers several advantages over the others, including the following: 1) a stronger ability to penetrate through clouds; 2) being directly related to soil water content by the soil dielectric constant; and 3) a lower sensitivity to land surface roughness or vegetation coverage [12], [13]. However, although passive microwave remote sensors are able to provide data with a much better spatial coverage, their spatial resolution is often too coarse for the retrieval of highly detailed soil moisture information. Visible and near-infrared (VNIR) and thermal infrared (IR) remote sensing data have been also applied to estimate soil moisture since the 1980s and are still widely employed in monitoring soil moisture due to their better availability and higher resolution in characterizing the reflective and emissive properties of soil and vegetation [8], [14], [15]. TI and temperature vegetation dryness index (TVDI) are two important derivatives from remotely sensed VNIR- and IR-band data, which are commonly used in developing soil moisture inversion models [2], [7], [8], [16], [17]. However, the TI-based methods are not applicable in densely vegetated areas and can hardly provide an accurate estimation of soil moisture in vegetated lands [10], [15], [18], [19]. The TVDI-based approaches were therefore proposed to cope with soil moisture estimation in vegetated areas [8], [20], [21]. Unlike remotely sensed microwave data, thermal IR and VNIR data are often interfered by clouds even in arid areas, and consequently, the TI- or TVDI-based inversion models (or the like) are not suitable for the continuous mapping of soil moisture at a large scale. The synthetic use of multiple remote sensors for mapping soil moisture regionally and globally has become a research hot spot in the community of quantitative remote sensing. This paper proposes a multisensor strategy that integrates Moderate Resolution Imaging Spectroradiometer (MODIS) and Advanced Microwave Scanning Radiometer for EOS (AMSR-E) data to build an effective inversion model for the accurate estimation of surface soil moisture (SSM) information across a large arid area. The rest of this paper is structured as follows. Section II describes the multisensor strategy and the integrated inversion model that we developed. Section III presents a case study of applying the model to mapping soil moisture in Xinjiang, a typical arid area in Northwest China. In Section IV, the modeled results are compared with those from a TVDI model and the U.S. National Snow and Ice Data Center (NSIDC) soil moisture products. Section V summarizes the findings and draws some conclusions.
0196-2892 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHANG et al.: ROBUST COINVERSION MODEL FOR SOIL MOISTURE RETRIEVAL FROM MULTISENSOR DATA
II. M ULTISENSOR R ETRIEVAL M ETHOD A. Multisensor Strategy The spectral data collected by different remote sensors can be synthetically utilized in the inversion of SSM. Generally speaking, three strategies are adopted: 1) independently processing each data set involved to obtain the derivatives required for the final inversion process [19], [22]; 2) building an integrated model to inverse soil moisture from multisensor data [23]; and 3) postprocessing (e.g., intersensor validation) with the inversion results from individual sensors to achieve higher accuracy [24], [25]. The approaches adopting the strategies 1) and 3) are relatively easier to implement but have little room for improvement because they do not focus on the core inversion process. In contrast, the integrated inversion models can handle multisensor data and are better suitable for mapping soil moisture across various regions. In particular, when a rapidly increasing number of remote sensors are launched and orbiting about the Earth, the theoretical and practical development for the integration of soil moisture inversion models is of great methodological and economic value [19], [23]. The variable of soil moisture can be assumed to be an addition of two terms during an observation period: baseline value and daily variation. Specifically, the baseline represents the relatively long-term level of soil water content that is determined by the regional climate and the natural environment, whereas the daily variation represents the short-term fluctuations caused by precipitation, evaporation, and/or other random meteorological factors. This concept is depicted as smij (t) = mij + Δmij (t)
(1)
where mij denotes the baseline soil moisture at a pixel (i, j) in an observation period T ; Δmij stands for the daily variation of soil moisture for the pixel (i, j) at time t in T ; and smij is the actual soil moisture (in volumetric fraction) at the pixel (i, j) as of time t. If T is relatively short (e.g., ten days), the local vegetation cover and land surface roughness normally do not significantly change for same area. In this study, MODIS and AMSR-E (both on board the National Aeronautics and Space Administration Aqua spacecraft) data are used to estimate the baseline value and the daily variation of soil moisture, respectively. The integrated analysis of data from these two sensors is expected to improve the soil moisture mapping. B. Inversion of SSM Baseline Levels In remote sensing of natural environments, land surfaces can be usually characterized by their normalized difference vegetation index (NDVI) and be broadly classified into three types: bare land, sparsely vegetated, and densely vegetated. For each type, a tailored inversion model is chosen from the following three to achieve a more accurate estimation of soil moisture baseline level. In Bare Lands: TI-Based Inversion: According to Stisen et al. [20], the TVDI-based models are not applicable if the NDVI value is below 0.1 because the theoretic basis for such models is the vegetation evaporation process. Comparatively, the TI-based model proposed by Price [7] is more suitable for
5231
estimating the baseline soil moisture levels in bare lands. In our approach, apparent TI (ATI) is derived from the MODIS thermal and visible band data and statistically correlated with SSM. Since clouds often partially contaminate the optical and thermal band data remotely sensed over a large area, the MODIS 16-day composite products are chosen to represent the local vegetation condition for their significantly better spatial continuity. ATI can be calculated using [7] ATIi,j =
1 − αi,j ΔTi,j
(2)
where αi,j represents land surface albedo, which can be derived from the MODIS VNIR data; ΔTi,j stands for the surface temperature range over 2 A . M . to 2 P. M . of local time, and i and j denote the coordinates of an image pixel. In Densely Vegetated Areas: TVDI-Based Inversion: The TVDI-based methods estimate SSM through the extraction of water stress index from the T s–NDVI feature space [8]. For a given region, the more the vegetation is covered, the greater the ability for the vegetation to turn the solar radiation energy into latent heat through evapotranspiration. The T s therefore tends to be lower due to less solar energy transformed into sensible heat [8], [26]. The amount of solar radiation absorbed by a land surface and its soil moisture are the two controlling factors of the T s. Given that the direct solar radiation is constant, a surface tends to be warmer if the soil is drier because the weakened evapotranspiration of vegetation will result in less latent heat but more sensible heat and vice versa. The drought in soil layers often stresses the above-growing vegetation by the undersupply of water and causes the foliage temperature to increase. Soil moisture can therefore be quantitatively estimated as a function of canopy temperature and NDVI [8]. The TVDI-based methods, also known as the “triangle” methods, were developed based on the analysis in the T s–NDVI space and have been successfully applied in the estimation of both evapotranspiration [20] and soil moisture [8]. TVDI can be calculated from MODIS data using TVDI = =
T s − T smin T smax − T smin T s − (a2 + b2 NDVI) a1 + b1 NDVI − (a2 + b2 NDVI)
(3)
where T s is derived from the MODIS thermal data. a1 , b1 , a2 , and b2 are the empirical parameters that need to be determined for modeling the wet and dry edges in the T s–NDVI feature space. T smax and T smin are the dry edge and the wet edge values in the T s–NDVI feature space, respectively. NDVI is the mean calculated from the time series of index values available during the period T and is therefore used to represent the averaged condition of local vegetation over the time span. It is well known that the TVDI model [8] has constraints in the soil moisture inversion from remotely sensed data over a large area. First, clouds can hinder the estimation of T s and NDVI, which are required in the TVDI calculation. Second, the variation of T s is a result from not only the vegetation evapotranspiration but also the solar radiation and the atmospheric condition, which adds uncertainty to the TVDI-based
5232
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014
soil moisture inversion. Third, Moran et al. [9] attempted to calculate the theoretical coordinates for the four vertices of the trapezoid in the T s–NDVI feature space and found that T smin was actually not constant but changed with vegetation cover fraction. Thus, the dry and wet edges have to be determined for each particular study. In order to model the edges, the NDVI values can be grouped into equal bins with a width of, for example, 0.01. The maximum T s and the minimum T s within each bin are linearly regressed on their corresponding NDVI, so that the coefficients a1 , a2 (intercepts) and b1 , b2 (slopes) can be determined, respectively. Thus, the TVDI for each pixel can be calculated using (3), and SSM can be finally regressed on TVDI to establish the statistical relationship between the two. It should be noted that the TVDI value ranges from 0 to 1. The MODIS cloud free products are chosen for the spatially continuous estimation of T s required by the TVDI model. More specifically, the MODIS 16-day maximum temperature composite data are mapped to the T s–NDVI space. It is understandable that the direct solar radiation reaching the surface varies with elevation and geographic latitude, which violates the assumption of homogeneous solar radiation in the TVDI model. Therefore, before being used to model and calculate TVDI, the temperature difference due to the geographic variation of solar radiation received at the surface is corrected through a regression analysis that is similar to the work by Stisen et al. [20]. Please refer to Zhao et al. for details [27]. In Sparsely Vegetated Areas: Averaging the Two: Due to the fact that the vegetation evaporative cooling process is very weak in sparsely vegetated areas, either an ATI-based or a TVDIbased model alone tends to produce inaccurate SSM inversion results. However, averaging the inversion results from the two models seems to be a better solution for estimating the baseline level of SSM in those transitional zones from “bare lands” to “densely vegetated.” C. Estimation of Daily Variation Daily variation of soil moisture over a large area can be derived from passive microwave remote sensing. For example, the spaceborne passive microwave sensor AMSR-E collected brightness temperature data through the four frequency channels of 6.9, 10.7, 18.7, and 36.5 GHz and is able to cover the entire globe every two days, with a Sun-synchronous orbit (ascending and descending at local time 1:30 P. M . and 1:30 A . M ., respectively) [12]. Due to the fact that the data at 36.5 and 18.7 GHz are heavily influenced by clouds, whereas the C-band (6.9 GHz) data are vulnerable to radio frequency interference, the X-band (10.7 GHz) data are left to be the most suitable for soil moisture inversion [13], [28]. The SSM inversion models using AMSR-E data are mostly based on analyzing the derivatives (e.g., polarization ratios) from the AMSR-E brightness temperature data. According to Njoku and Chan [29] and Zhang et al. [13], the daily variation of SSM at a location can be estimated using 2 Δmv = k1 (Pr − Prmin )Prkmin
(4)
where Δmv is the daily SSM variation. Pr stands for a polarization ratio (TBv − TBh )/(TBv + TBh ), in which TBv and
Fig. 1.
Study area of Xinjiang and the field survey area.
TBh are the vertically and horizontally polarized AMSR-E brightness temperatures, respectively. Prmin is the minimum polarization ratio value for a pixel in a certain observation period. k1 and k2 are the best fit coefficients. The change Δmv is mainly driven by the local land surface evaporation and precipitation and represents the daily fluctuation of SSM. D. Integrated SSM Inversion Model The integrated SSM inversion model follows three major steps as follows: Step 1) To estimate the baseline values of SSM from a MODIS 16-day composite image in the ways described in Section II-B; Step 2) To estimate the daily variation values of SSM in the 16-day period using the AMSR-E data at 10.7 GHz and (4); Step 3) To combine the baseline and the daily variation values of SSM obtained from Steps 1 and 2 to produce large-scale daily soil moisture maps at a 1-km resolution. III. A PPLYING THE M ODEL : A C ASE S TUDY A. Study Area Our study area is located in Northwest China, including the entire Xinjiang Uygur Autonomous Region (see Fig. 1). Compared with East China, Xinjiang is a large and sparsely populated region that occupies approximately one sixth of the country’s territory. It borders Tibet to the south; the Qinghai and Gansu provinces to the southeast; Mongolia to the east, Russia to the north; and Kazakhstan, Kyrgyzstan, Tajikistan, Afghanistan, and the Pakistan- and India-controlled parts of Kashmir to the west. The topography of Xinjiang can be described as “three mountains delineate two basins”: The Altai, Tianshan, and Kunlun Mountains lie on the north border, through the central area, and on the south border of Xinjiang, respectively. The Zhungeer Basin is bordered by the Tianshan and Altai Mountains, and the Tarim Basin is situated between the Kunlun and Tianshan Mountains.
ZHANG et al.: ROBUST COINVERSION MODEL FOR SOIL MOISTURE RETRIEVAL FROM MULTISENSOR DATA
5233
TABLE I PARAMETERS U SED IN THE I NVERSION OF S OIL M OISTURE IN X INJIANG
The study area has a typical central Asian arid climate with an average annual precipitation of 150 mm. The spatial distribution of the precipitation is quite heterogeneous. Typically, a sequence of landscapes (snow/ice covers → high-mountain meadows → forest → well-grown dense grassland neighboring the forest → very sparse hilly grassland → natural–artificial oasis → desert) can be seen for a mountain in Xinjiang when traversing from its top, through the hills, and to one of the central basins. The water resources in Xinjiang are extremely important for the regional ecoenvironmental and agricultural management and largely shape the landscapes’ evolution. Thus, it is of great value to closely and continuously monitor the soil moisture condition in the vast area of Xinjiang through satellite remote sensing. B. Data Acquisition and Preprocessing The MODIS 16-day synthesized NDVI product (MOD13A2) and the MODIS daily land surface temperature (LST) product (MYD11A1) for May and August 2009 were downloaded from the USGS website. The MODIS Reprojection Tool software provided by the USGS was initially used to preprocess the MODIS data. Clouds were subsequently masked out from the preprocessed data, and the MODIS LST data were further corrected for topographic elevation and geographic latitude [27]. A time series of the corrected LST data was then used to produce the 16-day composites of maximum temperature. The Interactive Data Language (IDL) programming language was used to implement the modules/tools that allowed us to perform the specific processing and calculations as needed in the study. The resampled brightness temperature (TB ) data sets (AMSR-E_L3_DailyLand_V06) for the same periods were obtained and used to estimate the daily variations of SSM across Xinjiang. The NSIDC’s soil moisture products were also downloaded for comparison purposes, as presented in Section IV. Additionally, for model evaluation and validation, field data were collected in an area that is representative of the Xinjiang’s landscapes: the northern slopes of the Tianshan Mountain. The field soil moisture data were taken by two means: mobile measurements using the WET instrument developed by Delta-T Devices and fixed observations using the WatchDog2400 Irrigation Station developed by Spectrum Technologies, Inc. Specifically, one WatchDog2400 station was set up within each of the five different land covers in the area: forest, meadow, sparse grassland, cotton farmland, and desert. These stations continuously collected the 5- to 10-cm SSM data on an hourly basis from May 7 through October 11, 2009. Moreover, the
Fig. 2. Inversed soil moisture distribution across Xinjiang through the integrated use of AMSR-E and MODIS data. (a) and (b) Daily variations derived from the ascent and descent overpasses of AMSR-E. (c) Baseline soil moisture derived from the MODIS data. (d) Integratively reversed soil moisture.
conventional yet reliable loss-on-drying method was employed to measure the real soil moisture content in the field samples and to calibrate the measurements by the two instruments. C. Implementation of the Integrated Model and the Results The whole SSM inversion model was implemented in the IDL programming environment. The procedures involved in the model are described in details as follows. Estimation of Baseline Soil Moisture: Using the methods described in Section II-B, the procedure for baseline SSM estimation is illustrated in a conceptual coding format as shown below. /∗ for bare land If NDVI < 0.1 Bsm1 = c1 + d1 ∗ ATI /∗ for densely vegetated Else If NDVI > 0.18 Bsm2 = c2 + d2 ∗ TVDI /∗ for sparsely vegetated (neither densely vegetated nor bare land) Else Bsm3 = (Bsm1 + Bsm2 )/2 where c1, c2, d1, and d2 are the coefficients obtained from the regression analysis of the in situ soil moisture measurements on the derived indices (ATI and TVDI) from the MODIS data.
5234
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014
Fig. 3. Soil moisture variation in Xinjiang in May and August 2009 (the maps were produced from the AMSR-E and MODIS data using the integrated inversion model).
The threshold values of NDVI (e.g., 0.1 or 0.18) for differentiating land cover types were determined by finding the abrupt changing points in the correlation coefficients between the in situ measured SSM values and the ATI or TVDI values. As can be seen, the estimation procedure applies the most suitable model for each pixel based upon its NDVI condition. Estimation of Daily Soil Moisture: The procedure for daily SSM variation estimation is illustrated below in the conceptual coding format as well. If Prmin ≤ Pr ≤ 3 Prmin Δmv = 72.58(Pr − Prmin ) Pr−0.625 min If Pr > 3 Prmin let Pr = 3 Prmin Δmv = 145.16 Pr0.375 min Note: please refer to (4) for the meanings of the denotations. The daily variation of SSM is mainly driven by the local precipitation and evaporation. The upper condition in the code is for the usually rapid increase of SSM in response to a precipitation process, whereas the lower condition is for the exceptions when the surface soil layers are saturated with water and the soil moisture will therefore stop increasing regardless of further water input. It should be pointed out that such exceptional scenarios are not well dealt with in the empirical models used by NSIDC to produce their soil moisture products. More detailed discussion about this can be found in Zhang et al. [13]. The model parameters derived from the in-situ measurements in Xinjiang are listed in Table I. Fig. 2 displays the inversion results of baseline SSM and daily SSM variation in Xinjiang on May 25, 2009. As shown in Fig. 2(a) and (b), the distribution of Δmv well matches the precipitation condition over northern Xinjiang on May 25, 2009. The precipitation data collected by the China Meteorological Administration operated weather stations (see Fig. 1) showed that there was strong precipitation taking place in northern Xinjiang on that day. It can be recognized that the inversed soil moisture baseline values [see Fig. 2(c)] form zones that spatially correspond to the different land covers in Xinjiang. Fig. 2(d) shows the final SSM inversion map of May 25, 2009 by combining the baseline and the daily variation. Similar maps
Fig. 4. Comparison of the in situ measurements, the soil moisture data sets produced by the integrated inversion model and the TVDI-based model, and the NSIDC soil moisture products for the different land covers in May 2009.
were also produced for May 5, 20, and 30, 2009 and August 5, 20, and 30, 2009, as displayed in Fig. 3. It can be interpreted that in both May and August 2009, the soil moisture in the Taklimakan Desert was relatively low overall, whereas both sides of the Tianshan and Kunlun Mountain ridges had a much higher level of soil moisture ranging from 10% to 40%. The soil moisture in the other areas of Xinjiang was between 5% and 15%. The soil moisture spatial distributions are consistent with our knowledge about the region and the field investigation. IV. ACCURACY A SSESSMENT AND D ISCUSSION A. Accuracy Assessment As described before, the field soil moisture data consist of the measurements taken by the WET and WathcDog2400 instruments in the study area. The WET data have a fairly good
ZHANG et al.: ROBUST COINVERSION MODEL FOR SOIL MOISTURE RETRIEVAL FROM MULTISENSOR DATA
5235
Fig. 5. Correlation analysis of the in situ measurements with the (a) TVDI-based inversion results, the (b) NSIDC soil moisture products, and the (c) integrated inversion results in May 2009.
Fig. 6. Correlation analysis of the in situ measurements with the (a) TVDI-based inversion results, the (b) NSIDC soil moisture products, and the (c) integrated inversion results in August 2009.
spatial coverage, whereas the WatchDog2400 data have a good temporal frequency of every hour. The WET data were first calibrated using the WatchDog2400 data with the aid of the data collected by the traditional loss-on-drying method and were then used as the reference in the assessment and comparison of the soil moisture products by the different approaches. The performance of the proposed inversion model was evaluated by comparing its results to those from the original TVDIbased inversion model, the NSIDC soil moisture products, and the in situ measurements (see Fig. 4). The assessment was also broken down by land cover type, comparing the model’s performance in forestland, dense grassland, and sparse grassland. Cotton fields and desert areas were excluded because the soil moisture in the former is dominated by human factors such as irrigation, and neither WET nor WatchDog2400 could function well to take reliable readings in the loose sand of desert. It is found that the inversion results of SSM with the proposed model are highly correlated with the in situ measurements in all the three types of land covers in both May (r2 = 0.85) and August (r2 = 0.76) 2009 (see Figs. 5 and 6). The root-mean-square errors (RMSEs) are 3.99% and 4.43% in the two months, respectively. In comparison, the correlation between the in situ measurements and the inversion results using the TVDI-based model (r2 = 0.36 in May and 0.55 in August) or the NSIDC soil moisture products (r2 = 0.33 in May and 0.65 in August) is much lower (see Figs. 5 and 6). The corresponding RMSEs are also significantly larger (TVDI-based model: 14.86% in May, 8.62% in August; NSIDC products: 8.96% in May, 8.95% in August), suggesting that the proposed model has stronger estimation power. It should be pointed out that the soil moisture maps produced using the original TVDIbased model are spatially discontinuous in all the three types of land covers primarily due to cloud contamination in the data (even for arid areas). This is a general limitation for estimating soil moisture from remotely sensed optical or thermal infrared
data. The temporal variation of SSM in the sparse grasslands is well captured in our inversion results and the NSIDC soil moisture products, which confirms the effectiveness of AMSR-E data in mapping soil moisture dynamics in sparsely vegetated areas. B. Discussion The assessment of the SSM inversion results from the AMSR-E and MODIS data is not very direct due to the spatial resolution (1 km) of the sensors. The problem is attributed to the uncertainty in characterizing the soil moisture of a 1 × 1 km2 cell, in which one point measurement or the average of a few scattered point measurements is usually used to represent the soil moisture of the whole cell. Such representation can sometimes be unreliable and uncertain to some degree. Moreover, the inversed soil moisture information from the AMSR-E X-band measurement only characterizes approximately the top 2-cm layer of the soil [12], [30]. The moisture in such a surficial soil layer changes too frequently with rainfall, wind, and temperature to be accurately measured in the field. The moisture in the top soil layer of 5–10 cm was alternatively measured in the field and used for the model assessment; the inversion results, consequently, bear some inherent errors when assessed using the field data as the reference [31]. It is understandable that vegetation cover tends to interfere the estimation of the daily variation of SSM with the passive microwave AMSR-E data. However, as shown in Fig. 4, the proposed inversion model performed quite well even in the densely vegetated areas such as forestland and highdensity grassland. This is largely attributed to our model’s ability to complement the strength of the sensors in SSM reversion while minimizing their limitations. Furthermore, the integrated inversion model proposed also benefits from the improvement made to Sandholt’s TVDI model. Three main
5236
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014
adjustments were made: 1) masking out of clouds and synthesizing T s from the 16-day observations in order to minimize cloud contamination; 2) correcting for elevation and latitude to standardize the solar radiation over a relatively large area; and 3) determining the wet and dry edges in the T s–NDVI space from the actual data rather than using the theoretical ones. These adjustments also greatly contributed to the superior performance of our proposed model than the original TVDI-based model in estimating the baseline SSM from the MODIS data. The heterogeneity of soil moisture distribution in arid areas such as the study area is relatively smaller than that in humid areas; the spatial variation of soil moisture within each AMSR-E image pixel, therefore, can be more effectively reflected by the baseline values extracted from the 1-km MODIS data. However, a hyperlocal weather event (e.g., thunderstorms affecting 3–5 km2 ) may dramatically elevate the soil moisture level in an area that is much smaller than the AMSR-E cell size (25 × 25 km). Such an elevation is therefore difficult to detect from the AMSR-E data. The future generation of passive microwave sensors with stronger spatial resolving power is expected to overcome the aforementioned limitation of our proposed model. V. C ONCLUDING R EMARKS This paper has proposed an integrated model for SSM inversion from multisensor data over a large geographic area. The basic idea underlying the model is to decompose the SSM value at an image pixel into two terms (baseline value and daily variation) and to model them independently. The baseline SSM in a relatively short period can be estimated from the MODIS optical and thermal infrared data, and the daily variation of SSM can be estimated using the AMSR-E X-band microwave data. The proposed inversion model was applied to estimate and map SSM in Xinjiang, China, at a 1-km resolution. The accuracy assessment has shown that, with a stronger correlation (r2 = 0.86 for May, r2 = 0.76 for August) with the reference data and a smaller RSME (3.99% for May and 4.43% for August), our model outperformed the original TVDI-based model that replies on MODIS data only; it also produced more accurate results than the NSIDC products that were solely derived from AMSR-E data. In addition, our model is generally applicable in areas that are densely vegetated (e.g., forest, grassland, and cropland), sparsely vegetated, or bare (e.g., desert). It therefore has a greater ability to produce spatially continuous SSM maps for a large region on a more frequent basis. The proposed model is highly promising for mapping SSM in other similar natural environments. The parameters required by the model should be calibrated using the local field data to adapt to a particular area. Caution needs to be exercised, particularly when estimating the surface soil moisture in agricultural areas due to their pronounced artificial processes such as irrigation and cultivation. Future research efforts should be made to more explicitly define the physical meanings of the baseline and daily variation of SSM.
ACKNOWLEDGMENT The authors would like to thank the anonymous reviewers for their valuable comments and suggestions. R EFERENCES [1] M. C. Anderson, J. M. Norman, G. R. Diak, W. P. Kustas, and J. R. Mecikalski, “A two-source time-integrated model for estimating surface fluxes using thermal infrared remote sensing,” Remote Sens. Environ., vol. 60, no. 2, pp. 195–216, May 1997. [2] A. Verhoef, “Remote estimation of thermal inertia and soil heat flux for bare soil,” Agr. Forest Meteorol., vol. 123, no. 3/4, pp. 221–236, Jun. 2004. [3] X. Zhan, P. R. Houser, J. P. Walker, and W. T. Crow, “A method for retrieving high-resolution surface soil moisture from hydros L-band radiometer and radar observations,” IEEE Trans. Geosci. Remote Sens., vol. 44, no. 6, pp. 1534–1544, Jun. 2006. [4] S. N. Goward, Y. Xue, and K. P. Czajkowski, “Evaluating land surface moisture conditions from the remotely sensed temperature/vegetation index measurements: An exploration with the simplified simple biosphere model,” Remote Sens. Environ., vol. 79, no. 2/3, pp. 225–242, Feb. 2002. [5] P. C. Doraiswamy, J. L. Hatfield, T. J. Jackson, B. Akhmedov, J. Prueger, and A. Stern, “Crop condition and yield simulations using Landsat and MODIS,” Remote Sens. Environ., vol. 92, no. 4, pp. 548–559, Sep. 2004. [6] B. Narasimhan and R. Srinivasan, “Development and evaluation of Soil Moisture Deficit Index (SMDI) and Evapotranspiration Deficit Index (ETDI) for agricultural drought monitoring,” Agr. Forest Meteorol., vol. 133, no. 1–4, pp. 69–88, Nov. 2005. [7] J. Price, “On the analysis of thermal infrared imagery: The limited utility of apparent thermal inertia,” Remote Sens. Environ., vol. 18, no. 1, pp. 59–73, Aug. 1985. [8] I. Sandholt, K. Rasmussen, and J. Andersen, “A simple interpretation of the surface temperature/vegetation index space for assessment of surface moisture status,” Remote Sens. Environ., vol. 79, no. 2/3, pp. 213–224, Feb. 2002. [9] M. S. Moran, T. R. Clarke, and Y. Inoue, “Estimating crop water deficit using the relation between surface air temperature and spectral vegetation index,” Remote Sens. Environ., vol. 49, no. 3, pp. 246–263, Sep. 2004. [10] W. W. Verstraeten, F. Veroustraete, C. J. van der Sande, I. Grootaers, and J. Feyen, “Soil moisture retrieval using thermal inertia, determined with visible and thermal spaceborne data, validated for European forests,” Remote Sens. Environ., vol. 101, no. 3, pp. 299–314, Apr. 2006. [11] R. Panciera, J. P. Walker, J. D. Kalma, D. Jetse, E. J. Kim, K. Saleh, and J. P. Wigneron, “Evaluation of the SMOS L-MEB passive microwave soil moisture retrieval algorithm,” Remote Sens. Environ., vol. 113, no. 2, pp. 435–444, Feb. 2009. [12] E. G. Njoku, T. Jackson, V. Lakshmi, T. Chan, and S. Nghiem, “Soil moisture retrieval from AMSR-E,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 2, pp. 215–229, Feb. 2003. [13] X. Zhang, J. Zhao, Q. Sun, X. Wang, Y. Guo, and J. Li, “Soil moisture retrieval from AMSR-E data in Xinjiang (China): Models and validation,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 4, no. 1, pp. 117–127, Mar. 2011. [14] L. Jiang and S. Islam, “Estimation of surface evaporation map over southern great plains using remote sensing data,” Water Resour. Res., vol. 37, no. 2, pp. 329–340, Feb. 2001. [15] S. Lu, Z. Ju, T. Ren, and R. Horton, “A general approach to estimate soil water content from thermal inertia,” Agr. Forest Meteorol., vol. 149, no. 10, pp. 1693–1698, Oct. 2009. [16] R. Kimura, “Estimation of moisture availability over the Liudaogou river basin of the Loess Plateau using new indices with surface temperature,” J. Arid Environ., vol. 70, no. 2, pp. 237–252, Jul. 2007. [17] G. Hulley, S. Hook, and A. Baldridge, “Investigating the effects of soil moisture on thermal infrared land surface temperature and emissivity using satellite retrievals and laboratory measurements inertia approach,” Remote Sens. Environ., vol. 114, no. 7, pp. 1480–1493, Jul. 2010. [18] M. Minacapilli, M. Iovino, and F. Blanda, “High resolution remote estimation of soil surface water content by a thermal inertia approach,” J. Hydrol., vol. 379, no. 3/4, pp. 229–238, Dec. 2009. [19] C. Mattar, J. Wigneron, J. Sobrino, N. Novello, J. Calvet, C. Albergel, P. Richaume, A. Mialon, D. Guyon, J. Jiménez-Muñoz, and Y. Kerr, “A combined optical–microwave method to retrieve soil moisture over
ZHANG et al.: ROBUST COINVERSION MODEL FOR SOIL MOISTURE RETRIEVAL FROM MULTISENSOR DATA
[20]
[21]
[22]
[23]
[24]
[25] [26] [27] [28] [29] [30]
[31]
vegetated areas,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 5, pp. 1404–1413, May 2012. S. Stisen, I. Sandholt, A. Nrgaard, R. Fensholt, and K. H. Jensen, “Combining the triangle method with thermal inertia to estimate regional evapotranspiration—Applied to MSG-SEVIRI data in the Senegal River basin,” Remote Sens. Environ., vol. 112, no. 3, pp. 1242–1255, Mar. 2008. Y. Han, Y. Wang, and Y. Zhao, “Estimating soil moisture conditions of the greater Changbai Mountains by land surface temperature and NDVI,” IEEE Trans. Geosci. Remote Sens., vol. 48, no. 6, pp. 2509–2515, Jun. 2010. S. Cros, A. Chanzy, M. Weiss, T. Pellarin, J. Calvet, and J. Wigneron, “Synergy of SMOS microwave radiometer and optical sensors to retrieve soil moisture at global scale,” IEEE Trans. Geosci. Remote Sens., vol. 46, no. 3, pp. 835–845, Mar. 2008. N. Das, D. Entekhabi, and E. Njoku, “An algorithm for merging SMAP radiometer and radar data for high-resolution soil-moisture retrieval,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 5, pp. 1504–1512, May 2011. M. Piles, A. Camps, M. Vall-llossera, I. Corbella, R. Panciera, C. Rüdiger, Y. Kerr, and J. Walker, “Downscaling SMOS-derived soil moisture using MODIS visible/infrared data,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 9, pp. 3156–3166, Sep. 2011. J. Kim and T. Hogue, “Improving spatial soil moisture representation through integration of AMSR-E and MODIS products,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 2, pp. 446–460, Feb. 2012. L. Unganai and F. Kogan, “Drought monitoring and corn yield estimation in southern Africa from AVHRR data,” Remote Sens. Environ., vol. 63, no. 3, pp. 219–232, Mar. 1998. J. Zhao, X. Zhang, and H. Bao, “Soil moisture retrieval from remote sensing data in arid areas using a multiple models strategy,” Adv. Intell. Soft Comput., vol. 105, no. 2, pp. 635–644, 2011. E. G. Njoku, P. Ashcroft, and L. Li, “Statistics and global survey of radio-frequency interference in AMSR-E land observations,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 5, pp. 938–947, May 2005. E. G. Njoku and T. K. Chan, “Vegetation and surface roughness effects on AMSR-E land observations,” Remote Sens. Environ., vol. 100, no. 2, pp. 190–199, Jan. 2006. C. S. Draper, J. P. Walker, P. J. Steinle, R. A. M. de Jeu, and R. H. T. Holmes, “An evaluation of AMSR-E derived soil moisture over Australia,” Remote Sens. Environ., vol. 113, no. 4, pp. 703–710, Apr. 2009. M. H. Cosh, T. J. Jackson, and R. Bindlish, “Watershed scale temporal and spatial stability of soil moisture and its role in validating satellite estimates,” Remote Sens. Environ., vol. 92, no. 4, pp. 427–435, Sep. 2004.
5237
Xianfeng Zhang received the Ph.D. degree in geography from the University of Western Ontario, London, ON, Canada, in 2005. He is currently an Associate Professor with the Institute of Remote Sensing and Geographical Information Systems, Peking University, Beijing, China. His research interests are in remote sensing of ecology, hyperspectral data processing and application, and geospatial data visualization. Prof. Zhang is a Referee for several international academic journals.
Jiepeng Zhao received the Master’s degree in photogrammetry and remote sensing with Peking University, Beijing, China. He is currently a Geomatics Engineer with the Beijing Institute of Surveying and Mapping, Beijing, China. His main research interests include quantitative remote sensing and remotely sensed data processing.
Jie Tian received the Master’s degree in remote sensing from the University of Western Ontario, London, ON, Canada, in 2004 and the Ph.D. degree in geographic information system (GIS) from Queen’s University, Kingston, ON, in 2009. He is an Assistant Professor of GIS with the Department of International Development, Community, and Environment, Clark University, Worcester, MA, USA. His research interests broadly focus on geospatial information science and technology and their application in environmental monitoring and modeling, landscape ecology, public health, air quality, and renewable energy.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014
A Robust Coinversion Model for Soil Moisture Retrieval From Multisensor Data Xianfeng Zhang, Jiepeng Zhao, and Jie Tian
Abstract—Optical remote sensing has been widely used to estimate soil moisture. However, modeling soil moisture dynamics across a large area based on remotely sensed optical data still poses a problem because of its spatial discontinuity due to cloud contamination. This study proposes a multisensor strategy for better mapping surface soil moisture on a daily basis at a regional scale. The basic idea is to decompose the surface soil moisture at any location into two terms, namely, baseline value in an observed period and daily variation, and to estimate for each term differently. For a certain day of interest, the corresponding 16-day composite of Moderate Resolution Imaging Spectroradiometer (MODIS) data is used to estimate the soil moisture baseline values across space, and the Advanced Microwave Scanning Radiometer for EOS (AMSR-E) data are employed to estimate the daily variations. The proposed model was applied to produce daily surface soil moisture maps at a 1-km resolution for the fairly large study area of Xinjiang, China, regardless of the local weather conditions. It was found that the integrated use of MODIS and AMSR-E data was able to achieve significantly higher accuracy in surface soil moisture estimation (with a root-mean-square error of 3.99% in May and 4.43% in August, 2009) than the approaches based on either data alone could. The proposed model is expected to perform well for mapping surface soil moisture in other arid areas after the required parameters are calibrated with the local field data. Index Terms—Advanced Microwave Scanning Radiometer for EOS (AMSR-E), coinversion, Moderate Resolution Imaging Spectroradiometer (MODIS), multisensor, soil moisture.
I. I NTRODUCTION
S
OIL moisture is a critical environmental variable for the global water and energy budgets that have a great impact on the climate change over land [1]–[3]; it is fundamental to vegetation growth and serves as an important parameter in monitoring agricultural drought and predicting crop yield [4]–[6]. Remote sensing has lately been used to map the spatial and temporal distribution of soil moisture at a large scale [3]. To the present, three types of remote-sensing-based approaches
Manuscript received March 14, 2012; revised October 6, 2012, March 6, 2013, July 4, 2013, and August 27, 2013; accepted September 30, 2013. This work was supported in part by the Chinese Ministry of Science and Technology under Grant 2012BAH27B03 and in part by the Natural Scientific Foundation of China under Grant 41071257. X. Zhang is with the Institute of Remote Sensing and Geographic Information System, Peking University, Beijing 100871, China (e-mail: xfzhang@pku. edu.cn; [email protected]). J. Zhao is with Beijing Institute of Surveying and Mapping, Beijing 100038, China. J. Tian is with the Department of International Development, Community, and Environment, Clark University, Worcester, MA 01610 USA. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2013.2287513
have been explored, which rely on thermal inertia (TI), spectral vegetation index, and microwave derivatives, respectively [7]–[11]. In comparison, the microwave remote sensing of soil moisture offers several advantages over the others, including the following: 1) a stronger ability to penetrate through clouds; 2) being directly related to soil water content by the soil dielectric constant; and 3) a lower sensitivity to land surface roughness or vegetation coverage [12], [13]. However, although passive microwave remote sensors are able to provide data with a much better spatial coverage, their spatial resolution is often too coarse for the retrieval of highly detailed soil moisture information. Visible and near-infrared (VNIR) and thermal infrared (IR) remote sensing data have been also applied to estimate soil moisture since the 1980s and are still widely employed in monitoring soil moisture due to their better availability and higher resolution in characterizing the reflective and emissive properties of soil and vegetation [8], [14], [15]. TI and temperature vegetation dryness index (TVDI) are two important derivatives from remotely sensed VNIR- and IR-band data, which are commonly used in developing soil moisture inversion models [2], [7], [8], [16], [17]. However, the TI-based methods are not applicable in densely vegetated areas and can hardly provide an accurate estimation of soil moisture in vegetated lands [10], [15], [18], [19]. The TVDI-based approaches were therefore proposed to cope with soil moisture estimation in vegetated areas [8], [20], [21]. Unlike remotely sensed microwave data, thermal IR and VNIR data are often interfered by clouds even in arid areas, and consequently, the TI- or TVDI-based inversion models (or the like) are not suitable for the continuous mapping of soil moisture at a large scale. The synthetic use of multiple remote sensors for mapping soil moisture regionally and globally has become a research hot spot in the community of quantitative remote sensing. This paper proposes a multisensor strategy that integrates Moderate Resolution Imaging Spectroradiometer (MODIS) and Advanced Microwave Scanning Radiometer for EOS (AMSR-E) data to build an effective inversion model for the accurate estimation of surface soil moisture (SSM) information across a large arid area. The rest of this paper is structured as follows. Section II describes the multisensor strategy and the integrated inversion model that we developed. Section III presents a case study of applying the model to mapping soil moisture in Xinjiang, a typical arid area in Northwest China. In Section IV, the modeled results are compared with those from a TVDI model and the U.S. National Snow and Ice Data Center (NSIDC) soil moisture products. Section V summarizes the findings and draws some conclusions.
0196-2892 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHANG et al.: ROBUST COINVERSION MODEL FOR SOIL MOISTURE RETRIEVAL FROM MULTISENSOR DATA
II. M ULTISENSOR R ETRIEVAL M ETHOD A. Multisensor Strategy The spectral data collected by different remote sensors can be synthetically utilized in the inversion of SSM. Generally speaking, three strategies are adopted: 1) independently processing each data set involved to obtain the derivatives required for the final inversion process [19], [22]; 2) building an integrated model to inverse soil moisture from multisensor data [23]; and 3) postprocessing (e.g., intersensor validation) with the inversion results from individual sensors to achieve higher accuracy [24], [25]. The approaches adopting the strategies 1) and 3) are relatively easier to implement but have little room for improvement because they do not focus on the core inversion process. In contrast, the integrated inversion models can handle multisensor data and are better suitable for mapping soil moisture across various regions. In particular, when a rapidly increasing number of remote sensors are launched and orbiting about the Earth, the theoretical and practical development for the integration of soil moisture inversion models is of great methodological and economic value [19], [23]. The variable of soil moisture can be assumed to be an addition of two terms during an observation period: baseline value and daily variation. Specifically, the baseline represents the relatively long-term level of soil water content that is determined by the regional climate and the natural environment, whereas the daily variation represents the short-term fluctuations caused by precipitation, evaporation, and/or other random meteorological factors. This concept is depicted as smij (t) = mij + Δmij (t)
(1)
where mij denotes the baseline soil moisture at a pixel (i, j) in an observation period T ; Δmij stands for the daily variation of soil moisture for the pixel (i, j) at time t in T ; and smij is the actual soil moisture (in volumetric fraction) at the pixel (i, j) as of time t. If T is relatively short (e.g., ten days), the local vegetation cover and land surface roughness normally do not significantly change for same area. In this study, MODIS and AMSR-E (both on board the National Aeronautics and Space Administration Aqua spacecraft) data are used to estimate the baseline value and the daily variation of soil moisture, respectively. The integrated analysis of data from these two sensors is expected to improve the soil moisture mapping. B. Inversion of SSM Baseline Levels In remote sensing of natural environments, land surfaces can be usually characterized by their normalized difference vegetation index (NDVI) and be broadly classified into three types: bare land, sparsely vegetated, and densely vegetated. For each type, a tailored inversion model is chosen from the following three to achieve a more accurate estimation of soil moisture baseline level. In Bare Lands: TI-Based Inversion: According to Stisen et al. [20], the TVDI-based models are not applicable if the NDVI value is below 0.1 because the theoretic basis for such models is the vegetation evaporation process. Comparatively, the TI-based model proposed by Price [7] is more suitable for
5231
estimating the baseline soil moisture levels in bare lands. In our approach, apparent TI (ATI) is derived from the MODIS thermal and visible band data and statistically correlated with SSM. Since clouds often partially contaminate the optical and thermal band data remotely sensed over a large area, the MODIS 16-day composite products are chosen to represent the local vegetation condition for their significantly better spatial continuity. ATI can be calculated using [7] ATIi,j =
1 − αi,j ΔTi,j
(2)
where αi,j represents land surface albedo, which can be derived from the MODIS VNIR data; ΔTi,j stands for the surface temperature range over 2 A . M . to 2 P. M . of local time, and i and j denote the coordinates of an image pixel. In Densely Vegetated Areas: TVDI-Based Inversion: The TVDI-based methods estimate SSM through the extraction of water stress index from the T s–NDVI feature space [8]. For a given region, the more the vegetation is covered, the greater the ability for the vegetation to turn the solar radiation energy into latent heat through evapotranspiration. The T s therefore tends to be lower due to less solar energy transformed into sensible heat [8], [26]. The amount of solar radiation absorbed by a land surface and its soil moisture are the two controlling factors of the T s. Given that the direct solar radiation is constant, a surface tends to be warmer if the soil is drier because the weakened evapotranspiration of vegetation will result in less latent heat but more sensible heat and vice versa. The drought in soil layers often stresses the above-growing vegetation by the undersupply of water and causes the foliage temperature to increase. Soil moisture can therefore be quantitatively estimated as a function of canopy temperature and NDVI [8]. The TVDI-based methods, also known as the “triangle” methods, were developed based on the analysis in the T s–NDVI space and have been successfully applied in the estimation of both evapotranspiration [20] and soil moisture [8]. TVDI can be calculated from MODIS data using TVDI = =
T s − T smin T smax − T smin T s − (a2 + b2 NDVI) a1 + b1 NDVI − (a2 + b2 NDVI)
(3)
where T s is derived from the MODIS thermal data. a1 , b1 , a2 , and b2 are the empirical parameters that need to be determined for modeling the wet and dry edges in the T s–NDVI feature space. T smax and T smin are the dry edge and the wet edge values in the T s–NDVI feature space, respectively. NDVI is the mean calculated from the time series of index values available during the period T and is therefore used to represent the averaged condition of local vegetation over the time span. It is well known that the TVDI model [8] has constraints in the soil moisture inversion from remotely sensed data over a large area. First, clouds can hinder the estimation of T s and NDVI, which are required in the TVDI calculation. Second, the variation of T s is a result from not only the vegetation evapotranspiration but also the solar radiation and the atmospheric condition, which adds uncertainty to the TVDI-based
5232
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014
soil moisture inversion. Third, Moran et al. [9] attempted to calculate the theoretical coordinates for the four vertices of the trapezoid in the T s–NDVI feature space and found that T smin was actually not constant but changed with vegetation cover fraction. Thus, the dry and wet edges have to be determined for each particular study. In order to model the edges, the NDVI values can be grouped into equal bins with a width of, for example, 0.01. The maximum T s and the minimum T s within each bin are linearly regressed on their corresponding NDVI, so that the coefficients a1 , a2 (intercepts) and b1 , b2 (slopes) can be determined, respectively. Thus, the TVDI for each pixel can be calculated using (3), and SSM can be finally regressed on TVDI to establish the statistical relationship between the two. It should be noted that the TVDI value ranges from 0 to 1. The MODIS cloud free products are chosen for the spatially continuous estimation of T s required by the TVDI model. More specifically, the MODIS 16-day maximum temperature composite data are mapped to the T s–NDVI space. It is understandable that the direct solar radiation reaching the surface varies with elevation and geographic latitude, which violates the assumption of homogeneous solar radiation in the TVDI model. Therefore, before being used to model and calculate TVDI, the temperature difference due to the geographic variation of solar radiation received at the surface is corrected through a regression analysis that is similar to the work by Stisen et al. [20]. Please refer to Zhao et al. for details [27]. In Sparsely Vegetated Areas: Averaging the Two: Due to the fact that the vegetation evaporative cooling process is very weak in sparsely vegetated areas, either an ATI-based or a TVDIbased model alone tends to produce inaccurate SSM inversion results. However, averaging the inversion results from the two models seems to be a better solution for estimating the baseline level of SSM in those transitional zones from “bare lands” to “densely vegetated.” C. Estimation of Daily Variation Daily variation of soil moisture over a large area can be derived from passive microwave remote sensing. For example, the spaceborne passive microwave sensor AMSR-E collected brightness temperature data through the four frequency channels of 6.9, 10.7, 18.7, and 36.5 GHz and is able to cover the entire globe every two days, with a Sun-synchronous orbit (ascending and descending at local time 1:30 P. M . and 1:30 A . M ., respectively) [12]. Due to the fact that the data at 36.5 and 18.7 GHz are heavily influenced by clouds, whereas the C-band (6.9 GHz) data are vulnerable to radio frequency interference, the X-band (10.7 GHz) data are left to be the most suitable for soil moisture inversion [13], [28]. The SSM inversion models using AMSR-E data are mostly based on analyzing the derivatives (e.g., polarization ratios) from the AMSR-E brightness temperature data. According to Njoku and Chan [29] and Zhang et al. [13], the daily variation of SSM at a location can be estimated using 2 Δmv = k1 (Pr − Prmin )Prkmin
(4)
where Δmv is the daily SSM variation. Pr stands for a polarization ratio (TBv − TBh )/(TBv + TBh ), in which TBv and
Fig. 1.
Study area of Xinjiang and the field survey area.
TBh are the vertically and horizontally polarized AMSR-E brightness temperatures, respectively. Prmin is the minimum polarization ratio value for a pixel in a certain observation period. k1 and k2 are the best fit coefficients. The change Δmv is mainly driven by the local land surface evaporation and precipitation and represents the daily fluctuation of SSM. D. Integrated SSM Inversion Model The integrated SSM inversion model follows three major steps as follows: Step 1) To estimate the baseline values of SSM from a MODIS 16-day composite image in the ways described in Section II-B; Step 2) To estimate the daily variation values of SSM in the 16-day period using the AMSR-E data at 10.7 GHz and (4); Step 3) To combine the baseline and the daily variation values of SSM obtained from Steps 1 and 2 to produce large-scale daily soil moisture maps at a 1-km resolution. III. A PPLYING THE M ODEL : A C ASE S TUDY A. Study Area Our study area is located in Northwest China, including the entire Xinjiang Uygur Autonomous Region (see Fig. 1). Compared with East China, Xinjiang is a large and sparsely populated region that occupies approximately one sixth of the country’s territory. It borders Tibet to the south; the Qinghai and Gansu provinces to the southeast; Mongolia to the east, Russia to the north; and Kazakhstan, Kyrgyzstan, Tajikistan, Afghanistan, and the Pakistan- and India-controlled parts of Kashmir to the west. The topography of Xinjiang can be described as “three mountains delineate two basins”: The Altai, Tianshan, and Kunlun Mountains lie on the north border, through the central area, and on the south border of Xinjiang, respectively. The Zhungeer Basin is bordered by the Tianshan and Altai Mountains, and the Tarim Basin is situated between the Kunlun and Tianshan Mountains.
ZHANG et al.: ROBUST COINVERSION MODEL FOR SOIL MOISTURE RETRIEVAL FROM MULTISENSOR DATA
5233
TABLE I PARAMETERS U SED IN THE I NVERSION OF S OIL M OISTURE IN X INJIANG
The study area has a typical central Asian arid climate with an average annual precipitation of 150 mm. The spatial distribution of the precipitation is quite heterogeneous. Typically, a sequence of landscapes (snow/ice covers → high-mountain meadows → forest → well-grown dense grassland neighboring the forest → very sparse hilly grassland → natural–artificial oasis → desert) can be seen for a mountain in Xinjiang when traversing from its top, through the hills, and to one of the central basins. The water resources in Xinjiang are extremely important for the regional ecoenvironmental and agricultural management and largely shape the landscapes’ evolution. Thus, it is of great value to closely and continuously monitor the soil moisture condition in the vast area of Xinjiang through satellite remote sensing. B. Data Acquisition and Preprocessing The MODIS 16-day synthesized NDVI product (MOD13A2) and the MODIS daily land surface temperature (LST) product (MYD11A1) for May and August 2009 were downloaded from the USGS website. The MODIS Reprojection Tool software provided by the USGS was initially used to preprocess the MODIS data. Clouds were subsequently masked out from the preprocessed data, and the MODIS LST data were further corrected for topographic elevation and geographic latitude [27]. A time series of the corrected LST data was then used to produce the 16-day composites of maximum temperature. The Interactive Data Language (IDL) programming language was used to implement the modules/tools that allowed us to perform the specific processing and calculations as needed in the study. The resampled brightness temperature (TB ) data sets (AMSR-E_L3_DailyLand_V06) for the same periods were obtained and used to estimate the daily variations of SSM across Xinjiang. The NSIDC’s soil moisture products were also downloaded for comparison purposes, as presented in Section IV. Additionally, for model evaluation and validation, field data were collected in an area that is representative of the Xinjiang’s landscapes: the northern slopes of the Tianshan Mountain. The field soil moisture data were taken by two means: mobile measurements using the WET instrument developed by Delta-T Devices and fixed observations using the WatchDog2400 Irrigation Station developed by Spectrum Technologies, Inc. Specifically, one WatchDog2400 station was set up within each of the five different land covers in the area: forest, meadow, sparse grassland, cotton farmland, and desert. These stations continuously collected the 5- to 10-cm SSM data on an hourly basis from May 7 through October 11, 2009. Moreover, the
Fig. 2. Inversed soil moisture distribution across Xinjiang through the integrated use of AMSR-E and MODIS data. (a) and (b) Daily variations derived from the ascent and descent overpasses of AMSR-E. (c) Baseline soil moisture derived from the MODIS data. (d) Integratively reversed soil moisture.
conventional yet reliable loss-on-drying method was employed to measure the real soil moisture content in the field samples and to calibrate the measurements by the two instruments. C. Implementation of the Integrated Model and the Results The whole SSM inversion model was implemented in the IDL programming environment. The procedures involved in the model are described in details as follows. Estimation of Baseline Soil Moisture: Using the methods described in Section II-B, the procedure for baseline SSM estimation is illustrated in a conceptual coding format as shown below. /∗ for bare land If NDVI < 0.1 Bsm1 = c1 + d1 ∗ ATI /∗ for densely vegetated Else If NDVI > 0.18 Bsm2 = c2 + d2 ∗ TVDI /∗ for sparsely vegetated (neither densely vegetated nor bare land) Else Bsm3 = (Bsm1 + Bsm2 )/2 where c1, c2, d1, and d2 are the coefficients obtained from the regression analysis of the in situ soil moisture measurements on the derived indices (ATI and TVDI) from the MODIS data.
5234
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014
Fig. 3. Soil moisture variation in Xinjiang in May and August 2009 (the maps were produced from the AMSR-E and MODIS data using the integrated inversion model).
The threshold values of NDVI (e.g., 0.1 or 0.18) for differentiating land cover types were determined by finding the abrupt changing points in the correlation coefficients between the in situ measured SSM values and the ATI or TVDI values. As can be seen, the estimation procedure applies the most suitable model for each pixel based upon its NDVI condition. Estimation of Daily Soil Moisture: The procedure for daily SSM variation estimation is illustrated below in the conceptual coding format as well. If Prmin ≤ Pr ≤ 3 Prmin Δmv = 72.58(Pr − Prmin ) Pr−0.625 min If Pr > 3 Prmin let Pr = 3 Prmin Δmv = 145.16 Pr0.375 min Note: please refer to (4) for the meanings of the denotations. The daily variation of SSM is mainly driven by the local precipitation and evaporation. The upper condition in the code is for the usually rapid increase of SSM in response to a precipitation process, whereas the lower condition is for the exceptions when the surface soil layers are saturated with water and the soil moisture will therefore stop increasing regardless of further water input. It should be pointed out that such exceptional scenarios are not well dealt with in the empirical models used by NSIDC to produce their soil moisture products. More detailed discussion about this can be found in Zhang et al. [13]. The model parameters derived from the in-situ measurements in Xinjiang are listed in Table I. Fig. 2 displays the inversion results of baseline SSM and daily SSM variation in Xinjiang on May 25, 2009. As shown in Fig. 2(a) and (b), the distribution of Δmv well matches the precipitation condition over northern Xinjiang on May 25, 2009. The precipitation data collected by the China Meteorological Administration operated weather stations (see Fig. 1) showed that there was strong precipitation taking place in northern Xinjiang on that day. It can be recognized that the inversed soil moisture baseline values [see Fig. 2(c)] form zones that spatially correspond to the different land covers in Xinjiang. Fig. 2(d) shows the final SSM inversion map of May 25, 2009 by combining the baseline and the daily variation. Similar maps
Fig. 4. Comparison of the in situ measurements, the soil moisture data sets produced by the integrated inversion model and the TVDI-based model, and the NSIDC soil moisture products for the different land covers in May 2009.
were also produced for May 5, 20, and 30, 2009 and August 5, 20, and 30, 2009, as displayed in Fig. 3. It can be interpreted that in both May and August 2009, the soil moisture in the Taklimakan Desert was relatively low overall, whereas both sides of the Tianshan and Kunlun Mountain ridges had a much higher level of soil moisture ranging from 10% to 40%. The soil moisture in the other areas of Xinjiang was between 5% and 15%. The soil moisture spatial distributions are consistent with our knowledge about the region and the field investigation. IV. ACCURACY A SSESSMENT AND D ISCUSSION A. Accuracy Assessment As described before, the field soil moisture data consist of the measurements taken by the WET and WathcDog2400 instruments in the study area. The WET data have a fairly good
ZHANG et al.: ROBUST COINVERSION MODEL FOR SOIL MOISTURE RETRIEVAL FROM MULTISENSOR DATA
5235
Fig. 5. Correlation analysis of the in situ measurements with the (a) TVDI-based inversion results, the (b) NSIDC soil moisture products, and the (c) integrated inversion results in May 2009.
Fig. 6. Correlation analysis of the in situ measurements with the (a) TVDI-based inversion results, the (b) NSIDC soil moisture products, and the (c) integrated inversion results in August 2009.
spatial coverage, whereas the WatchDog2400 data have a good temporal frequency of every hour. The WET data were first calibrated using the WatchDog2400 data with the aid of the data collected by the traditional loss-on-drying method and were then used as the reference in the assessment and comparison of the soil moisture products by the different approaches. The performance of the proposed inversion model was evaluated by comparing its results to those from the original TVDIbased inversion model, the NSIDC soil moisture products, and the in situ measurements (see Fig. 4). The assessment was also broken down by land cover type, comparing the model’s performance in forestland, dense grassland, and sparse grassland. Cotton fields and desert areas were excluded because the soil moisture in the former is dominated by human factors such as irrigation, and neither WET nor WatchDog2400 could function well to take reliable readings in the loose sand of desert. It is found that the inversion results of SSM with the proposed model are highly correlated with the in situ measurements in all the three types of land covers in both May (r2 = 0.85) and August (r2 = 0.76) 2009 (see Figs. 5 and 6). The root-mean-square errors (RMSEs) are 3.99% and 4.43% in the two months, respectively. In comparison, the correlation between the in situ measurements and the inversion results using the TVDI-based model (r2 = 0.36 in May and 0.55 in August) or the NSIDC soil moisture products (r2 = 0.33 in May and 0.65 in August) is much lower (see Figs. 5 and 6). The corresponding RMSEs are also significantly larger (TVDI-based model: 14.86% in May, 8.62% in August; NSIDC products: 8.96% in May, 8.95% in August), suggesting that the proposed model has stronger estimation power. It should be pointed out that the soil moisture maps produced using the original TVDIbased model are spatially discontinuous in all the three types of land covers primarily due to cloud contamination in the data (even for arid areas). This is a general limitation for estimating soil moisture from remotely sensed optical or thermal infrared
data. The temporal variation of SSM in the sparse grasslands is well captured in our inversion results and the NSIDC soil moisture products, which confirms the effectiveness of AMSR-E data in mapping soil moisture dynamics in sparsely vegetated areas. B. Discussion The assessment of the SSM inversion results from the AMSR-E and MODIS data is not very direct due to the spatial resolution (1 km) of the sensors. The problem is attributed to the uncertainty in characterizing the soil moisture of a 1 × 1 km2 cell, in which one point measurement or the average of a few scattered point measurements is usually used to represent the soil moisture of the whole cell. Such representation can sometimes be unreliable and uncertain to some degree. Moreover, the inversed soil moisture information from the AMSR-E X-band measurement only characterizes approximately the top 2-cm layer of the soil [12], [30]. The moisture in such a surficial soil layer changes too frequently with rainfall, wind, and temperature to be accurately measured in the field. The moisture in the top soil layer of 5–10 cm was alternatively measured in the field and used for the model assessment; the inversion results, consequently, bear some inherent errors when assessed using the field data as the reference [31]. It is understandable that vegetation cover tends to interfere the estimation of the daily variation of SSM with the passive microwave AMSR-E data. However, as shown in Fig. 4, the proposed inversion model performed quite well even in the densely vegetated areas such as forestland and highdensity grassland. This is largely attributed to our model’s ability to complement the strength of the sensors in SSM reversion while minimizing their limitations. Furthermore, the integrated inversion model proposed also benefits from the improvement made to Sandholt’s TVDI model. Three main
5236
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014
adjustments were made: 1) masking out of clouds and synthesizing T s from the 16-day observations in order to minimize cloud contamination; 2) correcting for elevation and latitude to standardize the solar radiation over a relatively large area; and 3) determining the wet and dry edges in the T s–NDVI space from the actual data rather than using the theoretical ones. These adjustments also greatly contributed to the superior performance of our proposed model than the original TVDI-based model in estimating the baseline SSM from the MODIS data. The heterogeneity of soil moisture distribution in arid areas such as the study area is relatively smaller than that in humid areas; the spatial variation of soil moisture within each AMSR-E image pixel, therefore, can be more effectively reflected by the baseline values extracted from the 1-km MODIS data. However, a hyperlocal weather event (e.g., thunderstorms affecting 3–5 km2 ) may dramatically elevate the soil moisture level in an area that is much smaller than the AMSR-E cell size (25 × 25 km). Such an elevation is therefore difficult to detect from the AMSR-E data. The future generation of passive microwave sensors with stronger spatial resolving power is expected to overcome the aforementioned limitation of our proposed model. V. C ONCLUDING R EMARKS This paper has proposed an integrated model for SSM inversion from multisensor data over a large geographic area. The basic idea underlying the model is to decompose the SSM value at an image pixel into two terms (baseline value and daily variation) and to model them independently. The baseline SSM in a relatively short period can be estimated from the MODIS optical and thermal infrared data, and the daily variation of SSM can be estimated using the AMSR-E X-band microwave data. The proposed inversion model was applied to estimate and map SSM in Xinjiang, China, at a 1-km resolution. The accuracy assessment has shown that, with a stronger correlation (r2 = 0.86 for May, r2 = 0.76 for August) with the reference data and a smaller RSME (3.99% for May and 4.43% for August), our model outperformed the original TVDI-based model that replies on MODIS data only; it also produced more accurate results than the NSIDC products that were solely derived from AMSR-E data. In addition, our model is generally applicable in areas that are densely vegetated (e.g., forest, grassland, and cropland), sparsely vegetated, or bare (e.g., desert). It therefore has a greater ability to produce spatially continuous SSM maps for a large region on a more frequent basis. The proposed model is highly promising for mapping SSM in other similar natural environments. The parameters required by the model should be calibrated using the local field data to adapt to a particular area. Caution needs to be exercised, particularly when estimating the surface soil moisture in agricultural areas due to their pronounced artificial processes such as irrigation and cultivation. Future research efforts should be made to more explicitly define the physical meanings of the baseline and daily variation of SSM.
ACKNOWLEDGMENT The authors would like to thank the anonymous reviewers for their valuable comments and suggestions. R EFERENCES [1] M. C. Anderson, J. M. Norman, G. R. Diak, W. P. Kustas, and J. R. Mecikalski, “A two-source time-integrated model for estimating surface fluxes using thermal infrared remote sensing,” Remote Sens. Environ., vol. 60, no. 2, pp. 195–216, May 1997. [2] A. Verhoef, “Remote estimation of thermal inertia and soil heat flux for bare soil,” Agr. Forest Meteorol., vol. 123, no. 3/4, pp. 221–236, Jun. 2004. [3] X. Zhan, P. R. Houser, J. P. Walker, and W. T. Crow, “A method for retrieving high-resolution surface soil moisture from hydros L-band radiometer and radar observations,” IEEE Trans. Geosci. Remote Sens., vol. 44, no. 6, pp. 1534–1544, Jun. 2006. [4] S. N. Goward, Y. Xue, and K. P. Czajkowski, “Evaluating land surface moisture conditions from the remotely sensed temperature/vegetation index measurements: An exploration with the simplified simple biosphere model,” Remote Sens. Environ., vol. 79, no. 2/3, pp. 225–242, Feb. 2002. [5] P. C. Doraiswamy, J. L. Hatfield, T. J. Jackson, B. Akhmedov, J. Prueger, and A. Stern, “Crop condition and yield simulations using Landsat and MODIS,” Remote Sens. Environ., vol. 92, no. 4, pp. 548–559, Sep. 2004. [6] B. Narasimhan and R. Srinivasan, “Development and evaluation of Soil Moisture Deficit Index (SMDI) and Evapotranspiration Deficit Index (ETDI) for agricultural drought monitoring,” Agr. Forest Meteorol., vol. 133, no. 1–4, pp. 69–88, Nov. 2005. [7] J. Price, “On the analysis of thermal infrared imagery: The limited utility of apparent thermal inertia,” Remote Sens. Environ., vol. 18, no. 1, pp. 59–73, Aug. 1985. [8] I. Sandholt, K. Rasmussen, and J. Andersen, “A simple interpretation of the surface temperature/vegetation index space for assessment of surface moisture status,” Remote Sens. Environ., vol. 79, no. 2/3, pp. 213–224, Feb. 2002. [9] M. S. Moran, T. R. Clarke, and Y. Inoue, “Estimating crop water deficit using the relation between surface air temperature and spectral vegetation index,” Remote Sens. Environ., vol. 49, no. 3, pp. 246–263, Sep. 2004. [10] W. W. Verstraeten, F. Veroustraete, C. J. van der Sande, I. Grootaers, and J. Feyen, “Soil moisture retrieval using thermal inertia, determined with visible and thermal spaceborne data, validated for European forests,” Remote Sens. Environ., vol. 101, no. 3, pp. 299–314, Apr. 2006. [11] R. Panciera, J. P. Walker, J. D. Kalma, D. Jetse, E. J. Kim, K. Saleh, and J. P. Wigneron, “Evaluation of the SMOS L-MEB passive microwave soil moisture retrieval algorithm,” Remote Sens. Environ., vol. 113, no. 2, pp. 435–444, Feb. 2009. [12] E. G. Njoku, T. Jackson, V. Lakshmi, T. Chan, and S. Nghiem, “Soil moisture retrieval from AMSR-E,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 2, pp. 215–229, Feb. 2003. [13] X. Zhang, J. Zhao, Q. Sun, X. Wang, Y. Guo, and J. Li, “Soil moisture retrieval from AMSR-E data in Xinjiang (China): Models and validation,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 4, no. 1, pp. 117–127, Mar. 2011. [14] L. Jiang and S. Islam, “Estimation of surface evaporation map over southern great plains using remote sensing data,” Water Resour. Res., vol. 37, no. 2, pp. 329–340, Feb. 2001. [15] S. Lu, Z. Ju, T. Ren, and R. Horton, “A general approach to estimate soil water content from thermal inertia,” Agr. Forest Meteorol., vol. 149, no. 10, pp. 1693–1698, Oct. 2009. [16] R. Kimura, “Estimation of moisture availability over the Liudaogou river basin of the Loess Plateau using new indices with surface temperature,” J. Arid Environ., vol. 70, no. 2, pp. 237–252, Jul. 2007. [17] G. Hulley, S. Hook, and A. Baldridge, “Investigating the effects of soil moisture on thermal infrared land surface temperature and emissivity using satellite retrievals and laboratory measurements inertia approach,” Remote Sens. Environ., vol. 114, no. 7, pp. 1480–1493, Jul. 2010. [18] M. Minacapilli, M. Iovino, and F. Blanda, “High resolution remote estimation of soil surface water content by a thermal inertia approach,” J. Hydrol., vol. 379, no. 3/4, pp. 229–238, Dec. 2009. [19] C. Mattar, J. Wigneron, J. Sobrino, N. Novello, J. Calvet, C. Albergel, P. Richaume, A. Mialon, D. Guyon, J. Jiménez-Muñoz, and Y. Kerr, “A combined optical–microwave method to retrieve soil moisture over
ZHANG et al.: ROBUST COINVERSION MODEL FOR SOIL MOISTURE RETRIEVAL FROM MULTISENSOR DATA
[20]
[21]
[22]
[23]
[24]
[25] [26] [27] [28] [29] [30]
[31]
vegetated areas,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 5, pp. 1404–1413, May 2012. S. Stisen, I. Sandholt, A. Nrgaard, R. Fensholt, and K. H. Jensen, “Combining the triangle method with thermal inertia to estimate regional evapotranspiration—Applied to MSG-SEVIRI data in the Senegal River basin,” Remote Sens. Environ., vol. 112, no. 3, pp. 1242–1255, Mar. 2008. Y. Han, Y. Wang, and Y. Zhao, “Estimating soil moisture conditions of the greater Changbai Mountains by land surface temperature and NDVI,” IEEE Trans. Geosci. Remote Sens., vol. 48, no. 6, pp. 2509–2515, Jun. 2010. S. Cros, A. Chanzy, M. Weiss, T. Pellarin, J. Calvet, and J. Wigneron, “Synergy of SMOS microwave radiometer and optical sensors to retrieve soil moisture at global scale,” IEEE Trans. Geosci. Remote Sens., vol. 46, no. 3, pp. 835–845, Mar. 2008. N. Das, D. Entekhabi, and E. Njoku, “An algorithm for merging SMAP radiometer and radar data for high-resolution soil-moisture retrieval,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 5, pp. 1504–1512, May 2011. M. Piles, A. Camps, M. Vall-llossera, I. Corbella, R. Panciera, C. Rüdiger, Y. Kerr, and J. Walker, “Downscaling SMOS-derived soil moisture using MODIS visible/infrared data,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 9, pp. 3156–3166, Sep. 2011. J. Kim and T. Hogue, “Improving spatial soil moisture representation through integration of AMSR-E and MODIS products,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 2, pp. 446–460, Feb. 2012. L. Unganai and F. Kogan, “Drought monitoring and corn yield estimation in southern Africa from AVHRR data,” Remote Sens. Environ., vol. 63, no. 3, pp. 219–232, Mar. 1998. J. Zhao, X. Zhang, and H. Bao, “Soil moisture retrieval from remote sensing data in arid areas using a multiple models strategy,” Adv. Intell. Soft Comput., vol. 105, no. 2, pp. 635–644, 2011. E. G. Njoku, P. Ashcroft, and L. Li, “Statistics and global survey of radio-frequency interference in AMSR-E land observations,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 5, pp. 938–947, May 2005. E. G. Njoku and T. K. Chan, “Vegetation and surface roughness effects on AMSR-E land observations,” Remote Sens. Environ., vol. 100, no. 2, pp. 190–199, Jan. 2006. C. S. Draper, J. P. Walker, P. J. Steinle, R. A. M. de Jeu, and R. H. T. Holmes, “An evaluation of AMSR-E derived soil moisture over Australia,” Remote Sens. Environ., vol. 113, no. 4, pp. 703–710, Apr. 2009. M. H. Cosh, T. J. Jackson, and R. Bindlish, “Watershed scale temporal and spatial stability of soil moisture and its role in validating satellite estimates,” Remote Sens. Environ., vol. 92, no. 4, pp. 427–435, Sep. 2004.
5237
Xianfeng Zhang received the Ph.D. degree in geography from the University of Western Ontario, London, ON, Canada, in 2005. He is currently an Associate Professor with the Institute of Remote Sensing and Geographical Information Systems, Peking University, Beijing, China. His research interests are in remote sensing of ecology, hyperspectral data processing and application, and geospatial data visualization. Prof. Zhang is a Referee for several international academic journals.
Jiepeng Zhao received the Master’s degree in photogrammetry and remote sensing with Peking University, Beijing, China. He is currently a Geomatics Engineer with the Beijing Institute of Surveying and Mapping, Beijing, China. His main research interests include quantitative remote sensing and remotely sensed data processing.
Jie Tian received the Master’s degree in remote sensing from the University of Western Ontario, London, ON, Canada, in 2004 and the Ph.D. degree in geographic information system (GIS) from Queen’s University, Kingston, ON, in 2009. He is an Assistant Professor of GIS with the Department of International Development, Community, and Environment, Clark University, Worcester, MA, USA. His research interests broadly focus on geospatial information science and technology and their application in environmental monitoring and modeling, landscape ecology, public health, air quality, and renewable energy.
