Automatic extraction of forests from historical maps ... - Semantic Scholar
Automatic extraction of forests from historical maps based on unsupervised classification in the CIELab color space P.-A. Herrault1,2, D. Sheeren 1, M. Fauvel 1, M. Paegelow 2 1DYNAFOR
Lab. UMR 1201 INP-ENSAT / INRA University of Toulouse
2GEODE
Lab. UMR 5602 UTM / CNRS University of Toulouse
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Olds maps contain specific information (historical places, historical land cover, building footprints) Interesting for various studies about long-term changes of landscapes,urban development or coastlines evolution For few years, a lot of maps are available thanks to National Archives
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Traditional approach to capture objects in historical maps are based on user intervention (for digitizing) As well-known, very time-consuming, very subjective and not reproducible on large areas
Need to develop automatic approaches
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Some problems to capture automatically features on historical maps
Overlapping of planimetric elements
Poor quality because of scanning procedure
Maps without any colors
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Several authors have already proposed automatic methods to capture geographical objects (Ansoult et al.1990; Li et al.1999; Leyk 2006)
Two major steps used in automatic extraction on scanned thematic maps Features extraction(Classification/Segmentation) Clean-up process (Before or/and after extraction process)
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Historical maps contain different kinds of « Noise » overlapping of planimetric elements Shading effect due to scanning procedure
Method based on image-filtering techniques Convolution filters Morphological filters
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Assigning a new value in each pixel using the pixel values in its neighborhood thanks to a mobile window for convolution thanks to a structuring element for morphological
Ex : Median Filtering
3, 3, 3, 4, 4, 5, 5, 5, 10
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Data capture on historical maps may concern various types of features Text (Cao and al.2002; Centeno 1998)
Regions (Shaw and al.2011; Chiang 2009)
Symbols (Gamba and Mecocci 1999; Boesch 1996)
Lines (Kaneko,1992; Mariani and al.1997)
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Developing automatic procedure to extract forests features from the historical ‘Map of France’ (19th century)
Today, no automatic method available for this map
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Developing automatic procedure to extract forests features from the historical ‘Map of France’ (19th century)
Reproducible on large areas
User intervention limited
Sufficiently generic in order to test it on other objects or other raster-color maps
INTRODUCTION
METHODOLOGY
RESULTS
Three excerpts of 1500*1500 pixels each
Differences in terms of quality, slope and relief Differences of colors for the forest features
CONCLUSION
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
Original map(s)
CONCLUSION
Filtered map(s)
Dilatation
Median filtering
Low-pass filtering
INTRODUCTION
METHODOLOGY
RESULTS
Original map(s)
CONCLUSION
Filtered map(s)
Dilatation
Median filtering
Low-pass filtering
Dilatation (square of 5*5 pixels) filling all possible holes within forests created by text, symbols, elevation contour lines
INTRODUCTION
METHODOLOGY
RESULTS
Original map(s)
CONCLUSION
Filtered map(s)
Dilatation
Median filtering
Low-pass filtering
Median filter (window of 5*5 pixels) reducing remaining elevation contour lines while preserving edges and colors
INTRODUCTION
METHODOLOGY
RESULTS
Original map(s)
CONCLUSION
Filtered map(s)
Dilatation
Median filtering
Low-pass filtering
Low-pass filter (window of 5*5 pixels) Removing the backround noise without blurring image
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Why? RGB (Red Green Blue) not always suited to perform automatic extraction non-uniformity of the luminosity lack of human perception (Angulo and Serra 2003) Other color-space well-known for graphic applications (HSV,HLS) but less suitable for image-processing
INTRODUCTION
METHODOLOGY
Specificities? an axis L (Luminosity) perpendicular to ‘ab’ planes one ‘ab’ plane for each value of Luminosity
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
Specificities? an axis L (Luminosity) perpendicular to ‘ab’ planes one ‘ab’ plane for each Value of Luminance
Advantages? possibility to consider each variation of one color like a succession of pure colors increasing uniformity of the image
CONCLUSION
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
Centroids +b
-b
-a
+a Données non-classées
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
Centroids
Example for K=3
+b
+b
-b
-b
-a
+a Data not classified
CONCLUSION
+a
-a Data classified
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Correcting the non-inclusion of some elements upstream the treatments
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Correcting the non-inclusion of some elements upstream the treatments Morphological opening
removing small isolated pixels which are non-forest features Contextual rules filling holes within forest features after classification
INTRODUCTION
Original Map
METHODOLOGY
RESULTS
CONCLUSION
Dilatation Median Filter
Filtered Map Low-pass filter
INTRODUCTION
METHODOLOGY
Carte originale
RESULTS
CONCLUSION
Dilatation Filtre médian
Carte filtrée Map Original Map
Filtre passe-bas
Filtered Map
INTRODUCTION
METHODOLOGY
Classification in RGB color space
Filtered Map
Classification in L*a*b color space
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
Classification in RGB color space
Filtered Map
Classification in L*a*b color space
RESULTS
CONCLUSION
INTRODUCTION
Original Map
METHODOLOGY
RESULTS
Pre-processings+classification
CONCLUSION
Extracted features not corrected
INTRODUCTION
Original Map
METHODOLOGY
RESULTS
Pre-processings+classification
CONCLUSION
Extracted features not corrected
Post-processings
Final extracted forest
Reconstruction
Extracted features corrected (binary)
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Excerpt 1
Validation by comparing with manual extraction
Excerpt 2
High Global Accuracy Trend to under-detect forests features Excerpt 3
Original Map
Binary extraction layer
Final extraction layer
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Excerpt 1
Validation by comparing with manual extraction
Excerpt 2
High Global Accuracy Trend to under-detect forests features Excerpt 3
Original Map
Binary extraction layer
Final extraction layer
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Excerpt 1
Validation by comparing with manual extraction
Excerpt 2
High Global Accuracy Trend to under-detect forests features Excerpt 3
Original Map
Binary extraction layer
Final extraction layer
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Method relatively robust L*a*b color space well suited to low-quality maps
Trend to under-detect forest features ************************************************************************
Improving post-processings steps (contextual rules)
Testing the method on other objects or others rastercolor maps
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Thank you for your attention
Lab. UMR 1201 INP-ENSAT / INRA University of Toulouse
2GEODE
Lab. UMR 5602 UTM / CNRS University of Toulouse
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Olds maps contain specific information (historical places, historical land cover, building footprints) Interesting for various studies about long-term changes of landscapes,urban development or coastlines evolution For few years, a lot of maps are available thanks to National Archives
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Traditional approach to capture objects in historical maps are based on user intervention (for digitizing) As well-known, very time-consuming, very subjective and not reproducible on large areas
Need to develop automatic approaches
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Some problems to capture automatically features on historical maps
Overlapping of planimetric elements
Poor quality because of scanning procedure
Maps without any colors
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Several authors have already proposed automatic methods to capture geographical objects (Ansoult et al.1990; Li et al.1999; Leyk 2006)
Two major steps used in automatic extraction on scanned thematic maps Features extraction(Classification/Segmentation) Clean-up process (Before or/and after extraction process)
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Historical maps contain different kinds of « Noise » overlapping of planimetric elements Shading effect due to scanning procedure
Method based on image-filtering techniques Convolution filters Morphological filters
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Assigning a new value in each pixel using the pixel values in its neighborhood thanks to a mobile window for convolution thanks to a structuring element for morphological
Ex : Median Filtering
3, 3, 3, 4, 4, 5, 5, 5, 10
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Data capture on historical maps may concern various types of features Text (Cao and al.2002; Centeno 1998)
Regions (Shaw and al.2011; Chiang 2009)
Symbols (Gamba and Mecocci 1999; Boesch 1996)
Lines (Kaneko,1992; Mariani and al.1997)
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Developing automatic procedure to extract forests features from the historical ‘Map of France’ (19th century)
Today, no automatic method available for this map
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Developing automatic procedure to extract forests features from the historical ‘Map of France’ (19th century)
Reproducible on large areas
User intervention limited
Sufficiently generic in order to test it on other objects or other raster-color maps
INTRODUCTION
METHODOLOGY
RESULTS
Three excerpts of 1500*1500 pixels each
Differences in terms of quality, slope and relief Differences of colors for the forest features
CONCLUSION
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
Original map(s)
CONCLUSION
Filtered map(s)
Dilatation
Median filtering
Low-pass filtering
INTRODUCTION
METHODOLOGY
RESULTS
Original map(s)
CONCLUSION
Filtered map(s)
Dilatation
Median filtering
Low-pass filtering
Dilatation (square of 5*5 pixels) filling all possible holes within forests created by text, symbols, elevation contour lines
INTRODUCTION
METHODOLOGY
RESULTS
Original map(s)
CONCLUSION
Filtered map(s)
Dilatation
Median filtering
Low-pass filtering
Median filter (window of 5*5 pixels) reducing remaining elevation contour lines while preserving edges and colors
INTRODUCTION
METHODOLOGY
RESULTS
Original map(s)
CONCLUSION
Filtered map(s)
Dilatation
Median filtering
Low-pass filtering
Low-pass filter (window of 5*5 pixels) Removing the backround noise without blurring image
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Why? RGB (Red Green Blue) not always suited to perform automatic extraction non-uniformity of the luminosity lack of human perception (Angulo and Serra 2003) Other color-space well-known for graphic applications (HSV,HLS) but less suitable for image-processing
INTRODUCTION
METHODOLOGY
Specificities? an axis L (Luminosity) perpendicular to ‘ab’ planes one ‘ab’ plane for each value of Luminosity
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
Specificities? an axis L (Luminosity) perpendicular to ‘ab’ planes one ‘ab’ plane for each Value of Luminance
Advantages? possibility to consider each variation of one color like a succession of pure colors increasing uniformity of the image
CONCLUSION
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
Centroids +b
-b
-a
+a Données non-classées
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
Centroids
Example for K=3
+b
+b
-b
-b
-a
+a Data not classified
CONCLUSION
+a
-a Data classified
INTRODUCTION
1
2
3
4
METHODOLOGY
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Correcting the non-inclusion of some elements upstream the treatments
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Correcting the non-inclusion of some elements upstream the treatments Morphological opening
removing small isolated pixels which are non-forest features Contextual rules filling holes within forest features after classification
INTRODUCTION
Original Map
METHODOLOGY
RESULTS
CONCLUSION
Dilatation Median Filter
Filtered Map Low-pass filter
INTRODUCTION
METHODOLOGY
Carte originale
RESULTS
CONCLUSION
Dilatation Filtre médian
Carte filtrée Map Original Map
Filtre passe-bas
Filtered Map
INTRODUCTION
METHODOLOGY
Classification in RGB color space
Filtered Map
Classification in L*a*b color space
RESULTS
CONCLUSION
INTRODUCTION
METHODOLOGY
Classification in RGB color space
Filtered Map
Classification in L*a*b color space
RESULTS
CONCLUSION
INTRODUCTION
Original Map
METHODOLOGY
RESULTS
Pre-processings+classification
CONCLUSION
Extracted features not corrected
INTRODUCTION
Original Map
METHODOLOGY
RESULTS
Pre-processings+classification
CONCLUSION
Extracted features not corrected
Post-processings
Final extracted forest
Reconstruction
Extracted features corrected (binary)
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Excerpt 1
Validation by comparing with manual extraction
Excerpt 2
High Global Accuracy Trend to under-detect forests features Excerpt 3
Original Map
Binary extraction layer
Final extraction layer
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Excerpt 1
Validation by comparing with manual extraction
Excerpt 2
High Global Accuracy Trend to under-detect forests features Excerpt 3
Original Map
Binary extraction layer
Final extraction layer
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Excerpt 1
Validation by comparing with manual extraction
Excerpt 2
High Global Accuracy Trend to under-detect forests features Excerpt 3
Original Map
Binary extraction layer
Final extraction layer
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Method relatively robust L*a*b color space well suited to low-quality maps
Trend to under-detect forest features ************************************************************************
Improving post-processings steps (contextual rules)
Testing the method on other objects or others rastercolor maps
INTRODUCTION
METHODOLOGY
RESULTS
CONCLUSION
Thank you for your attention
