Polarimetry-based land cover classification with Sentinel-1 data Banqué, Xavier (1); Lopez-Sanchez, Juan M (2); Monells, Daniel (1); Ballester, David (2); Duro, Javier (1); Koudogbo, Fifame (1) 1. Altamira-Information S.L., Barcelona, Spain 2. Universidad de Alicante, Alicante, Spain PolInSAR 2015 1
OUTLINE Introduction Sentinel-1 data classification potential Methodology Land cover classifiers Results Conclusions 2
INTRODUCTION OBJECTIVE: the goal of the presented research work is to assess the capability of Sentinel-1 data for land-cover classification Sentinel-1 coherent dual polarimetric acquisitions present the possibility of exploiting polarimetric features for classification Sentinel-1 mission short revisit time (6 days) together with classification capabilities would foster a myriad of EO applications (Land Use, Emergency management, ) Sentinel-1 good trade-off between spatial resolution and swath allows better terrain coverage than previous spaceborne SAR sensors. The free, full and open data policy adopted for the Copernicus programme foresees access available to all users for the Sentinel data products. This will result in bigger impact of EO applications into society. 3
SENTINEL-1 DATA CLASSIFICATION POTENTIAL Area of interest is the rural zone in south-west Ansbach, Germany. It contains several land cover types including agricultural fields, forest and water bodies (e.g. Altmühlsee lake). The date of data acquisition is 15/11/2014. Google Earth snapshot of the AOI, Ansbach, Germany 4
SENTINEL-1 DATA CLASSIFICATION POTENTIAL S1 SLC RGB composite suggests the capacity of identifying among different land cover types S1 SLC RGB composite (R: VV, G: VH, B: VH/VV ) 5
SENTINEL-1 DATA CLASSIFICATION POTENTIAL SAR data used and location Altmühlsee, Weißenburg-Gunzenhausen, Germany (15/11/2014) S1 IW Single Look Complex data: PARAMETER Coordinate System Pixel Value Bits Per Pixel Polarization VALUE Slant Range Complex 16 I and 16 Q Dual (VV+VH) Ground Range Coverage [km] 251.8 Slant Range Resolution [m] Azimuth Resolution [m] Slant Range Pixel Spacing [m] Azimuth Pixel Spacing [m] 2.7 m (IW1) 21.7 m 2,3 m 17,4 m Incidence angle [ ] 32.9 NESZ [db] < -23.7 db 6
METHODOLOGY 1. Debursting : data handling so that azimuth time is continuous in the SLC to be processed (w.r.t. original SLC product) S1 bursted Image example S1 debursted Image example 2. Calibration: data has been calibrated with L1 Calibration Annotation Data Set according to Sentinel-1 Product Specification document in order to obtain: σ 0,VV σ 0,VH 7
METHODOLOGY 3. De-speckle filtering: Several alternatives were evaluated Boxcar Adaptive: Refined Lee PolSAR Speckle Filter NL-means (Deledalle et al.) 4. Classification: Several approaches have been tested: Supervised Classifiers: Maximum Likelihood, Minimum Distance Supervised Wishart Model Based Classifier 8
LAND COVER CLASSIFIER Supervised approach Five main land cover classes are targeted to be classified: Google Earth snapshot with the identified classes Water (inland) Bare Soil Forest Urban/Man made Crops Span with the chosen training classes ROIs are chosen for training purposes in supervised classifiers as well as for a polarimetric features evaluation to build up the model based classifier 9
LAND COVER CLASSIFIER Supervised Classifiers: Minimum Distance & Maximum likelihood They show similar performance, though Minimum Distance seems to be slightly better It seems that crops are underestimated. Accuracy and confusion matrix (diagonal order is: water, forest, crops, urban, bare soil) Minimum Distance classification result 99,97 0,00 0,00 0,00 0,03 0,00 99,61 0,29 0,03 0,07 0,00 3,89 84,72 11,39 0,00 0,00 19,06 19,56 61,15 0,22 0,31 0,18 0,00 0,00 99,51 RGB composite 10
LAND COVER CLASSIFIER Supervised Classifiers: Wishart Clearly better performance than minimum distance. It seems some agricultural fields are misclassified as urban Accuracy and confusion matrix (diagonal order is: water, forest, crops, urban, bare soil) Wishart classification result RGB composite 99,87 0,00 0,00 0,00 0,08 0,00 97,66 0,31 4,93 0,07 0,00 2,34 97,44 33,49 0,00 0,00 0,00 2,25 61,09 0,00 0,13 0,00 0,00 0,49 99,86 11
LAND COVER CLASSIFIER Model based classifier: Based on the different polarimetric features of the chosen classes. Polarimetric features evaluated: σ 0,VV σ 0,VV Ratio Span Alpha1 Entropy Normalized correlation 12
LAND COVER CLASSIFIER Water Both VV and VH are expected to be very low due to water low backscattering. Bare Soil Both VV and VH are expected to be low due to general low backscattering, just above water backscattering power. By using these two information channels seems feasible to identify between this class and the rest Forest Volumetric backscattering mechanism suggests that VH will have more relevance than for the remaining classes with respect to the total span. Moreover, entropy is expected to be high due to the clear presence of more than one backscattering mechanism in the scatterers, and total span has to be moderate. Urban Urban presents in general high reflectivity values in both channels, but with more heterogeneity, and not presenting reflection symmetry. The polarimetric coherence is expected to be high. In addition, Alpha1 has been observed to be higher than in the case of crops. Crops This class consists of agricultural fields with the main characteristic of presenting a high VV backscatter and much lower cross-polar channel, which involves a very low Alpha1 value. Crops present low coherence and very low Alpha1 angle with respect to urban. 13
LAND COVER CLASSIFIER σ 0,VV < wt_copol σ 0,VH < wt_crosspol SPAN < wt_span σ 0,VV < bs_copol σ 0,VH < bs_crosspol SPAN < bs_span SPAN < frt_span Ratio < frt_ratio Entropy > frt_entropy WATER BARE SOIL FOREST Upon analysis this decision tree is proposed for class differentiation so that probability of correct detection is maximized. Coher > urb_coh α 1 > urb_alpha URBAN CROPS 14
RESULTS Model based classifier results are compared with minimum distance and Wishart for a small test area with presence of all the defined land covers. RGB composite patch for the classification performance comparison 15
RESULTS Results Comparison: Minimum Distance vs Model based (confusion matrix) water forest crops urban bare soil 99,97 0,00 0,00 0,00 0,03 0,00 99,61 0,29 0,03 0,07 0,00 3,89 84,72 11,39 0,00 0,00 19,06 19,56 61,15 0,22 0,31 0,18 0,00 0,00 99,51 water forest crops urban bare soil 99,83 0,00 0,00 0,00 0,17 0,00 99,03 0,91 0,05 0,00 0,00 3,39 96,50 0,11 0,00 0,00 20,99 55,83 23,04 0,14 0,78 0,61 0,08 0,00 98,53 Minimum Distance classification result Model based classification result 16
RESULTS Results Comparison: Wishart vs Model Based (confusion matrix) water forest crops urban bare soil 99,87 0,00 0,00 0,00 0,08 0,00 97,66 0,31 4,93 0,07 0,00 2,34 97,44 33,49 0,00 0,00 0,00 2,25 61,09 0,00 0,13 0,00 0,00 0,49 99,86 water forest crops urban bare soil 99,83 0,00 0,00 0,00 0,17 0,00 99,03 0,91 0,05 0,00 0,00 3,39 96,50 0,11 0,00 0,00 20,99 55,83 23,04 0,14 0,78 0,61 0,08 0,00 98,53 Wishart classification result Model based classification result 17
RESULTS Overall Result: Model based 18
RESULTS Overall Result: Model based 19
CONCLUSIONS AND NEXT STEPS CONCLUSIONS Sentinel-1 SLC data present good potential for land cover classification Preliminary obtained result present a good classification performance (over ~95% for all the classes except urban, which is the real challenge of the classifier) Different evaluated classifiers present similar results Model based classifier has the potential to become unsupervised w.r.t. the other classifiers evaluated NEXT STEPS Evaluate over an area with a good land-cover map (ground truth for performance evaluation) Evaluate the influence of the incidence angle in the classification performance Evaluate the impact of meteorological conditions during the data acquisition in the classification results Potential for crop type mapping evaluation. 20
Thanks for your attention! 21