Using support vector machine (SVM) for agriculture land use mapping with SAR data: Preliminary results from western Canada

Multi-temporal C-band RADARSAT-2 fine quad-pol data (5 scenes) and X-band TerraSAR-X stripmap data (8 scenes) were collected during the 2009 growing season from late May to late end of August. The study site was located in Carman, Manitoba in western Canada. A crop survey was conducted for 441 fields covering a wide range of annual and perennial crops (15 classes) including beans, canola, corn, fallow, field peas, flaxseed, hay/pasture, potato, soybean, sunflower, barley, oat, spring wheat and winter wheat. A supervised support vector machine method was used for image classification. Results revealed that the use of single-date single-frequency SAR data can only achieve marginal success in separating crop classes over regions with complex crop mixtures (15 classes). Better classification accuracies usually occur at the end of the growing season. The C-band SAR image data are generally superior to the X-band dataset. The overall classification accuracies using multi-temporal, single-frequency SAR data are within 50% for mid-season and 75% at the end of the growing season. The integration of RADARSAT-2 and TerraSAR-X data produced more favorable classification results.