Evaluation of machine learning algorithms to Sentinel SAR data

The present study uses multi-temporal Sentinel-1 SAR dataset for classification of Saharanpur area in the Indo-Gangetic plains with December, January and February month datasets of VV and VV/VH polarization. Support Vector Machine (SVM), Random Forest (RF), Decision Tree (DT) and Artificial Neural Network (ANN) algorithms with six different band combinations was used to classify the data in 6 classes. The highest accuracy was achieved with SVM for December–January combination with Overall Accuracy of 74.36% and a kappa coefficient of 0.6905. SVM algorithm performed the best followed by DT, ANN and RF. It was observed that the accuracy of classification increased with multi-temporal datasets. In SVM and RF the accuracy increased by almost 8% from single to dual date, but no increase in accuracy was observed irrespective of taking three dates. For DT and ANN, the accuracy from single to dual date increased by > 10% and by approximately 3% (Marginal) for three dates. The single date ANN achieved very poor results but with an increase in the datasets, good accuracy was attained. This study, therefore, reveals that with single and dual datasets, SVM and RF performs well and with multi-temporal datasets, DT and ANN can also achieve good accuracy.