Comparison of Accuracy Between Support Vector Machine and Random Forest Classifiers for Land Use and Crop Mapping Using Multi-Temporal Sentinel-2 Images

Land-cover/land-use maps are necessary for monitoring land changes and proper planning for managers in agriculture, natural resources and environment fields each year. The method of field data collection using GPS and land survey is time-consuming and costly. Therefore satellite images which have entire coverage and repetition of collection, low cost and real-time data, are usually used so that land-cover/land-use maps are produced. Accurate mapping using technique suitable for today is a key factor. Although in the past, conventional classification methods have been applied to images such as Landsat, using new satellite images and modern classifiers specially machine learning has been growing recently and their effectiveness in preparing land-cover/land-use maps has been very successful. Another advantage of satellite images is repetitious collection and according to that, vegetation changes through time can be used to differentiate land cover types. The Sentinel-2 satellite with the superiority of a pixel rating of 10 meters is one of the appropriate tools to discriminate land cover types. In the current study, Support Vector Machine and Random Forest classifiers on multi-temporal Sentinel-2 images were used to differentiate land use and crop types of Sanjabi plain in Ravansar and their accuracies were compared. To do so, after sampling, Principal Component Analysis was performed for four dates in crops’ growing season and PC1,2,3 bands of the images were combined. The two techniques were implemented on the layerstacks of PC1,2,3 bands of the images and the training samples. Results of accuracy assessments showed that Support Vector Machine, with overall accuracy of 91.36% and Kappa coefficient of 0.8927, produces a more precise land use and crop map rather than Random Forest method.