Crop classification from Sentinel-1 time-series images using extreme Gradient Boosting Algorithm

In this thesis, machine learning algorithms were used and made comparison to make classification of summer crop types in Gediz plain by using multitemporal Sentinel-1 SAR satellite images. Extreme Gradient Boosting (XGBoost) and Random Forest (RF) machine learning algorithms were used in the classification. A series of eight Sentinel-1 SAR images acquired during the crop growth season of 2017 (April 10, May 3, June 2, August 1, September 7, October 10 and November 16) were selected as the images. In addition to original bands of Sentinel-1, the VH/VV ratio band, and the computed texture bands contrast, homogeneity, dissimilarity, entropy, angular second moment, variance, and correlation were included in the classification operation. All time series images and texture bands were used as a layer stack in the classification. To compute classification accuracy, the data collected through field survey and the farmer's registry system database were used as reference data. The crop types used in classification are wheat, tomato, pasture, corn, cotton, wineyard, clover, and olive trees. The reference dataset was randomly divided into two halves, one half for training and the other half for validation. Based on the results obtained, the overall accuracy (83.67%) and value (0.8059) of the XGBoost algorithm were sligthly higher than the overall accuracy (83.55%) and value (0.8056) of the RF algorithm. At crop type level, cotton provided the highest user's accuracy of 99.13%, the highest producer's accuracy of 96.59%, and the highest balanced accuracy of 98.25% in classification using the XGBoost algorithm. In classification using the RF algorithm, again cotton provided the highest user's accuracy of 99.34%, the highest producer's accuracy of 96,87%, and the highest balanced accuracy of 98,40%. On the other hand, clover provided the lowest user's accuracy of %23,96, the lowest producer's accuracy of %10,39, and the lowest balanced accuracy of %55,03 in classification using the XGBoost algorithm. In classification using the RF algorithm, again clover provided the lowest user's accuracy of %25,17, the lowest producer's accuracy of %5,74, and the lowest balanced accuracy of %52,79.