Different Machine Learning Applications for Crop Classification via Multi-Temporal Sentinel-2 Images and Farmer Declaration Parcels

In this study, the performances of Random Forests (RF), Artificial Neural Networks (ANN), Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost) algorithms in agricultural crop classification were compared using multi-temporal Sentinel-2 images. The study area is located within an area of approximately 45* 45 km2 within the borders of Çukurova Plain. Within the scope of the study, agricultural crop pattern classification including corn, cotton, wheat, sunflower, sunflower, watermelon, peanut, and citrus trees, as well as double crop corn, soya, and cotton crops planted after wheat using RF, ANN, SVM, XGBoost algorithms with multi-time Sentinel-2 images of 2021. The study used parcels registered in the Farmer Registration System (FRS) as reference data. Before defining the parcels that are called Farmer Declaration Parcels (FDP) as ground truth data, pre-editing and rule-based deletion operations were performed, afterwards false and incorrect declarations were eliminated.10 bands (B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12) of Sentinel-2 imagery from 11 different dates and Enhanced Vegetation Index (EVI) were used. All classification results showed a high overall accuracy (OA) ranging from 85% to 92%. As a result of the classification using a total of 121 different features, the highest overall accuracy OA (92.14%) belongs to the XGBoost algorithm. XGBoost algorithm was followed by RF (89.15%), SVM (86.14%) and ANN (85.48%). In all algorithms, although cotton, maize and wheat gave the highest accuracy in the classification results, double crops (wheat-maize, wheat-soybean, wheat-cotton) especially those with very close phenological periods, could not reach very high accuracy values. As a result of the study, it was observed that crops at different phenological stages achieved high classification accuracy using all algorithms, while crops at very similar phenological stages lower classification results. With the methodology applied in this study, a large number of ground truth data were generated from FDP plots and used for classification.