Research Information System
Remote Sensing in Earth Systems Sciences, vol.9, no.2, 2026 (Scopus)
This study presents the evaluation of the potential of agricultural crop type mapping using a single-date Sentinel-2 imagery and segmentation-based Convolutional Neural Network (CNN) classification. To assess the effectiveness, we compared three CNN models, VGG-16, ResNet-50, and U-Net. Image segmentation was carried out based on multi-resolution segmentation (MRS) method. Image classification was then performed through segment-based logic. An experiment was conducted on an agricultural area selected in Türkiye to evaluate the crop mapping performance and compare the performance of the selected CNN models. Among the models tested, U-Net achieved the highest training accuracy of 92.9%, ResNet-50 attained the highest validation accuracy of 86.3%, and VGG-16 yielded the highest test accuracy of 79.4%. Regarding the accuracy of classified image, VGG-16 yielded the highest overall accuracy (OA) of 82.1%. ResNet-50 ranked second with an OA of 80.6%, while U-Net demonstrated the lowest performance with an OA of 79.5%. We also compared the results of segment-based classification to a pixel-based classification. The pixel-based classification results showed a similar trend with the results of segment-based classification. VGG-16 achieved the highest OA of 75.8%, followed by ResNet-50 with 74.0%, and U-Net with 71.9%. Overall, segment-based classification exhibited approximately 7% higher performance than that of pixel-based classification. For both classification methods, VGG-16 achieved the highest classification accuracy, followed by ResNet-50 and U-Net. Our findings demonstrate that a single-date Sentinel-2 imagery offers an efficient method for crop type mapping, reinforcing its applicability for detecting agricultural patterns within a specific seasonal window in the context of Türkiye.