CropNet: Leveraging SegFormer for Efficient and Scalable Crop Mapping with Sentinel-2 Data

Abstract

This research investigates a deep learning-based methodology for crop classification by integrating Sentinel-2 satellite imagery with SegFormer, a state-of-the-art transformer-based semantic segmentation model. The study focuses on five dominant land cover types: rice fields, sugarcane, cassava, para rubber, and pond areas within a part of Khu Mueang District, Buriram Province, Thailand. The main objectives are to develop an efficient classification method using Sentinel-2 satellite data and to evaluate the predictive performance of SegFormer in the agricultural field. Satellite images were acquired via Google Earth Engine (GEE) during the harvest season (Nov 2023–Jan 2024), complemented by ground truth data collected from field surveys and high-resolution drone imagery. Preprocessing steps included cloud filtering, image normalization, and manual pixel-level labeling in QGIS software. The dataset was divided into 512×512 pixel patches, resulting in 780 image–mask pairs allocated for training (480), validation (120), and testing (180). The SegFormer model was trained using Optuna to find the best hyperparameter settings. The model achieved 0.967 pixel-wise accuracy with a validation loss of 0.075 (cross-entropy) on the training and validation datasets, demonstrating strong learning performance during model development. It showed strong classification performance for para rubber and sugarcane. However, it faced challenges in distinguishing cassava, ponds, and bare soil due to class imbalance and spectral similarity.