Parcel-scale summer crop classification based on multi-source remote sensing data and deep learning

Accurate crop structure information is fundamental for precision agriculture, particularly in regions with highly fragmented farmland. This study focuses on Jiangsu Province, China, and proposes a multi-source remote sensing approach to address the classification challenges posed by complex agricultural landscapes. Optical Sentinel-2, synthetic aperture radar (SAR) Sentinel-1, and high-resolution Gaofen-2 (GF-2) imagery were integrated to exploit their spatial and temporal complementarities. A modified multi-scale attention U-Net (MSA-UNet) was developed to extract cropland parcels from GF-2 imagery, which served as the basic classification units. Time-series spectral and polarimetric features derived from Sentinel-1/2 data were used to train a convolutional neural network – bidirectional long short-term memory (CNN – BiLSTM) model for parcel-level summer crop classification. The proposed MSA-UNet achieved an intersection over union (IoU) of 73.89% and an F1-score of 84.66%, outperforming the baseline U-Net by 5.87 and 4.07% points, respectively. Incorporating both optical and SAR features improved the precision to 91.28%, representing gains of 2.37 and 13.87% points compared to models using single-source inputs. The CNN – BiLSTM architecture effectively captured both intra-seasonal dynamics and long-term dependencies in the time series. These results demonstrate that the proposed method significantly improves the accuracy of crop mapping in fragmented agricultural regions and offers a practical solution for large-scale agricultural monitoring and decision support.