A deep hybrid learning approach for spatio-temporal crop classification with UAV-derived field structures and Sentinel-2 time series: case study in the Hangbu River Basin

Taking the Hangbu River Basin in Anhui Province, China, as the study area, this region is characterized by abundant agricultural resources and extensive cultivated land. However, influenced by local terrain and land-use practices, the farmland exhibits a high degree of fragmentation, featuring small, irregularly-shaped plots, diverse crop types, and complex planting patterns, all of which pose significant challenges to remote sensing classification.To address the challenges of limited vegetation classification accuracy in complex farmland scenarios, this study proposes a hybrid Convolutional Neural Network - Recurrent Neural Network (CNN-RNN)model that integrates Unmanned Aerial Vehicle (UAV)-derived high-resolution imagery and Sentinel-2 Normalized Difference Vegetation Index (NDVI) time-series data, incorporating attention mechanisms and a chunking training strategy to synergize spatial features and temporal phenological information. Focusing on the Hangbu River Basin in Anhui Province, China, UAV-extracted field boundaries were combined with multi-temporal Sentinel-2 NDVI data. The model employs a CNN module (ResNet50) to extract spatial texture and spectral features, an RNN module Long Short-Term Memory (LSTM) to capture vegetation growth dynamics, and attention mechanisms to enhance critical spatio-temporal feature selection. Comparative experiments were also done with the independent CNN model and the feed-forward neural network Multilayer Perceptron (MLP) used as the basis, and the experimental results demonstrate that the hybrid model achieves an overall classification accuracy of 95.00% with a Kappa coefficient of 0.95, significantly outperforming standalone CNN (94.00% accuracy) and MLP (49.70% accuracy) approaches. The model excels in fragmented farmland classification, particularly for rice (recall: 100%) and rapeseed (F1-score: 0.99). However, corn classification performance remains suboptimal (F1-score: 0.26), primarily due to limited sample size and spectral similarity with other crops. The study validates the effectiveness of multi-source remote sensing data fusion, though challenges persist in spatial-temporal resolution disparities between UAV and satellite data, as well as acquisition costs. Future work will explore high-resolution heterogeneous data fusion algorithms to further improve classification accuracy in complex agricultural landscapes.