Complex Landscape Rice Extraction Using Integrated Sentinel-2 Spectral–Temporal–Spatial Imagery and a Hybrid Deep Learning Architecture

Rice extraction in complex landscapes is a challenging issue in remote sensing, particularly in areas with diverse land-use types and spatiotemporal variability. To enhance the accuracy of rice extraction, this study proposes a novel approach integrating Sentinel-2 spectral-temporal–spatial imagery with a hybrid deep learning architecture for extracting single-cropping and double-cropping rice. First, a time-series dataset of spectral and texture features was constructed to capture the seasonal variations of rice. Second, an active learning strategy was employed to select high-confidence samples, and spectral, temporal, and spatial information was integrated into a unified dataset. Finally, a hybrid deep learning model, convolutional-Transformer hybrid network (CTH-Net), was developed, combining convolutional neural networks (CNNs) and Transformer networks. The model incorporates a fusion module to effectively integrate multiscale temporal, spatial, and spectral features and a residual module to improve gradient flow, mitigating the vanishing gradient problem in deep networks. Results demonstrate that the CTH-Net achieved 99.69% overall accuracy in rice extraction, maintaining >96% accuracy for single-cropping rice, double-cropping rice, and abandoned land. It outperformed models like CNNs, Transformers, long short-term memory (LSTM), and support vector machines (SVMs) in handling fragmented rice distributions and mixed land types, significantly improving extraction accuracy. This study provides an efficient and reliable solution for rice extraction in complex landscapes, supporting agricultural monitoring and management.