Abstract:

[Objective] Rice-crayfish co-culture represents a highly efficient ecological agricultural system that simultaneously provides substantial economic returns and ecological benefits. Accurately obtaining spatial distribution information on rice-crayfish fields is of great importance for promoting the optimal allocation of agricultural resources, supporting ecological protection, and facilitating sustainable agricultural development. In regions characterized by complex terrain, fragmented plots, and diverse planting structures, traditional extraction approaches are often influenced by spectral confusion and indistinct field boundaries, making it difficult to achieve high-precision identification of rice-crayfish fields. The aim of this research is to develop a deep learning model that integrates multi-temporal and multi-scale feature information to improve the accuracy of remote sensing identification of rice-crayfish fields. [Methods] Multi-temporal GF-2 satellite imagery was employed as the primary data source, and a deep learning-based extraction method named the dual-branch attention pyramid network (DBAP-Net) was developed. The proposed model was established upon the U-Net framework and designed with a dual-branch encoder architecture to extract temporal features from different phenological stages, thereby fully capturing the spectral variations of rice-crayfish fields between the paddy flooding and rice-growing periods. The convolutional block attention module (CBAM) was embedded in each encoder layer and in the skip connections to adaptively adjust feature weights along both the channel and spatial dimensions, enhancing the network's capacity to emphasize critical spatial structures of rice-crayfish fields while effectively suppressing background noise and redundant information. During the feature fusion stage, an atrous spatial pyramid pooling (ASPP) module was incorporated to aggregate contextual information from multiple receptive fields through multi-scale atrous convolutions, improving the model's capability for multi-scale spatial information representation. [Results and Discussions] A quantitative performance evaluation of each DBAP-Net component was conducted through ablation experiments. The results demonstrated that the introduction of a dual-branch structure improved overall accuracy (OA), F₁-Score, intersection over union (IoU), and Matthews correlation coefficient (MCC) by 0.48, 0.57, 0.94, and 0.99 percentage points, respectively. Incorporating the CBAM module further enhanced these metrics by 0.52, 0.85, 1.40, and 1.20 percentage points, while the addition of the ASPP module yielded further increases of 0.59, 1.09, 1.81, and 1.49 percentage points, respectively. The DBAP-Net model achieved the highest comprehensive performance, with an OA of 94.45%, F₁-Score of 91.79%, IoU of 84.82%, and MCC of 87.60%. These values represented respective improvements of 1.83, 2.55, 4.25, and 3.98 percentage points compared with the baseline U-Net model. These findings indicated that each enhancement module made a substantial contribution to improving both feature representation and spatial boundary delineation. DBAP-Net was further compared with five representative semantic segmentation networks, U-Net, PSPNet, DeepLabV3+, SegFormer, and TransUNet, to comprehensively evaluate its generalization and segmentation performance. The results demonstrated that DBAP-Net consistently achieved higher overall accuracy, precision, F₁-Score, and IoU than all other comparison models. Specifically, compared with U-Net, PSPNet, and DeepLabV3+, the F₁-Scores increased by 2.55, 2.98, and 2.75 percentage points, while the IoU values improved by 4.25, 4.96, and 4.57 percentage points, respectively. In comparison with the more recent models, SegFormer and TransUNet, DBAP-Net's F₁-Score was higher by 3.09 and 1.67 percentage points, and its IoU was enhanced by 5.12 and 2.81 percentage points. Visualization of the segmentation results further revealed that DBAP-Net produced clear segmentation boundaries with complete fields, significantly reducing misclassification and omission rates. In contrast, other models exhibited varying degrees of boundary blurring and fragmentation. When applied across the entire study area, DBAP-Net demonstrated strong robustness and stability. The model achieved an overall accuracy of 96.00%, with a Kappa coefficient of 0.920. The producer's accuracy and user's accuracy were 95.13% and 97.38%, respectively. Compared with traditional approaches such as the seasonal water-difference method, random forest classification, and temporal index thresholding, DBAP-Net significantly improved both extraction precision and spatial completeness, particularly under conditions of complex terrain and fragmented agricultural landscapes. [Conclusions] DBAP-Net, by integrating multi-temporal spectral and multi-scale spatial information, significantly improves the accuracy and completeness of remote sensing extraction of rice-crayfish fields from high-resolution remote sensing imagery. The model provides a reliable and adaptable technical framework for fine-scale monitoring, precise mapping, and sustainable management of rice-crayfish co-culture systems, offering valuable methodological support for agricultural resource assessment and ecological protection.

Key words: remote sensing, semantic segmentation, deep learning, rice-crayfish, GF-2