A data- and knowledge-driven cropland parcel recognition method based on segment anything model (SAM)

In modern agricultural practices, information on cropland parcels serves as an indispensable foundation of agricultural production management. However, manual annotation of remote sensing image datasets is time-consuming and labor-intensive. Traditional models tend to focus on pixel-level segmentation while neglecting the fine-grained boundary features of cropland parcels. To address these challenges, this study first utilizes the Segment Anything Model (SAM) to generate cropland parcel label samples based on multi-source remote sensing imagery and the Third National Land Survey (TNLS) data. Morphological processing and texture feature screening are employed to enhance the quality of pseudo-labels. Subsequently, we propose a multi-task intelligent cropland parcel extraction model, Squeeze-and-Excitation Multiple Task Decoder Network (SE-MTDNet). This model adds a Normalized Difference Vegetation Index (NDVI) channel to RGB channels to extract hierarchical features. The atrous spatial pyramid pooling (ASPP) module and channel attention are introduced to aggregate multi-scale contextual information. In the decoder, multi-task joint training is adopted to prevent over-reliance on the single results. Experimental results demonstrate that SE-MTDNet significantly improves both the accuracy and boundary clarity of cropland parcel recognition. In Xi County, SE-MTDNet achieved an overall classification accuracy of 94.25%, an Intersection over Union (IoU) of 88.67%, and an F1-score of 92.76%. Furthermore, the model also outperformed other comparison models in transfer validation conducted in Lishu County. Ablation experiments further confirm that key modules and the NDVI feature contribute significantly to improved cropland parcel segmentation accuracy. Our method provides valuable insights into high-precision and automated cropland parcel recognition.