PMNET: Fine-Grained Extraction of Tidal Flat Farmland Based on Sandglass Module Codec for Remote Sensing Fundamental Model

As a critical land cover type, tidal flats play an essential role in carbon sequestration, ecological conservation, and disaster risk reduction. However, tidal flat areas exhibit significant spatiotemporal variability due to the dynamic influences of tides, waves, and precipitation. Moreover, the complex characteristics of tidal flat farmlands in high-resolution remote sensing imagery, coupled with the limited capabilities of conventional extraction methods, preclude large-scale, high-precision farmland delineation. To address this challenge, this study develops a comprehensive tidal flat farmland dataset using high-resolution remote sensing imagery and proposes a novel remote sensing foundation model based on a Sandglass module encoder-decoder architecture. The key contributions of this work encompass four major innovations:1. Development of a novel lightweight Multi-Scale Edge-Aware Module (MSEAM) that integrates learnable gradient operators with parallel multi-scale convolutions, effectively suppressing background noise while adaptively enhancing critical edge regions; 2. Construction of an efficient Sandglass encoder-decoder structure employing a lightweight Sandglass network as the feature extraction backbone, combined SCSE modules, achieving enhanced feature representation capabilities while reducing parameter count and computational complexity; 3. Introduction of an Adaptive Feature Fusion Module (AFM) that facilitates effective integration of high-level and low-level semantic information through a dynamic weight allocation mechanism, emphasizing useful channel features; 4. Integrating the enhanced Sandglass encoder-decoder model with the large-scale vision-language model CLIP. Experimental validation demonstrates that the proposed model achieves an accuracy of 96.75% and mIoU of 90.42% on the high-resolution tidal flat farmland dataset. So the model can realize the extraction of fine geographic patches of farmland in the coastal tidal flat area, and the extracted plot data can be used as the base map data for the observation and research of crop phenology, which is of great significance for the precise delineation of crop planting areas.