NIR-RGB-M²Net: A Fusion Model for Precise Agricultural Field Segmentation Using Multisource Remote Sensing Data

Precise extraction of agricultural field parcels is critical for resource management and yield prediction. Multi-source remote sensing combines near-infrared (NIR) and visible light (RGB) data to leverage complementary features, but fusing these modalities often requires complex networks that risk losing vegetation signals and boundary details. To address this issue, this paper proposes NIR-RGB-M2Net, a novel fusion model for precise agricultural field segmentation using multi-source remote sensing data. The primary innovations of NIR-RGBM2Net are as follows: A U-Net–based fusion model that integrates a Convolutional Block Attention Module (CBAM) in the NIR branch to highlight vegetation-related channels and spatial regions, and a Dilated Convolution Module (DCM) in the RGB branch to expand the receptive field without sacrificing resolution, capturing fine boundary textures. This dual-path design enables simultaneous extraction of deep vegetation cues and precise object contours. Evaluated on the Hi-CNA dataset and Belgium dataset, the accuracy rate, pixel precision, IoU and F1 scores of the proposed model were 94.54%, 93.53%, 89.65% and 94.54%, as well as 86.35%, 82.46%, 73.58% and 84.78%, respectively. This method provides a comprehensive and accurate solution for precise farmland extraction, offering significant implications for agricultural resource management, crop yield prediction, and sustainable agricultural practices.