Publication Year:
Author(s): Li, J., Zhou, Z, Deng, M.; Zhou, X, Chen, J.
Abstract
This article presents a pseudomultitask (PMT) segmentation neural network (PMTNet) for cropland mapping in mountainous regions using high-resolution remote sensing images. PMTNet extends BsiNet by introducing two key innovations: 1) a pixel-level mask and edge features fusing module using distance features (MEF_D), and 2) a PMT module that replaces the conventional multibranch-task predictions. The MEF_D module leverages spatial attention guided by distance features as weighting indicators to effectively fuse mask and edge features at the pixel level, leading to improved boundary representation. The PMT module, serving as the core prediction component, consists of a single branch dedicated to mask prediction. The two auxiliary tasks—edge detection and distance mapping—are derived directly from the mask output using the Canny edge detecting algorithm and Euclidean distance transformation, respectively. The model was trained and evaluated using cropland samples from Chongqing and Fenghuang, China, based on high-resolution remote sensing images. Comparative experiments were conducted against two representative multitask neural networks (BsiNet and SEANet) and two transformer-based semantic segmentation models (HRFormer + OCR and LRFormer). The results demonstrated that PMTNet consistently outperformed these baselines, achieving the highest scores across multiple metrics, including precision, recall, F1-score, intersection over union, overall accuracy, and the Kappa coefficient—all within a compact model size. Applicability analysis confirmed that PMTNet can effectively identify croplands of diverse types, shapes, and cultivation stages, as long as their boundaries in the images are visually distinguishable. Moreover, the MEF_D module provides a robust strategy for fusing mask and edge features at the pixel level, significantly enhancing the edge delineation of the extracted results.
Citation
Li, J., Zhou, Z., Deng, M., Zhou, X., & Chen, J. (2026). A pseudomultitask neural network classification model for cropland mapping in mountainous areas using high-resolution remote sensing images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 19, 8092–8107. https://doi.org/10.1109/JSTARS.2026.3667717