DACTransNet: A Remote Sensing Image Cropland Parcel Segmentation Model Based on Low-Pass Filtering De-Aliasing and CTrans Encoder

To address the issues of “mismerging” and incomplete parcel extraction in current mainstream remote sensing image semantic segmentation models, we propose DACTransNet, a novel model incorporating a Low-pass Filtering De-aliasing Downsampling (LFDD) module and CTrans encoder. Through a systematic analysis of the relationships among pooling aliasing, downsampling, and “mismerging”, we establish their quantitative correlation and design the LFDD module to effectively reduce “mismerging”. By combining the strengths of Convolutional Neural Network (CNN) and Vision Transformer (ViT) architectures, the CTrans encoder enhances both local and global feature extraction capabilities. To resolve the lack of highquality datasets, we constructed a new dataset using cropland remote sensing images from Sanjiang Plain and Liaohe Plain. Experimental results on both our custom dataset and the LoveDA dataset show that DACTransNet achieves the best performance, with mIoU scores of 77.75% and 50.33%, respectively. It significantly reduces “mismerging” and substantially improves global and local feature extraction compared to current mainstream models, confirming its superiority in cropland remote sensing image parcel segmentation.