CTDM-Net: A CNN–Transformer Dynamic Memory Network for Cropland Semantic Change Detection

Abstract

Cropland semantic change detection (CSCD) is crucial for monitoring agricultural land dynamics by identifying pixel-level “from-to” transitions in bitemporal high-resolution remote sensing (RS) images. Unlike general semantic change detection (SCD), CSCD requires distinguishing genuine land cover changes from pseudochanges caused by phenological variations and seasonal effects, which introduce significant appearance ambiguities without semantic shifts. To address these challenges, we propose a novel CNN–Transformer dynamic memory network (CTDM-Net) for robust CSCD. CTDM-Net employs a dual-backbone Siamese encoder, leveraging vision Transformer (ViT) for high-level semantic abstraction and a lightweight convolutional neural network (CNN) for fine-grained spatial details, ensuring consistent bitemporal feature representations. A gated difference extraction module (GDEM) adaptively captures and fuses change-aware spatial cues and semantic context, enhancing detection of subtle changes. In addition, a dynamic memory refinement module (DMRM) with a memory bank iteratively refines change features, suppressing pseudochanges and improving semantic consistency. Experimental evaluations on three CSCD datasets and one general SCD dataset demonstrate that CTDM-Net achieves state-of-the-art (SOTA) performance in change localization and semantic transition identification, with superior robustness and generalization across diverse scenarios. The source code is available at https://github.com/ZhuOO5/CTDM-Net