WFCDCLIP: A CLIP-Based Framework for Weakly Supervised Farmland Change Detection

Farmland change detection is of significant importance for food security, land management, and precision agriculture. However, farmland scenarios generally exhibit high texture similarity and complex background noise, which not only makes manual annotation costly and easily disturbed but also complicates the automatic identification of fine-grained changes. Although weakly supervised methods based on image-level labels can significantly reduce annotation costs, they still face three major limitations under fine-grained changes and noisy environments: semantic mismatch between images and text descriptions, instability of pseudo labels due to noise, and insufficient depth of multimodal information fusion. To address these challenges, this paper proposes a CLIP-based weakly supervised farmland change detection framework, WFCDCLIP, whose core objective is to comprehensively enhance the modeling and discrimination of fine-grained changes through a multimodal semantic alignment mechanism. Specifically, based on the CLIP architecture, this study introduces an attribute-anchored averaged text prompt (AAATPrompt), which anchors text prompt embeddings to class-level change/no-change semantic prototypes and allows only residual adaptation to alleviate semantic mismatch between appearance variations and true semantic changes. A Monte Carlo sampling-based CAM improvement method (MC-CAM) is designed to suppress background interference through multiple samplings and weighted aggregation, thereby improving the stability of pseudo labels in noisy environments. Simultaneously, a text–vision bidirectional interaction mechanism (TVS) is incorporated at the decoder stage to deeply optimize the interaction between images and text, overcoming the insufficient fusion problem of existing methods under complex backgrounds. Finally, a local dense contrastive loss (LLEA) is proposed to enhance the discriminability of multimodal features in local regions, enabling better capture of fine-grained changes in farmland scenarios. These modules collectively form a multi-level multimodal alignment enhancement framework, which explicitly enforces semantic consistency between textual change concepts and fine-grained visual differences, thereby improving the reliability of weakly supervised farmland change detection. Experimental results show that our method significantly outperforms existing state-of-the-art approaches on the HI-CNA and CLCD datasets, especially in terms of Intersection over Union (IoU) and F1 score.