An SAM Fine-Tuning Framework With Frequency-Domain Interactive LoRA for Remote Sensing Change Detection
Top Cited Papers
- 5 January 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64 (01962892) , 1-19
Abstract
Achieving high-accuracy remote sensing change detection (RSCD) algorithms requires high-quality semantic feature extraction from remote sensing images (RSIs). Due to its powerful general-purpose feature extraction capability, the Segment Anything Model (SAM) has found wide application across diverse fields. However, SAM may not be optimally suited for RSIs. To address this limitation, we propose a Frequency-domain Interactive LoRA Fine-tuning Architecture (FILFArch) to enhance the performance of SAM in RSCD tasks. Based on FILFArch, we then develop two task-specific algorithms, the FILFBCD for Binary Change Detection (BCD), and the FILFSCD for Semantic Change Detection (SCD). To enhance the capability of SAM in capturing bi-temporal RSIs feature relationship, the Bi-temporal Feature Interactive LoRA (BIF-LoRA) is designed with Siamese architecture. Within BIF-LoRA, Frequency-Domain Feature Interaction (FDFI) utilizes Fast Fourier Transform Block (FFTB) to fuse bi-temporal frequency-domain features. This enables cross-temporal frequency-domain interaction, effectively discriminating spatio-temporal feature differences. Additionally, we use a shared BCD Decoder to serves as the binary change detector for both FILFBCD and FILFSCD. The BCD Decoder first applies a Coarse Difference Feature Extraction (CDFE) to coarsely fuse deep semantic features, yielding a coarse-grained change feature map. Subsequently, a Frequency-Domain Feature Enhancement (FDFE) refines these abstract features to generate a fine-grained change map. In FILFSCD, FDFE is further utilized to recover semantic change information of each temporal RSIs. Experimental results demonstrate that FILFBCD achieves the highest F1 scores of 83.53%, 66.75%, and 83.79% on BCD datasets MLCD, S2Looking, and SYSU-CD, respectively. Meanwhile, FILFSCD achieves the highest F1 scores of 64.05% and 87.02% on SCD datasets SECOND, and DSCD, respectively. These results demonstrate the effectiveness and versatility of the proposed FILFArch for RSCD tasks. The code is available at https://github.com/juncyan/filora.Keywords
Funding Information
- Macao Polytechnic University (RP/FCA-12/2025)
- Submission Control (sc fca.9624.4fed.4)
This publication has 70 references indexed in Scilit:
- ProFiT: A prompt-guided frequency-aware filtering and template-enhanced interaction framework for hyperspectral video trackingISPRS Journal of Photogrammetry and Remote Sensing, 2025
- A Federated Learning-Friendly Approach for Parameter-Efficient Fine-Tuning of SAM in 3D Segmentation2025
- When Fast Fourier Transform Meets Transformer for Image Restoration2024
- RM-UNet: UNet-like Mamba with rotational SSM module for medical image segmentationSignal, Image and Video Processing, 2024
- Adapting the segment anything model for multi-modal retinal anomaly detection and localizationInformation Fusion, 2024
- Large kernel convolution application for land cover change detection of remote sensing imagesInternational Journal of Applied Earth Observation and Geoinformation, 2024
- Cross-domain landslide mapping from large-scale remote sensing images using prototype-guided domain-aware progressive representation learningISPRS Journal of Photogrammetry and Remote Sensing, 2023
- Attention-guided siamese networks for change detection in high resolution remote sensing imagesInternational Journal of Applied Earth Observation and Geoinformation, 2023
- ChangeMask: Deep multi-task encoder-transformer-decoder architecture for semantic change detectionISPRS Journal of Photogrammetry and Remote Sensing, 2022
- A deeply supervised image fusion network for change detection in high resolution bi-temporal remote sensing imagesISPRS Journal of Photogrammetry and Remote Sensing, 2020