HA2F: Dual-module Collaboration-Guided Hierarchical Adaptive Aggregation Framework for Remote Sensing Change Detection

TL;DR

HA2F introduces a CNN-Transformer hybrid framework for remote sensing change detection that combines a Dynamic Hierarchical Feature Calibration Module (DHFCM) and a Noise-Adaptive Feature Refinement Module (NAFRM) to address multi-scale feature misalignment and noise-induced spurious changes. The architecture fuses global context from ViT with local detail from ResNet18, then dynamically calibrates and refines multi-scale features before producing a change map. Across LEVIR-CD, WHU-CD, and SYSU-CD, HA2F achieves state-of-the-art metrics (P, R, OA, F1, IoU) with superior efficiency, supported by extensive ablations demonstrating the contributions of DHFCM and NAFRM. The approach offers robust performance in diverse remote sensing scenes and provides a foundation for further lightweight and semi-supervised extensions.

Abstract

Remote sensing change detection (RSCD) aims to identify the spatio-temporal changes of land cover, providing critical support for multi-disciplinary applications (e.g., environmental monitoring, disaster assessment, and climate change studies). Existing methods focus either on extracting features from localized patches, or pursue processing entire images holistically, which leads to the cross temporal feature matching deviation and exhibiting sensitivity to radiometric and geometric noise. Following the above issues, we propose a dual-module collaboration guided hierarchical adaptive aggregation framework, namely HA2F, which consists of dynamic hierarchical feature calibration module (DHFCM) and noise-adaptive feature refinement module (NAFRM). The former dynamically fuses adjacent-level features through perceptual feature selection, suppressing irrelevant discrepancies to address multi-temporal feature alignment deviations. The NAFRM utilizes the dual feature selection mechanism to highlight the change sensitive regions and generate spatial masks, suppressing the interference of irrelevant regions or shadows. Extensive experiments verify the effectiveness of the proposed HA2F, which achieves state-of-the-art performance on LEVIR-CD, WHU-CD, and SYSU-CD datasets, surpassing existing comparative methods in terms of both precision metrics and computational efficiency. In addition, ablation experiments show that DHFCM and NAFRM are effective. \href{https://huggingface.co/InPeerReview/RemoteSensingChangeDetection-RSCD.HA2F}{HA2F Official Code is Available Here!}

Figures (12)

• Figure 1: Scheme of the motivation diagram. The proposed HA2F optimizes multi-scale difference feature fusion and purifies noise impurities.
• Figure 2: Overview of the proposed HA2F. Bitemporal features are first extracted using ViT and ResNet18. The DHFCM then integrates cross-layer difference information to generate highly discriminative enhanced advanced features. Next, the NAFRM achieves feature alignment by spatially adaptive transformation and filters change sensitive regions to optimize noise robustness. Finally, the classifier generates a change graph that reflects the actual architecture changes.
• Figure 3: Scheme of the DHFCM. In this module, the DHFCM mainly includes two parts: triple cross-attention mechanism and HAFS. Among them, the HLF represents high-level features and LLF represents low-level features.
• Figure 4: Scheme of the HAFS in DHFCM.
• Figure 5: Scheme of the NAFRM. In this module, SAT generates spatial bias field, after which HAFS filters noise information by double screening strategy.