HLMamba: Hybrid Lightweight Mamba-Based Fusion Network for Dense Prediction of Remote Sensing Images

Convolutional neural network (CNN)- and transformer-based methods have made significant advances in dense prediction tasks for remote sensing images (RSIs). Nevertheless, both architectures have inherent limitations: CNNs are constrained by a finite receptive field, rendering capturing long-range dependencies challenging, whereas transformers face challenges in terms of ensuring computational efficiency. Recently, the mamba architecture, based on state-space models (SSMs), has emerged as a promising alternative by enabling efficient modeling of long-term dependencies while reducing linear computational complexity. In this study, we present HLMamba, a hybrid lightweight mamba fusion network that integrates a CNN-based encoder with mamba-based fusion modules and a mamba-based decoder. Specifically, after feature extraction using a CNN-based encoder, we introduce a comprehensive interaction mamba (CIM) to enhance modality fusion by capturing intramodal feature dependencies and facilitating intermodal feature interactions, resulting in more comprehensive fusion features. Moreover, we design a wavelet complementary hybrid mamba (WCHM) to integrate fused features by exploring their correlations in feature and wavelet space for accurate dense prediction. Experimental results on two widely used datasets, ISPRS Vaihingen and ISPRS Potsdam, demonstrate the effectiveness and potential of the proposed HLMamba. By leveraging linear complexity and global modeling capabilities, HLMamba achieves better efficiency and accuracy in RSI analysis than existing transformer-based and CNN-based models. The source code is available at: https://github.com/MAXHAN22/HLMamba