A novel multimodal fusion deep network for identifying hillside urban expansion

Hillside urban expansion (HUE) threatens fragile and sensitive mountainous and hilly ecosystems, increasing geological and ecological risks that adversely affect human well-being. Accurate HUE identification is therefore essential for understanding land-use conflicts and optimizing urban spatial planning in hilly regions. However, the fragmented and small-scale nature of HUE patches, coupled with severe land-cover confusion, makes spatial identification highly challenging. To address this issue, we proposed a multimodal fusion deep network (HUENet) that integrates spatial- and frequency-domain features for improved HUE identification in complex terrains. First, a prior-based masking strategy was used to construct training and validation datasets. Then, a multi-scale frequency-domain spatial and channel attention (MFSCA) module was introduced to enhance feature representation and fusion. Experimental results show that HUENet significantly outperforms mainstream machine learning and deep learning models, with mIoU improvements of 3.28%–14.67% and notable gains in efficiency. HUENet exhibits high spatial accuracy, achieving an overall accuracy exceeding 80% across related datasets. In addition, HUENet exhibits strong transferability across various topographic conditions and potential for long-term monitoring of HUE dynamics. Our study offers a novel methodological approach for precise HUE identification and provides scientific insights to support sustainable urban development in mountainous and hilly regions.