DecAttNet: A decoder-centric multi-task attention network for rapid, large-scale land use change detection and 2 M resolution products generation

Abstract

To address the issues of insufficient resolution and a lack of fine-grained categories in traditional land use and land cover (LULC) products, this study integrates lightweight deep learning, a decoder-centric multi-task learning architecture, and a significant change attention mechanism to propose a deep learning framework (Decoder-Centric Attention Network, DecAttNet) for large-scale, high-precision land use change mapping. By mitigating cross‑task feature conflicts within the multitask‑learning framework via its significant‑change attention module, DecAttNet can adaptively focus on key change areas while sharing features and improving computational efficiency. Additionally, this study utilized DecAttNet to generate a 2023 land‑use product for North China (ten provinces, municipalities and autonomous regions) (LULC_NorthChina_2023) comprising 28 fine‑grained land‑cover classes based on 2‑meter high‑resolution imagery. The results of this study demonstrate that (1) DecAttNet significantly outperforms other deep learning models in terms of both accuracy and computational efficiency, with an average overall accuracy of 90.28 %, which is higher than the figures achieved by the FCNN model (83.12 %), the ChangeMask model (87.73 %), the BGSNet model (86.63 %), and the TripleS model (88.24 %), while improving computational efficiency by approximately 44 % compared to the FCNN model. (2) Based on an independent field validation dataset, the DecAttNet model achieved high extraction precision (precision = 94.86 %) for changed land cover features. (3) The LULC_NorthChina_2023 dataset constructed in this study, which is based on the DecAttNet model, combines high precision (precision = 91.85 %) with high spatial detail. (4) The DecAttNet model demonstrates significant temporal generalization capabilities, achieving average classification accuracies of 89.23 % and 88.78 % in selected sample test regions in 2024 and 2025, respectively, with interannual variability of less than 1.5 %, and exhibits high stability across different geographic re. The DecAttNet framework provides an efficient and accurate solution for large-scale, high-resolution land use dynamic monitoring, and the resulting dataset serves as a valuable benchmark for fine-scale environmental research.