Explainable Mapping of the Irregular Land Use Parcel With a Data Fusion Deep-Learning Model

Abstract

Real land parcels exhibit high variability in size and complexity. Existing deep learning-based models often rely on grids or simple resampling to ensure consistent input sizes, which makes it hard to accurately represent real land distribution while preserving information. Current research also lacks the interpretability of model recognition and data fusion. In order to address these limitations, this article introduces the irregular parcel classification model (IPCM), a novel multisource fusion approach. IPCM uses Poisson disk sampling to regularize samples while retaining essential information. IPCM’s interpretability is analyzed by exploring its classification process and data fusion using gradient-weighted class activation mapping++ (Grad-CAM++) and Explainable Boosting. The results highlight the model’s attention to different functional categories in multimodal data and reveal the phenomenon of poorer fusion results for certain categories due to information mismatch during the fusion process; furthermore, the model optimizes data fusion weight to enhance the correct information to mitigate the negative impact when multisource data information has excessive mismatch. Optimized IPCM achieves 0.892 test accuracy and 0.862 Kappa on irregular parcels. This research can serve as an important reference for high-precision land use mapping and understanding the data fusion process.