Spatiotemporal Prediction of River Fish Biodiversity by Multimodal Deep Learning of eDNA and Satellite Imagery.

High-resolution spatiotemporal biodiversity maps are essential for conservation and environmental management of freshwater ecosystems. Yet existing eDNA-based approaches remain largely static and spatially extrapolative, discarding fine-grained landscape structure and overlooking interannual dynamics. Here, we developed a multimodal deep learning framework that integrates raw satellite imagery with environmental tabular covariates to predict fish species richness across river networks, using eDNA-derived richness from 1302 sites across six Chinese river basins as training labels. The framework couples a U-Net encoder-decoder with a novel water-guided spatial attention mechanism that leverages MNDWI-derived water masks to prioritize aquatic and riparian pixels, and an attention autoencoder that compresses high-dimensional environmental variables into a latent representation. Ablation experiments confirmed that multimodal fusion of imagery and tabular data substantially outperformed single-modality baselines (R2 from 0.54 to 0.76), and that the water-guided attention mechanism yielded further gains (R2 = 0.81), demonstrating that raw imagery preserves ecologically critical spatial information lost by conventional buffer-based summarization. The model was validated on fully independent basins (middle Yangtze and Yuan River) and multiyear field surveys (2018-2023). It achieved strong spatial transferability (r = 0.83 and 0.69) without region-specific retraining, and captured interannual recovery trends following the Yangtze fishing ban without any direct policy input. By shifting from static spatial extrapolation to prediction that generalizes across space and time, this framework offers a scalable pathway for freshwater biodiversity monitoring and conservation evaluation under the Kunming-Montreal Global Biodiversity Framework.