Spatiotemporal deep learning models for predictive analytics of crop health deterioration based on satellite remote sensing and unmanned aerial vehicle data

Abstract

In sustainable agriculture and food security, proper and timely forecasting of the degradation of crop health is vital. The new deep learning model described in this paper is Multi-View Dual-Attention Spatiotemporal Crop Health Forecasting Network (MVDA-STNet), which combines satellite multispectral data, UAV-based hyperspectral data, and successive environmental parameters and carries out fine-grained predictive analytics. It has the advantage of using a two-stream time-space encoder which is used independently to encode coarse satellite data and fine details of UAV imagery followed by use of Dual Attention Fusion Module that aligns spatial and temporal data using cross-view spatial attention and temporal context attention processes. The predicted maps of crop health in the future are estimated by a Transformer-enhanced decoder, and corrected iteratively via a residual feedback loop that uses vegetation indices. The regression accuracy, temporal consistency, spectral fidelity and the cross-view coherence are imposed into a hybrid multi-objective loss to enhance the robustness of the model. Through this combined architecture, crop stress, start of the disease and subsequent reduction in productivity can be detected early and the insights can be specific in terms of obvious agronomic measures to be undertaken. The salt treatment program focuses on scalability of different crop types, different regions with real-time implementation enabled by cloud-AI infrastructure. The suggested MVDA-STNet attained a total prediction accuracy of 94.6% when projecting the decline of crop health.