Satellite data reconstruction under cloud cover for corn yield forecasting via multimodal fusion

Abstract

A comprehensive intelligent system for corn yield prediction based on the synergy of a Spatio-Temporal Generative Adversarial Network (ST-cGAN) and a recurrent CNN-LSTM architecture has been developed and tested. A multimodal fusion of weather-independent Sentinel-1 radar data, ERA5-Land meteorological factors, and historical Sentinel-2 optical observations was applied. The problem of "biochemical blindness" in radar signals, where radar captures physical plant structure but fails to detect photosynthetic activity, and cloud cover limitations in optical remote sensing was successfully resolved. To achieve this, an early fusion strategy was implemented. The ST-cGAN simultaneously processes current SAR data for the structural macro-state, historical optical data for the last known biochemical baseline, and a 10-day history of meteorological priors to enforce physiological constraints. Consequently, if radar indicates high biomass but meteorological data reveals severe drought, the neural network mathematically recognizes biological stress and synthesizes proportionally depressed NDVI and NDRE indices, preventing the hallucination of falsely healthy crops. Dynamic synthesis of missing NDVI and NDRE vegetation indices was conducted under prolonged continuous cloud cover (up to 21 days), achieving a high structural similarity index (SSIM = 0.89). Temporal discriminators for frame sequence analysis were introduced into the architecture, improving the edge-preserving index by 18 % and minimizing spatial artifacts at field boundaries. Pixel-level regression was performed for a 15,000-hectare test area in the Western Forest-Steppe of Ukraine based on reconstructed time series covering 15 critical phenological stages. It was established that the proposed architecture reduces the root mean square error (RMSE) to 0.48 t/ha with a coefficient of determination (R2) of 0.90. Statistical analysis proved the significant superiority of the developed method over industry-standard gap-filling algorithms (e.g., STARFM). A scalable decision-support system for optimizing harvest logistics and financial planning under any atmospheric conditions was presented. Future research directions involving neural network knowledge distillation for IoT devices and UAV data integration were outlined.

Code and data availability

The article describes a multimodal phenotyping framework (^F-KgZ) for corn yield forecasting, but no blocks contain explicit data or code availability statements, public repository names, accessions, or author-provided URLs. No paper-specific public asset is identifiable.

No evidence-backed public reproduction asset is currently recorded.