Synthesizing crop modelling and deep learning for remote estimation of wheat biomass dynamics from multispectral and weather observations

Improving crop productivity while maintaining low environmental impact is essential for sustainable food production under increasing population pressure and irreversible climate changes. Dynamic biomass prediction is critical for effective crop growth monitoring and management, yet existing approaches struggle to provide consistent and reasonable predictions across diverse environments in a rapid, economic, and practical manner. Here, we propose SpecWeaNet, a biophysics-informed neural network framework that integrates explicit biophysical principles governing biomass accumulation with implicit mechanisms learned from representative training data. The framework enables dynamic prediction of wheat biomass from sowing to harvest using daily weather data and limited in-season spectral observations, without requiring model recalibration. From a systematic perspective, SpecWeaNet is designed as a flexible framework, from which we further developed three ready-to-use pre-trained variants with different input configurations tailored to commonly used sensors. Our comprehensive evaluation demonstrates the robustness and generalizability of pre-trained models for seasonal prediction of biomass dynamics from non-daily spectral observations (with random interval between two consecutive observations) and daily weather data, with coefficient of determination (R 2 ) higher than 0.99, relative mean absolute error (RMAE) within 26% and relative root mean square error (RRMSE) within 35% on more than 250,000 in-silico simulation scenarios across diverse environmental conditions, including different years, geographical locations, and crop varieties. Furthermore, validation on multiple field experiments showcases the capability of pre-trained models to provide reliable predictions at both trial (R 2 = 0.77–0.88, RMAE = 20–26%, RRMSE = 29–40%) and plot (R 2 = 0.83–0.93, RMAE = 12–19%, RRMSE = 15–26%) scales, utilizing daily weather observations and available satellite or drone-based imagery. This work demonstrates how integrating crop modelling with artificial intelligence can enable scalable estimation of crop biomass dynamics, advancing remote sensing–based crop phenotyping and monitoring for sustainable agricultural systems.