A novel approach to monitor peanut equivalent water thickness through modular training and transfer learning of an improved PROSAIL model using a Wasserstein generative adversarial network

Abstract

Empirical and physical models are widely used for monitoring equivalent water thickness (EWT) to adjust plant moisture management. However, model transferability to different times and locations, and insufficient training data remain the two key challenges of field spectroscopy analysis. Therefore, this study aims to construct a hybrid model, which combines the physical models optimized by Wasserstein Generative Adversarial Nets (WGAN) and empirical models for performing hyperparameter searches (the process of finding optimal model settings) to monitor the peanut EWT. Specifically, we develop a large spectral dataset consisting of field-measured data which including 246 peanut varieties in five peanut farms across China and synthetic datasets generated from the physical models optimized by WGAN. Furthermore, the PWLEH was constructed by hyperparameter tuning and pre-training which using synthetic datasets, and then fine-tuned by modular training with field data of peanut canopy water content. Comparing the model constructed with field data (R² = 0.5618, mean squared error (MSE) = 0.0725) and PROSAIL (a widely used canopy radiative transfer model) (R² = 0.7105, MSE = 0.0473), PWLEH achieved high accuracy in predicting peanut water content (R² = 0.7650, MSE = 0.0519). Unlike pure data-driven approaches, the new hybrid model incorporated radiative transfer knowledge and obtained higher predictive performance with fewer field data. This study demonstrates the potential of applying an optimized PROSAIL, hyperparameter search and modular training to improve the accuracy and transferability of the EWT prediction model, providing a new approach for sustainable agricultural management.

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.