Predicting stomatal conductance of chili peppers using TPE-optimized LightGBM and SHAP feature analysis based on UAVs’ hyperspectral, thermal infrared imagery, and meteorological data

Abstract

Effective water management is crucial for ensuring the healthy growth and high yield of crops, and it relies on accurate monitoring of plant water status. As a core indicator of plant gas exchange capacity, stomatal conductance (Gs) directly determines the efficiency of photosynthesis and transpiration, significantly impacting crop growth and yield formation. Therefore, timely and accurate prediction of stomatal conductance is essential for optimizing water management strategies and improving crop yield and quality. However, stomatal conductance is influenced by a variety of environmental factors and plant physiological traits, making its variability complex and dynamic. These challenges result in difficulties in selecting key features, insufficient prediction accuracy, and a lack of transparency in model decision-making processes. To address these issues, this study proposes a novel approach that combines a random forest (RF) feature selection method with a Tree-structured Parzen Estimator (TPE)-optimized light gradient boosting machine (LightGBM) model (TPE-LightGBM). This approach leverages UAV-based hyperspectral, thermal infrared imagery, and meteorological data to improve predictive performance. Additionally, SHAP (SHapley Additive exPlanations) analysis is incorporated to offer insights into the model’s decision-making process by revealing feature dependencies. In the feature selection process, we compared four common methods, including mutual information (MI), successive projection algorithm (SPA), recursive feature elimination (RFE), and least absolute shrinkage and selection operator (LASSO) to ensure the significance and effectiveness of the selected features. To comprehensively evaluate model performance, we also compared five predictive models: ridge regression (RR), partial least squares regression (PLSR), random forest regression (RFR), random search-optimized LightGBM (Random-LightGBM), and grid search-optimized LightGBM (Grid-LightGBM). The experimental results revealed that the combination of RF and TPE-optimized LightGBM significantly outperformed all other models, achieving the highest prediction accuracy. The optimal number of features was determined to be N = 15, with a coefficient of determination (R²) of 0.862, a root mean square error (RMSE) of 0.037, and a mean absolute error (MAE) of 0.029. Through SHAP analysis, the study not only identifies key influencing factors such as photosynthetically active radiation (PAR), canopy temperature (CT), and red-edge spectral bands, but also sheds light on how these factors interact with each other to influence stomatal conductance. The proposed model provides an innovative approach to effectively predicting stomatal conductance, enabling agricultural managers to better understand and regulate chili pepper’s water status, thereby promoting healthy chili pepper growth and efficient resource management.

Plant phenotyping relevance

UAVのハイパースペクトル・熱赤外画像からチリ pepper の気孔コンダクタンスという生理形質を推定するモデルを開発・比較・解釈しており、表現型取得・推定手法が研究の中心です。

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.