A novel framework for developing accurate and explainable leaf nitrogen content estimation model for aquilaria sinensis seedlings using canopy RGB imagery

Abstract

Leaf nitrogen content (LNC) is crucial for the cultivation and health management of the endangered tree species Aquilaria sinensis. Although RGB imagery combined with machine learning has been effective for non-destructive LNC estimation, current models often neglect colour index texture features and face feature selection and interpretability challenges. This study introduces a framework to address these issues. Firstly, the canopy RGB imagery colour indices and the texture features of Aquilaria sinensis seedlings were collected as an initial feature set. Then, an improved hybrid feature selection algorithm combining SHapley Additive exPlanation (SHAP) with a dynamic ranking strategy was applied with a regression algorithm. This approach was tested using random forest (RF), support vector regression (SVR), and deep neural network (DNN) models. Optimal feature subsets were identified for each model, and performance comparisons determined the best LNC estimation model. Results show that texture features derived from colour indices significantly enhance LNC estimation accuracy. The dynamic SHAP ranking method outperformed RF and fixed SHAP rankings in feature selection. The optimal model, a DNN with an R² of 0.946 and RMSE of 1.859 g kg⁻¹ included two colour indices and five colour index texture features. While the normalised red colour index had the highest contribution, texture features contributed more overall to model accuracy. This method can be extended to other biophysical and biochemical parameter estimations.