A Comparative Analysis of Black-Box and Glass-Box Models for Poplar Plantation Mapping with Remote Sensing Data

Abstract

Poplar trees are essential for industrial afforestation applications due to their globally recognized plantation practices, reputation, ability to produce a large quantity of raw material in a short time, diverse applications in wood production, suitability for hybridisation and breeding implementations, and the availability of various species and clones adapted to the soil and climate conditions. Accurate identification and mapping of poplar afforestation areas are therefore crucial for planners and decision-makers to manage inventory records and maximize economic value. In this study, the poplar tree mapping and feature selection performances of the glass-box Explainable Boosting Machine (EBM) algorithm were investigated using satellite images having different resolutions (i.e., Sentinel-2 and PlanetScope) and texture features. A robust black-box algorithm, XGBoost, was utilized as a benchmark algorithm to compare the performance of EBM. The results showed that the EBM algorithm outperformed the standard XGBoost algorithm by up to 2% in classifying poplar trees when both spectral bands and calculated texture features were used, for both satellite images. Additionally, using the high spatial resolution PlanetScope imagery resulted in a significant decrease in the classification accuracy of popular areas (about %10) compared to Sentinel-2 imagery. The study also assessed the most important features influencing the classification process. For this, while 15 features were selected employing the visual charts provided by EBM for interpreting the decision-making process, the SHAP technique was applied to examine the most prominent features in the XGBoost model structure. In this scenario, EBM and XGBoost presented greater performances for both satellite data compared. These findings emphasize EBM's consistent superiority, indicating that its enhanced interpretability can facilitate more precise feature selection and model refinement, particularly for Sentinel-2 imagery. © 2025 Elsevier B.V., All rights reserved.