This research investigates the rate of spectral, polarimetric and vertical structure indices derived from freely available satellite data by utilizing ensemble machine learning and Explainable AI (XAI) techniques. Accordingly, by employing Google Earth Engine, the effective indices were extracted for warm and cold seasons from Sentinel-1, Sentinel-2, and tree Canopy Height Models (CHM). Then various ensemble tree based models including Random Forest, Extra Trees, AdaBoost, XGBoost and LightGBM were employed to classify tree species at various levels (L0 to L3) based on the collected dataset. The result demonstrated that LightGBM obtained the best performance across all four levels. At the higher levels (L2 and L3), the classification become more complex by increasing the number of tree species and intra-class variation. However, the model could successfully identify 20 similar tree species at the L3 level with a satisfactory F1-score of 67.2%, using only free satellite data and sensor fusion techniques. By adding spectral and vertical structure indices to polarimetric indices, the F1-score at the L2 level improved by 26.36% . This study also investigated the effects of seasonal variations on classification. The results indicated that evergreen tree species, due to their higher tolerance thresholds and stable vegetation characteristics, exhibited similar spectral and polaremetric pattern in both warm and cold periods, resulting in consistent accuracy. In contrast, deciduous species showed more distinctive spectral reflectance in the warm period due to physiological changes. Also, considering the diverse range of indices, the model's behavior was systematically analyzed using Feature Importance and SHapley Additive exPlanations (SHAP) across three levels: (1) exploring the contributions of different sensors (Sentinel-1, Sentinel-2, CHM) to classification performance, (2) understanding how feature importance varied between different tree species groups, and (3) assessing how changes in individual feature values influenced the model's output. Based on feature importance analysis, the indices DPSVI, DPSVIo, Cross Ratio, SWIR2/SWIR1, and SWIR2/Red ranked higher than others. From the perspective of SHAP analysis, the indices MIN, DPSVIo, and Cross Ratio were identified as the most influential features. Finally the implementation of proposed seasonal fusion technique realized the generation of species classification maps at various level with desirable accuracy, facilitating regular forest monitoring.
