Forest decline poses a critical threat to the ecological function and long-term preservation of forest ecosystems. In dearth of comprehensive field measurements, remote sensing data support forest decline monitoring by offering consistent spatial and temporal observations. However, many approaches rely on costly data or lack structured feature management, limiting their applicability and performance. Freely available data like Sentinel-2 and Landsat, along with well-designed input features, enhance the efficiency and scalability of tree decline modeling. Here, we analyzed both a classified decline dataset (four discrete classes) and a continuous Phenological Decline Index (PDI) derived from Unmanned aerial vehicle (UAV) imagery, where hybrid feature selection and feature exteraction optimized inputs for the subsequent Random Forest (RF) regression and classification. The 9-year decline trend was predicted and smoothed using locally estimated scatterplot smoothing (LOESS) trend analysis, while trends of ERA5 climate data were assessed via the Sequential Mann-Kendall test. Results highlight the importance of feature selection even for non-parametric models like RF, improving R² from 0.40 to 0.60 and reducing RMSE from 0.13 to 0.10. Predicting PDI resulted in more consistent trends than the classification, revealing its effectiveness. Moreover, decline patterns proved highly complex and not directly aligned with climatic trends, which indicates that trees may have adapted to environmental stress. The study confirms the effectiveness of PDI methods and shows that tree decline patterns are only partially linked to climate variables.