Impact and evaluation of different temporal compositing strategies on the accuracy of tea plantation remote sensing mapping

Accurate mapping of tea plantations is essential for agricultural management and ecological monitoring. Although multi-temporal and phenological data are widely used, a systematic evaluation of feature efficacy under different temporal strategies remains lacking. To address this gap, this study integrated Sentinel-1 radar, Sentinel-2 optical imagery, and auxiliary environmental variables to rigorously compare three distinct temporal strategies: annual composite, phenological time-series, and multi-temporal stacking. These strategies were evaluated across both binary and multi-class Random Forest (RF) classification tasks to determine the optimal approach for high-precision extraction. The results indicate that feature importance is highly strategy-dependent. The multi-temporal approach excels by capturing instantaneous biophysical signatures such as Land Surface Temperature (LST) and Modified Normalized Difference Water Index (MNDWI) during critical growth stages, whereas annual and phenological strategies rely on broader metrics. Notably, the multi-temporal strategy demonstrated superior sample efficiency and the lowest decision uncertainty, maintaining robust performance even with limited training data. While multi-class models suffered from spectral confusion with orchards and forests, the multi-temporal binary strategy performed optimally, achieving an Overall Accuracy (OA) of 0.9740 and an F1-score of 0.9698. Area estimates from this configuration showed the highest consistency with official statistical bulletins. This research confirms that a multi-source, multi-temporal binary framework is highly effective for precise tea plantation extraction in topographically and climatically complex regions.