Seasonal variation monitoring of fluvial scene using PlanetScope data and machine learning method

Riverscapes vary across seasons as discharge, substrate exposure, and riparian phenology shift, making consistent, high-resolution monitoring difficult. We implemented a simple, transferable workflow to map seasonal fluvial scenes using 3m PlanetScope imagery and a pixel-level stacked ensemble with majority voting (PSEMV) implemented in Google Earth Engine. Seasonal PlanetScope composites (winter, spring, summer, autumn) were generated for 2021–2025 using fixed seasonal windows and median compositing to reduce residual noise while preserving sharp bar–water boundaries. Four classes were mapped (water, sediment, tree, and grass/shrub) using a two-source labeling strategy that combines expert-curated Dynamic World–derived high-confidence masks with an independent Autumn 2025 UAV/field survey. PSEMV trained RF, KNN, and Gradient Tree Boosting across deterministic seeds with repeated 60/20/20 splits, screened weak fits, and combined retained per-pixel predictions via majority vote; vote fractions provide an interpretable confidence surrogate. Across all seasonal scenes, the ensemble achieved the highest and most stable performance (mean accuracy ≈ 0.94), improving robustness relative to individual classifiers and delivering the largest benefits for challenging vegetation discrimination. Drone-based validation confirmed strong spatial agreement, with residual errors concentrated at mixed boundaries and deep shadows.