Abstract
Coastal sand dunes serve as critical protective barriers and ecological systems, yet the escalating intensity of tropical cyclones in regions such as Southwest Florida necessitates large-scale, costly restoration initiatives. Despite these investments, traditional field-based monitoring of expansive dune systems is frequently hampered by logistical complexity and limited spatial coverage, which can underrate evaluations of restoration success. Remote sensing offers a powerful, cost-effective alternative to traditional surveys, providing repeatable observations capable of capturing vegetation dynamics across large coastal areas. However, the effective monitoring of localized restoration efforts presents a "resolution problem," as standard satellite platforms (e.g., 10–30 m) lack the spatial detail required to characteristically resolve the sub-meter heterogeneity of young or fragmented vegetation establishment. This study systematically evaluates the comparative efficacy of Landsat 8/9, Sentinel-2, PlanetScope, and centimeter-scale multispectral Unmanned Aerial Vehicle (UAV) sensors for monitoring dune recovery across three sites in Collier County, Florida. Ground-truth vegetation cover was derived using a validated U-Net deep learning model (Mean Accuracy = 91.8%), and twenty-three vegetation indices were tested using hierarchical Linear Mixed-Effects Models to isolate independent predictive power across various spatial and temporal scales. The UAV-derived Green Normalized Difference Vegetation Index (GNDVI) emerged as the optimal metric, capturing the highest total variance (R2c = 0.621) and minimizing predictive error (RMSE = 21.43%). Competitive model evaluation identified a non-linear Logistic (Sigmoid) functional form as the superior mathematical fit (AIC = 16,133.76), providing a mechanical explanation for the predictive failure of coarser satellite platforms. These platforms become mathematically "trapped" in a spectral lag phase where high sand albedo masks early-stage establishment. Furthermore, statistically optimized thresholds were established at GNDVI = 0.122 to exclude non-vegetated surfaces and GNDVI = 0.308 to isolate actively photosynthetic green canopy. Longitudinal analysis successfully captured intra-seasonal phenological shifts, allowing for a nuanced distinction between healthy winter dormancy and potential mortality. By bridging the scale gap between satellite and UAV data, this research provides coastal managers with a standardized, automated framework for rapid post-storm assessment and the long-term evaluation of geomorphic and ecological resilience.
