Robust hyperspectral retrieval of mangrove functional traits across species and spatial scales using GA-ELM

Plant functional traits provide critical indicators of ecosystem functioning, biodiversity maintenance, and plant adaptive strategies under environmental change. However, hyperspectral retrieval of mangrove functional traits remains challenging because trait-related spectral signals are strongly influenced by species composition, canopy structural heterogeneity, and intertidal background effects, leading to unstable trait-spectrum relationships across species and scales. To address these limitations, this study developed a genetic algorithm-optimized extreme learning machine (GA-ELM) framework for robust retrieval of mangrove functional traits under heterogeneous mangrove spectral conditions. Three key traits, including chlorophyll content (Cab), equivalent water thickness (EWT), and leaf mass per area (LMA), were retrieved for four dominant mangrove species using leaf-level hyperspectral measurements and canopy-scale GF-5 imagery. GA-ELM generally outperformed PLSR, RF, and conventional ELM in species-specific comparisons. Leave-one-species-out validation yielded R 2 ranges of 0.61–0.75 for Cab, 0.60–0.72 for EWT, and 0.66–0.76 for LMA, indicating moderate predictive capability for previously unseen species, with performance varying among species and traits. At the canopy scale, GA-ELM achieved stable estimation accuracy for Cab, EWT, and LMA based on GF-5 imagery ( R 2 > 0.75), demonstrating that the framework can be independently calibrated for canopy-scale retrieval. Independent cross-regional validation yielded R 2 values of 0.60, 0.49, and 0.44 for Cab, EWT, and LMA, respectively, indicating moderate but reduced cross-regional transferability. Spectral sensitivity and band-selection analyses revealed species- and trait-dependent patterns that provided context for interpreting variation in predictive performance. These results remain conditional on one GF-5 acquisition per site and 30 m mixed-pixel conditions, and broader tidal-stage and cross-sensor validation is still required. These findings demonstrate the potential of GA-ELM for hyperspectral retrieval and spatial mapping of mangrove functional traits.