Assessing local and transferable spectral endmembers for disturbance detection in Pinus sylvestris L. forests across Spain

Differences in forest structure, species composition, and climate across biogeographical regions challenge the reliability of spectral models for forest disturbance monitoring. This study assessed whether region-specific spectral endmembers are required to detect silvicultural interventions in Pinus sylvestris forests across Spain. Sentinel-2 spectral signatures were characterized across 18 Provenance Regions and forest development stages (young, pole, and mature stands). PRs were grouped by spectral similarity using K-means clustering, and local, transferred, and global endmember sets were incorporated into the Fusion Near Real Time (FNRT) algorithm through Linear Spectral Mixture Analysis (LMSA) and the Normalized Difference Fraction Index (NDFI). Detection performance was evaluated using 70 reference polygons representing clear-cutting and thinning intervention. Spectral analysis identified three main spectral-ecological clusters: Atlantic, Continental, and Mediterranean. Detection accuracy differed significantly among clusters, reaching 86.63 ± 15.09% in the Continental, 78.63 ± 20.87% in the Atlantic, and 70.23 ± 17.77% in the Mediterranean cluster. Silvicultural treatment also significantly affected performance, with clear-cutting achieving higher accuracy (89.60 ± 10.76%) than high- (73.02 ± 20.22%) and low-intensity thinning (72.74 ± 19.95%). In contrast, endmember origin had no significant effect on detection accuracy (p = 0.617): local endmembers achieved 80.22 ± 18.47%, compared with 78.44 ± 19.04% for global and 76.82 ± 20.22% for transferred endmembers. These results show that, despite regional spectral variability, global and transferred endmembers maintain comparable detection performance, supporting their use for large-scale operational forest monitoring while reducing the need for region-specific calibration.