Peach Orchard Identification Using Multi-Temporal Sentinel-2 NIR–SWIR Features: A Case Study of Ganyu District, Lianyungang, China

In complex agroforestry mosaics, similarities in canopy structure and spectral response among tree species constrain medium-resolution peach orchard identification. Using 2023 multi-temporal Sentinel-2 imagery from Ganyu District, Lianyungang, Jiangsu, we assessed NIR–SWIR features, vegetation indices, and phenological parameters for classifying peach orchards, poplar, pine, and willow. First, spectral curves, Jeffries–Matusita (JM) distances, and bootstrap confidence intervals compared separability across leaf-off, early-spring, and leaf-on periods. Subsequently, HANTS reconstructed the time series from which SOS and EOS were extracted; multi-temporal bands, MSI, NDVI, EVI, and NDRE were used in Random Forest schemes. Furthermore, McNemar tests, repeated random splits, and permutation feature importance assessed differences, stability, and feature contributions; the 2023 model was transferred to 2025 imagery. Results showed phenological dependence, with leaf-off and early-spring performance exceeding leaf-on. MSI and NDII showed similar separability and performance, indicating effects from shared B8 and B11 information rather than index formulation. The optimal scheme yielded OA = 94.545%, Kappa = 0.924, peach orchard PA = 97.619%, and UA = 91.111%. Ten splits gave OA = 92.83% ± 2.08% and peach orchard F1-score = 94.34% ± 2.27%. At fixed validation locations, cross-year transfer achieved OA = 90.184% and Kappa = 0.863, indicating temporal transfer potential within the forestland and orchard mask.