Improved mapping of perennial crop types based on intra-annual biophysical changing patterns of spectral endmembers

Perennial crops are vital to economic growth, environmental sustainability, and human well-being. However, due to the diversity and complexity of natural environments and agricultural management practices, there is currently no widely transferable mapping strategy for these crop types, particularly in regions with diverse perennial species. To address this gap, we propose a novel perennial crop mapping strategy based on intra-annual changing patterns of spectral endmembers (CPSEM). This strategy integrates a unified spectral endmember (EM) space with a harmonic model to characterize and quantify the biophysical processes and morphology of vegetation. Using Linear Spectral Mixture Analysis (LSMA), Sentinel-2 time-series data (2020−2022) were unmixed into a unified spectral EM space comprising green vegetation (GV), non-photosynthetic vegetation (NPV), soil (SL), and dark surfaces (DA), enabling the reconstruction of land surface component (LSC) trajectories at the pixel level. We developed two EM-based morphology indices to capture structural and compositional relationships among EMs. A harmonic model was applied to extract key parameters from the EM fractions and EM-based morphology indices, representing vegetation biophysical processes. Finally, a Random Forest model was used to classify perennial crop types. The results show that perennial crops of the same type exhibited similar biophysical processes and morphology, while distinct types exhibited substantial differences. Our method effectively maps perennial crop types across diverse environments and planting conditions, achieving classification accuracies of 87.27 %–90.91 %. Compared to traditional spectral-based methods, the proposed strategy improves perennial crop classification by 1.7 %–3.9 % and overall vegetation classification by 5.3 %–8.4 %. Additionally, this strategy effectively addressed the limitations inherent in traditional phenological indices for accurately classifying perennial crops, demonstrating robust performance even in complex classification scenarios. Incorporating synthetic aperture radar (SAR) features did not further improve classification accuracy. This strategy enhances interpretability and transferability through the use of a unified spectral EM space and detailed biophysical characterization. Thus, the CPSEM-based perennial crop mapping strategy provides a robust and scalable approach for accurately identifying perennial crops and land cover at large scales.