Parcel-level crop classification in small and irregular fields using structured Sentinel-2 data and clustering based analysis

Abstract Satellite image time-series (SITS) is one of the most useful tools for crop type classification. However, the low resolution of satellite image is a challenge for small, irregular shaped parcels such as farms in South Korea. In this study, we conducted a comprehensive evaluation of a parcel-level crop classification framework based on structured Sentinel-2 SITS data. The study was conducted across four regions in South Korea (Naju-si, Yeongcheon-si, Uiseong-gun, and Namhae-gun), covering 20 different crop types. The data was derived from 6-month Sentinel-2 imagery and drone-based ground truth surveys. For preprocessing, the satellite imagery was structured into a 2D matrix (Time × Bands), representing the 10-day average reflectance across spectral bands.. Ten machine-learning and deep-learning models were compared, showing that XGBoost achieved the highest performance with an Micro F1 of 0.7175, Macro F1 of 0.5474 and Kappa of 0.6746, while CBAM model recorded the second highest performance with an Micro F1 of 0.6892, Macro F1 of 0.5512 and Kappa of 0.6470, suggesting that decision tree-based models have advantage with tabular data. A post-hoc SHAP analysis based on XGBoost suggested that the cultivation characteristics and growing seasons of the crops were well-reflected in the spectro-temporal features. By analyzing the similarity between crops using K-means + + clustering, we could identify their spectro-temporal similarities, which effectively explained the reasons for low accuracy for specific crops in the classification models. This research is expected to be applicable in the future for nationwide crop mapping, supporting agricultural policy-making and decision-making.