Weakly Supervised Phenology-Guided Refinement of WorldCover Cropland Maps in Hyperarid Regions

Reliable cropland maps are critical for agricultural monitoring, food security, and water resource management. However, global 10 m land cover products, such as the European Space Agency (ESA) WorldCover, tend to overestimate cropland in hyper‑arid regions because irrigated non‑cropland vegetation (e.g., urban greenery and amenity landscapes) is frequently misclassified as cropland. The objective of this study is to improve the reliability of the ESA WorldCover cropland product in hyper‑arid environments by developing and evaluating a phenology-guided refinement framework that specifically targets commission errors arising from irrigated non‑cropland vegetation. Using the United Arab Emirates as a case study, we derive multi‑temporal Sentinel‑2 vegetation indices and phenological metrics to automatically generate high‑confidence annual‑cropland and irrigated non‑cropland labels, train a Random Forest classifier, and apply hysteresis thresholding to produce a refined 10 m resolution cropland map. Validation confirms the classifier achieves balanced performance at the operational threshold (F1‑score = 0.93, overall accuracy = 0.93, intersection‑over‑union = 0.87). Independent evaluation with very‑high‑resolution imagery shows that 90–97% of pixels removed from the original WorldCover cropland class are correctly identified as non‑cropland, as captured by the False Positive Removal Accuracy (FPRA) metric. Emirate‑level comparison with official statistics indicates that the refinement removes more than 50 km² of false positives in intensive farming emirates, reducing area‑weighted commission error by 78% and lowering estimated irrigation water demand in Abu Dhabi by approximately 24 million m³ per year. The proposed framework provides a robust and transferable approach for refining global cropland products in hyper‑arid regions, supporting more realistic cropland statistics and water budgeting.