Large-Scale Maize Condition Mapping to Support Agricultural Risk Management

Crop condition mapping and yield loss detection are highly relevant scientific fields due to their economic importance. Here, we report a new, robust six-category crop condition mapping methodology based on five vegetation indices (VIs) using Sentinel-2 imagery at a 10 m spatial resolution. We focused on maize, the most drought-affected crop in the Carpathian Basin, using three selected years of data (2017, 2022, and 2023). Our methodology was validated at two different spatial scales against independent reference data. At the parcel level, we used harvester-derived precision yield data from six maize parcels. The agreement between the yield category maps and those predicted from the crop condition time series by our Random Forest model was 84.56%, while the F1 score was 0.74 with a two-category yield map. Using a six-category yield map, the accuracy decreased to 48.57%, while the F1 score was 0.42. The parcel-level analysis corroborates the applicability of the method on large scales. Country-level validation was conducted for the six-category crop condition map against official county-scale census data. The proportion of areas with the best and worst crop condition categories in July explained 64% and 77% of the crop yield variability at the county level, respectively. We found that the inclusion of the year 2022 (associated with a severe drought event) was important, as it represented a strong baseline for the scaling. The study's novelty is also supported by the inclusion of damage claims from the Hungarian Agricultural Risk Management System (ARMS). The crop condition map was compared with these claims, with further quantitative analysis confirming the method's applicability. This method offers a cost-effective solution for assessing damage claims and can provide early yield loss estimates using only remote sensing data.