Sentinel-2 Based Crop Class Monitoring of Small and Narrow Parcels Based on Modelling Pixel Composition

Earth observation (EO) and image classification have become essential tools for the enforcement and management of agriculture subsidies in the European Union (EU) under the Common Agricultural Policy (CAP). Crop classification is a fundamental approach to subsidy monitoring to validate the farmer’s claims. Similar to many areas in Europe, in Hungary a significant number of agricultural parcels are small in size due to the historical reason, ownership structure and landscape properties. This poses several challenges when monitoring with EO, primarily related to the spatial resolution of satellite imagery, the complexity of land use in smaller areas and the difficulty of distinguishing crop types on a finer scale. The Hungarian CAP monitoring showed that 8% of the claims had to be excluded and could not be monitored using Sentinel-2 data. For operational satellite-based monitoring in Hungary, the current limit was 800 m², similarly to what the European Commission’s Joint Research Centre recommended.

The aim of this project is to mitigate this issue by proposing modern machine learning (ML) methods using temporally dense Sentinel-2 image stacks to perform sub-pixel analysis and classification. This research develops, tests, and optimises algorithms for the classification of small and narrow parcels, followed by validation of the results against farmer claims and on-the-spot checks. The accuracy is assessed across different parcel sizes and crop types. The evaluation of compliance must provide an output at the parcel level in order to be comparable with the farmer’s claims. This includes developing and validating the aggregation of the crop probability to the level of individual parcels and quantifying the uncertainty of the results. Additionally, a benchmark study will be conducted to compare the developed approach with the current state-of-the-art, including the usage of higher-resolution commercial satellite imagery. The best-performing algorithm will be evaluated in an operational production context across the entire area of Hungary to support CAP monitoring.