Delineation of summer crop areas using Sentinel-2 imagery derived indices and georeferenced data collected through a Volunteered Geographic Information approach

Accurate delineation of cropped areas is crucial for efficient agriculture, food security, and policy development. Numerous global crop-mapping initiatives provide valuable insights into cropping activities worldwide. However, the spatiotemporal resolutions of these maps constrain their usefulness at the local level. Furthermore, map accuracy relies heavily on reference data, which are often prone to inconsistency and inaccuracies due to various data collection procedures. Publicly available satellite imagery, Volunteered Geographic Information (VGI)-based reference data collection, and machine learning (ML) approaches enable crop classification at the sub-national scale. However, their integration approaches are lacking. This study examined the potential of VGI reference data and Spectral Indices (SIs) derived from Sentinel-2 imagery to delineate summer crop areas using a Random Forest classifier. The classifier utilised 800 reference data samples collected by farmers and volunteers using the Open Data Kit (ODK) on mobile devices. Classification accuracy of 89% and f1-scores of 89%, 88%, 80%, 96%, and 89%, respectively, for Built-up Area/Bare Land, Paddy, Sugarcane, Tree Cover, and Water Body were obtained. The proposed method demonstrates promising results in smallholder farm settings and is replicable and cost-effective for sub-national level mapping in low-income countries.