Cloud-based mapping of fragmented tobacco fields using multi-source remote sensing to support autonomous agricultural operations

Abstract. Traditional autonomous agricultural systems face significant challenges in performing continuous operations within fragmented field regions. To address this issue, it is essential to upgrade these systems to automatically acquire high-precision field boundaries. This study tests the hypothesis that fragmented tobacco parcels can be reliably mapped using a cloud-based, multi-source remote sensing framework and that the resulting products can directly support autonomous field operations. Using Xuchang City, Henan Province, China, as a case study, we developed a cloud-edge-integrated tobacco mapping workflow on the Google Earth Engine (GEE) platform by fusing Sentinel-2 optical imagery, Sentinel-1 synthetic-aperture radar data, and topographic variables. A comprehensive feature set, including spectral bands, vegetation indices, radar backscatter, texture metrics, and terrain attributes, was used to train and compare three machine learning classifiers: random forest (RF), gradient boosting decision tree (GBDT), and classification and regression tree (CART). RF achieved the highest performance, with an overall accuracy of 93.08 % and a kappa coefficient of 0.92, outperforming GBDT (90.60 %, 0.89) and CART (87.60 %, 0.85). The RF-derived tobacco planting area showed the closest agreement with official statistics, with a consistency ratio of 94.12 %. Model robustness was further demonstrated by direct transfer to the adjacent Pingdingshan City without re-training, yielding a 97.70 % consistency with reported acreage. By shifting field-boundary extraction from manual delineation to automated cloud-based processing, this study provides a scalable solution for mapping fragmented tobacco fields and delivering parcel-level geospatial data to autonomous agricultural systems, with broader applicability to other cash crops in fragmented landscapes.