Field-scale sugarcane mapping in Thailand by fusing annual satellite embeddings with a global field-boundary model, cross-checked against mill weighbridge records

<jats:p>Sugarcane in Thailand is grown almost entirely by smallholders who deliver to mills under seasonal contracts, so the quantities that matter to planning are properties of individual fields rather than of pixels. Published Thai cane maps are per-pixel classifications, and global cane products perform markedly worse in Thailand than elsewhere because cane is confused with cassava. Whether the global field-delineation models released in the past two years can supply the missing parcel geometry for Thai smallholder cane has not been tested. We combined the AlphaEarth Satellite Embedding dataset with the Fields of The World (FTW) global boundary product to build a field-scale cane parcel map over a mill catchment of about 8,300 km² in central Thailand, and cross-checked it against 74,516 weighbridge delivery tickets covering 1.54 million tonnes from 8,924 known fields. Cane was classified per pixel from three years of 64-band annual embeddings using two-step positive-unlabelled learning (Liu et al., 2002; Elkan &amp;amp; Noto, 2008; Bekker &amp;amp; Davis, 2020); parcels were taken from FTW and labelled as cane when at least half of their area fell inside the predicted mask. The detector recalled 93.2% of held-out registered cane pixels at a 5.3% false-positive rate and fired on 97.0% of 1,500 fields already carrying cane before they entered the registry. The fused layer resolved 67,251 parcels of median size 8.0 rai against 0.6 rai for raw FTW parcels; 59.4% showed a harvest-scale NDVI collapse in both milling seasons. Visual interpretation of paired seasonal imagery over the 299 highest-ranked unregistered parcels returned a precision of 0.946 (95% CI 0.915–0.967); a stratified assessment of the full layer is in progress and no map-level accuracy is claimed until it is complete.</jats:p>