Towards efficient global grape mapping: a two-stage spatio-temporal filtering framework for training sample collection

Grapes are among the top four fruit crops with the largest planting area and the highest yield worldwide. Timely and accurate vineyard mapping is critical for cultivation management and yield estimation. However, the spectral similarity between grapevines and other vegetation types in satellite imagery poses significant challenges for the classification, compounded by the inherent difficulty of acquiring high-confidence training samples. In this study, we proposed a two-stage spatio-temporal filtering framework for training sample collection, starting with 5,000 potential vineyard samples generated from existing land cover datasets. In the first stage, temporal similarity screening was applied to reference samples using NDVI time-series profiles, and in the second stage, spatial pattern matching was applied using ultra-high-resolution imagery. Approximately 500 high-confidence samples were obtained, achieving over 88% accuracy when validated against ground truth and showing 24–87% greater diversity than field-collected samples. Being applied to the seven largest grape-producing countries, the framework achieved an overall accuracy of 77−84% using random forest classifier, with the highest accuracy was found in the USA and Chile, and the lowest accuracy was found in China due to the fragmented landscape. The classification results show excellent spatial agreement with the actual land cover distribution, though salt-and-pepper noise occurred in regions with smallholder cultivation. Area estimates deviated by 2.7−13.2% from statistics, primarily due to spectral confusion between vineyards and shrubland orchards. The framework of this study was proved useful for large-scale vineyard mapping and could support the land cover mapping of other land cover types and crop types, especially for places where training samples are insufficient.