Large-area cotton mapping and yield forecasting via Semi-supervised temporal-spatial-spectral method (TSpaSpe) using Sentinel-2 data

Mastering information on cotton yield in large areas has an irreplaceable role in agricultural production and balancing the relationship between supply and demand. Owing to temporal and spatial phenological differences and spectral inconsistencies in large areas, cotton information extraction poses additional challenges. Traditional prediction methods only use a single piece of information and rely on a large number of samples, which are unsuitable for large-scale cotton extraction research because of their poor generalization ability and limited sample availability in large areas. To address these challenges, this study introduces a semi-supervised temporal-spatial-spectral method (TSpaSpe) that leverages Sentinel-2 data. This method integrates temporal-spatial-spectral features along with phenological information to accentuate the distinctions between cotton and other crops and weaken the model dependence on samples. The TSpaSpe method was tested in three renowned high-yield cotton-producing regions (i.e., Aksu in Xinjiang, China; Irving in Texas, the United States; and Querencia, Mato Grosso State, Brazil), and sample datasets were generated using Google Earth high-resolution images. Compared with eight other prediction methods derived from various data configurations, the two methods grounded in the TSpaSpe framework demonstrated superior performance in distinguishing cotton from other crops. The application of the proposed methods to cotton extraction using Sentinel-2 images across the three study areas revealed that the TSpaSpe_CNNDW approach achieved the highest average R2 and overall performance relative to the alternative extraction techniques. Further validation confirmed that TSpaSpe_CNNDW outperformed the other methods in predicting cotton yield across different years within the three regions. These findings suggest that TSpaSpe_CNNDW can effectively extract spatiotemporal distribution patterns and growth information for cotton in diverse environments. Moreover, this technique shows great promise for advancing the field of large-area cotton yield prediction, offering a robust tool for agricultural planning and policy making.