Highly transferable paddy field identification model based on SAR index and transformer

Artificial intelligence models are powerful tools for accurately identifying paddy fields from remote sensing data, which is crucial for effective agricultural management and decision-making. Due to historical and funding constraints, we often have very limited samples available, and samples’ features cannot fully cover all regions and time periods. This requires the corresponding model to have good transferable capabilities. However, building a transferable model also requires a large number of samples. This contradiction severely limits our ability to perform paddy field identification over large spatial and temporal ranges. To address these challenges, we propose a transferable index transformer based deep model for paddy field identification based on Sentinel-1 time series data (Transferability-Index). Transferability-Index constructs a novel structure that integrates existing SAR paddy field identification indexes into the neural network. This allows the model to inherit the “identification experience” contained in the index from the very beginning of its construction. This structure, combined with the transformer structure, can build a highly transferable model using only a small number of samples. In this study, experiments were conducted over a large area from 2018 to 2021, the training data only used samples from 2021. In comparison with seven traditional methods, for the 2021 test data, Transferability-Index achieved an overall accuracy of 92.50%, which is higher than other methods. For the 2018–2020 test data, Transferability-Index achieved 87.75%, 88.75% and 89.38%, which is significantly ahead of the compared methods. The high overall accuracy and stable performance of Transferability-Index demonstrate its strong transferable capabilities in paddy field identification. The Transferability-Index model can serve as a potent tool for paddy field mapping applications, especially when training data are scarce.