Deep Learning-based Near-real-time Monitoring of Autumn Irrigation Extent at Sub-pixel Scale in a Large Irrigation District

Flood irrigation is widely applied in harvested croplands of arid regions and can be classified as autumn/winter irrigation (AI) depending on the time of application. Due to its high water consumption, real-time monitoring of the AI extent is crucial to improve its scheduling. We proposed a remote sensing-based long short-term memory (LSTM) model for near-real-time monitoring of AI extent at sub-pixel scale. The model loosely coupled MODIS data with Sentinel-2 data to solve the mixed pixel issue of MODIS data, and calibrated Sentinel-2 thresholds for AI identification by a random forest (RF) module to extract large-scale reference data with high temporal frequencies. The variable importance estimated by RF is used as a reference for feature screening in the LSTM model. As Sentinel-2 images are not available daily, LSTM models trained with incomplete sequences were validated using multiple validation approaches. The model was applied to the Hetao Irrigation District, the largest irrigation district in arid region of China, and delivered reasonable performance with coefficient of determination of over 0.82 and mean absolute error of around 10.7%. Classification of irrigation patterns using simulated time series of irrigation area fractions revealed eight different irrigation patterns from 2010 to 2020 in the study region. Results indicate that the maximum fraction of AI area upstream is closely related to the cropland distribution pattern. The AI patterns changed significantly over the 11 years, with a more pronounced reduction in irrigation duration for individual pixels in the downstream area. These changes are associated with two important land policies implemented in many regions of China, land consolidation and land transfer. The proposed model demonstrates the potential for near-real-time monitoring of autumn irrigation extent within large irrigation districts, which can aid in AI scheduling and provide insight into irrigation patterns and practices.