A Dynamic Irrigation Event Identification Method Driven by Spatiotemporal Fusion and Dynamic Samples Using an Attention-Enhanced Bi-LSTM Model

Under the dual pressures of water scarcity and rising food demand, dynamic irrigation identification is essential for optimizing water allocation and improving agricultural water-use efficiency. However, constrained by the spatiotemporal resolution trade-off of single-source remote sensing data, limited capability in characterizing irrigation dynamics, and the difficulty of generating dynamic samples, current approaches remain insufficient for continuous and automated detection of irrigation events at fine scales. To address these limitations, this study proposes a dynamic irrigation identification method integrating spatiotemporal fusion and dynamic sample generation. First, daily 10 m re-modified perpendicular drought index (RPDI) time series are reconstructed using the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM), and combined with in situ soil moisture observations to characterize RPDI responses to irrigation events. Second, canal water release records and RPDI response features are integrated to generate dynamic irrigation samples with per-time-step labels. Finally, a multi-head attention-enhanced Bi-LSTM model is developed for automated identification of irrigation signals. Application in the Yongji Irrigation District showed that the proposed model outperformed LSTM, Bi-LSTM, and single-head attention-enhanced Bi-LSTM in both per-time-step irrigation response state identification and actual irrigation event identification, with average test-set OA and F1 values of 0.965 and 0.945. Independent validation showed no false-positive or false-negative actual irrigation events and accurate irrigation frequency retrieval. By contrast, although the baseline models achieved relatively high per-time-step accuracy, with OA values ranging from 0.887 to 0.960, they failed to accurately retrieve irrigation frequency, with OAfreq ranging from 0.211 to 0.868. These results demonstrate that the proposed method is capable of dynamically identifying irrigation events and effectively retrieving irrigation frequency information.