A joint learning Im-BiLSTM model for incomplete time-series Sentinel-2A data imputation and crop classification

• DOI: 10.1016/j.jag.2022.102762
• Publication Date: 2022-04-02T16:08:29Z
• Authors: Baili Chen; Hongwei Zheng; Lili Wang; Olaf Hellwich; Chunbo Chen; Liao Yang; Tie Liu; Geping Luo; Anming Bao; Xi Chen
• Keywords: Imputation (statistics); Interpretability

Multi-temporal deep learning approaches can make full use of crop growth patterns and phenological characteristics, resulting in excellent classification performance large areas. However, obtaining complete time-series remote sensing images during the growing season is challenging due to cloud contamination. Hence, given multispectral data, it important impute missing data accurately classify crops. A novel Imputation-BiLSTM model (Im-BiLSTM) was developed based on Bidirectional Long Short-term Memory network (BiLSTM) jointly perform imputation classification. The Im-BiLSTM regards as variables, which are efficiently updated backpropagation. treats interaction between tasks, reducing error uncertainty caused by separation operation Furthermore, we improved interpretability evaluating importance input features visualizing hidden state units. In Shawan County, Xinjiang, China, acquired a total 10 Sentinel-2A from April October 2016, 3 lost partial cover. applied incomplete containing time-steps for pixel-level classification, BiLSTM constructed cloud-free comparison. proposed tested four different cases rates. results showed that outperformed model, overall accuracy maximum 4.2%, F1-scores spring corn tomato 16.1% 21.4%, respectively. Therefore, effectively improve imputing data. (the coefficient determination values range 0.4 ∼ 0.9) indicated bands with larger contribution had higher accuracy. Feature evaluation captured key periods input. visualization units demonstrated accumulated useful information over time, learned high-level made crops more separable than original inputs.