Data imputation in large and small-scale spatiotemporal time series gaps using BackForward Bi-LSTM

Abstract In recent years, tracking systems have transformed the understanding of spatiotemporal dynamic processes. However, these systems often face challenges due to missing data caused by technical limitations or intentional manipulation, leading to classification bias and hidden suspicious activities. In response, we propose a deep learning algorithm based on sequential Bidirectional Long Short-Term Memory (BiLSTM) to impute missing trajectories in spatiotemporal datasets. Our approach employs both backward and forward BiLSTMs resulting in BF-BiLSTM architecture to model temporal dependencies before and after the missing trajectories. We evaluated the algorithm on a realistic marine and aerial dataset, considering missing data trajectories on vessels and flights with missing data rates ranging from 5% to 30%. The model improves performance in terms of MAE (0.011°), MSE (0.056%), MAPE (0.054%) and ADE (0.017°) when compared to state-of-the-art approaches. By effectively addressing challenges in spatiotemporal datasets and improving existing benchmarks, our algorithm provides a robust solution for enhancing trajectory imputation in the context of monitoring systems potentially across diverse application domains.