Deep Data Imputation for UAV Low-Altitude Sensing Considering Spatial– Temporal Interaction

With the rapid development of unmanned aerial vehicle (UAV) technology, UAV low-altitude sensing has been gradually applied in multiple fields of data collection and monitoring. However, limited conditions in complex sensor environment lead to the serious deviation or partial loss of data collection. In order to build a complete database for UAV low-altitude sensing, the imputation method needs to be designed. Previous related work usually focused on separate spatial domain or single time series dimension, regardless of spatial–temporal coordination. In addition, traditional methods such as simple interpolation cannot meet the requirements of special monitoring environment. In this article, we propose a hybrid deep learning model to achieve spatial–temporal data imputation for UAV low-altitude sensing. Specially, the proposed method can be divided into following parts: 3-D dilated convolution (3D-D-CNN), squeeze with excitation (SE), optimization considering impact of missing data, the targeted bidirectional long short-term memory (T-Bi-LSTM), and the final fusion module. In addition, we improve the proposed overall architecture by using mathematical methods. On the one hand, we conduct the data preprocessing via ${Z}$ -score normalization before deep model input. On the other hand, we also take regularized regression based on sparsity and low rank to improve the efficiency. Experiments show that the proposed hybrid deep learning model can effectively finish UAV low-altitude sensing data imputation, and shows its superiority in the development of UAV acquisition and monitoring.