Spatio-Temporal Missing Data Reconstruction by Using Deep Neural Networks in Agricultural Monitoring Systems

This study addresses the issue of missing, distorted, or inaccurate data in agricultural monitoring systems. Such data issues can lead to errors in prediction and management models, affecting various applications in agricultural meteorology and remote sensing. Conventional missing data completion algorithms often fail to effectively leverage the inherent relationships between temporal and spatial data in agricultural observation systems. Machine learning techniques, specifically deep learning, offer a promising solution by considering factors such as time windows, seasons, and plant characteristics to fill missing data. However, managing the non-linear nature of agricultural monitoring within a machine learning framework poses a challenge. This study proposes a new deep learning approach called Predictive Error Compensated Network that addresses missing data reconstruction while mitigating overfitting. Predictive Error Compensated Network utilizes feature extraction networks and Discrete Wavelet Transform to incorporate different data types and time windows, improving performance. Evaluation against traditional methods demonstrated superior results with Predictive Error Compensated Network, achieving a significant reduction in reconstruction Root Mean Squared Error across different time windows.