Foliar wetness prediction using sensor network data and WaveNet-based deep learning models

Environmental and meteorological data are often collected as time series from distributed sensor networks. Leveraging inter-parameters correlations can also minimize sensor deployment by inferring information from physical quantities not actually measured. This study introduces a novel data reconstruction approach based on convolutional neural networks (CNNs). The model architecture, namely a set of WaveNet, is applied to real-world data collected from five sensor nodes in a vineyard. The primary goal is to reconstruct binarized foliar wetness using temperature and relative humidity measurements, effectively classifying the leaf wet or dry status. Preprocessing involved handling missing data through linear regression-based imputation. The WaveNets, individually optimized for each node, demonstrated strong classification performance, with the best node achieving accuracy, precision, recall and F1-score higher than 0.88, demonstrating how deep learning techniques can be applied for reducing sensor density.