Convolutional Denoising Autoencoder for Temperature Time Series Reconstruction in Agricultural IoT

This paper proposes a reconstruction methodology based on a one-dimensional Convolutional Denoising Autoencoder (1D-CDAE) for recovering missing air temperature values in agricultural IoT time-series data affected by extended transmission losses. The approach is assessed using real-world microclimatic measurements acquired from a wireless sensing node operating within a vineyard monitoring system. Both single-variable and multivariate reconstruction schemes are investigated. While the single-variable formulation relies exclusively on historical temperature data, the multivariate approach leverages relative humidity through a dual-encoder architecture combined with a cross-attention fusion mechanism, allowing the selective exploitation of physically correlated environmental information to enhance temperature reconstruction, particularly under extended missing intervals. Experimental results on real vineyard data show that the proposed multivariate 1D-CDAE consistently outperforms the temperature-only configuration, with particularly significant accuracy improvements for long missing-data segments, highlighting its effectiveness as a robust preprocessing tool for agricultural IoT decision-support applications.