Generative AI-Driven Anomaly Detection in Soil Electrical Conductivity Using Temporal Autoencoders
Book Chapter
Etar, A, Soni, J, Upadhyay, H et al. (2027). Generative AI-Driven Anomaly Detection in Soil Electrical Conductivity Using Temporal Autoencoders
. 685 LNICST 23-32. 10.1007/978-3-032-22542-9_2
Etar, A, Soni, J, Upadhyay, H et al. (2027). Generative AI-Driven Anomaly Detection in Soil Electrical Conductivity Using Temporal Autoencoders
. 685 LNICST 23-32. 10.1007/978-3-032-22542-9_2
Environmental monitoring datasets such as apparent electrical conductivity (ECa) maps are critical for applications like irrigation and salinity management. Autoencoders have proven useful for learning the distribution of normal sensor data and detecting anomalies by reconstruction errors. In this study, we train four Autoencoder models which includes a fully connected Linear Autoencoder, a temporal feedforward Time Linear Autoencoder, a Temporal Convolutional Autoencoder and a Convolutional Autoencoder on the 80% of the apparent soil electrical conductivity data from California farmland (1991–2017). We then use a Variational Autoencoder (Anomaly-VAE) trained on the normal data distribution to inject synthetic anomalies into the remaining test set. For each model, we compute the reconstruction error (delta) on the anomalous test set. We evaluate model performance using two metrics, the coefficient of determination (R2) and the fraction of errors exceeding the 90th percentile. Our results show that all models effectively learn the underlying ECa distribution but the Time Linear Autoencoder achieves the best accuracy under anomaly injection whereas standard Linear Autoencoder and Convolutional Autoencoder show higher error (R2). These findings demonstrate the utility of incorporating temporal information in autoencoder architectures for anomaly detection in soil conductivity data.
