A Deep Learning Approach for Improving Soil Property Prediction with Unannotated Hyperspectral DESIS Imagery

An increasing number of hyperspectral satellite platforms are becoming available as exemplified by the DESIS platform.  This generates an immense amount of spectral information about the earths surface which is unlabelled.  In this work we use DESIS imagery to compare two deep learning approaches for utilising all of this unlabelled data for predicting topsoil properties. The first is transfer learning from laboratory based spectral libraries.  The second is a novel self-supervise learning approach which employs a transfer-based autoencoder architecture for unsupervised learning, analyzing spectra patterns and cultivating powerful latent representations for the downstream task of soil property analysis. Using a dataset from eastern Australia we show that the self-supervised learning approach gives superior predictions than transfer learning.