Joint Energy-based Modelling for Remote Sensing Image Processing

Abstract

The peta-scale, continuously increasing amount of publicly available remote sensing information forms an unprecedented archive of Earth observation data. Although advances in deep learning provide tools to exploit big amounts of digital information, most supervised methods rely on accurately annotated sets to train models. Access to large amounts of high-quality annotations proves costly due to the human labor involved. Such limitations have been studied in semi-supervised learning where unlabeled samples aid the generalization of models trained with limited amounts of labeled data. The Joint Energy-based Model (JEM) is a recent, physics-inspired approach simultaneously optimizing a supervised task along with a generative process to train a sampler approximating a data distribution. Although a promising formulation of such models, current JEM implementations are predominantly applied to classification tasks. Their potential improving semantic segmentation tasks remains locked. Our work investigates JEM training behavior from a conceptual perspective, studying mechanisms of loss function divergences that numerically destabilizes the model optimization. We explore three regularization terms imposed on energy values and optimization gradients to alleviate the training complexity. Our experiments indicate that the proposed regularization mitigates loss function divergences for remote sensing imagery classification. Regularization on energy values of real samples performed the best. Additionally, we present an extended definition of JEM for image segmentation, sJEM. In our experiments, the generation branch did not perform as expected. sJEM was unable to generate realistic remote-sensing-like samples. Correspondingly performance is biased for the sJEM segmentation branch. Initial model optimization runs demand additional research to stabilize the methodology given spatial auto-correlations in remote sensing multi-spectral imagery. Our insights pave the way for the design of follow-up research to advance sJEM for Earth observation.