Scalable Deep Learning for Remote Sensing with High Performance Computing

Advances in remote sensing (RS) missions in recent decades have greatly increased the volume of data that is continually acquired and made available to end users, who can utilize it in a variety of Earth observation (EO) applications. land cover (LC) maps play a key role in monitoring the Earth’s surface, providing scientists and policymakers with an accurate view of the evolution of the landscape and helping them address pressing questions, from efficient resource planning to resilience to climate change. Due to the use of classical machine learning (ML) and more recently of deep learning (DL) methods, the information content of RS data can be exploited to an unprecedented degree, fostering research, development, and deployment of workloads to address open challenges for EO applications, including LC classification. However, the larger size of the datasets needed to train state-of-the-art (SotA) DL models and the need to utilise them at scale increases the time to deployment, which can hinder their effective utilization. Adopting strategies for distributed deep learning (DDL) on high performance computing (HPC) systems provides the opportunity to speed up the training of the models, allowing faster development times for researchers. Since space agencies operate a variety of missions, data acquired by different sensors can be used to increase the temporal resolution at which a certain area is observed, with potential improvements in the accuracy of the ML/DL models. The thesis objectives are formulated with these premises in mind and were investigated using a combination of methodologies to exploit the dedicated resources of HPC systems, contributing to addressing new questions on the adoption of DDL methods for EO applications and to familiarize the RS community with such approaches, which can be of great value.

‍

View the recording PhD defence

‍