Landscape Analysis for Multi-Objective Hardware-Aware Neural Architecture Search in Earth Observation Applications

Abstract

Recent advancements in the field of AutoML, particularly in Neural Architecture Search (NAS), offer exciting possibilities for both researchers and industry practitioners by automating the design of neural networks. While many NAS studies have focused on well-established datasets, Earth Observation (EO) datasets introduce unique challenges and complexities distinct from those datasets. This study’s primary objective is to establish a dedicated benchmark dataset tailored specifically for EO data and to conduct a landscape analysis on the created benchmark dataset. This dataset will be valuable for the EO researchers who wish to apply NAS methods in their research. Within the NAS framework, Hardware Aware Neural Architecture Search (HW-NAS) holds particular significance as it enables the optimization of network designs tailored to specific hardware configurations, especially in resource-constrained settings. Considering the possible needs of HW-NAS on EO domain, our work also includes hardware-dependent metrics in the benchmark dataset. Additionally, we explore the potential of developing compact surrogate models for data-centric initialization in HW-NAS, thereby trying to enhance the efficiency of the architecture search process. This research not only addresses the unique demands posed by EO data in the context of NAS but also holds a promise for applications spanning various domains.