An OGC TrainingDML-AI approach for making EO training datasets ready in deep learning frameworks

Training data (TD) are vital for artificial intelligence deep learning or machine learning (AI DL/ML) in remote sensing image interpretation. With the proliferation of a large number of Earth Observation (EO) datasets, the availability of a wide range of datasets has introduced challenges in ensuring the FAIR (Findable, Accessible, Interoperable, Reusable) use of training datasets. This paper proposes an approach that leverages the OGC Training Data Markup Language for AI (TrainingDML-AI) standard to make training data ready to be consumable by existing DL frameworks. It presents a training data pipeline approach to integrate TD in DL. The approach enables the retrieval and transformation of training data for compatibility with existing deep learning frameworks.

I. Introduction

Training data (TD) play a vital role in data-driven approaches for remote sensing image interpretation using artificial intelligence (AI) machine learning (ML) and deep learning (DL). As a result, substantial efforts have been made by EO experts, research teams, and organizations, to generate massive training datasets for various image interpretation tasks [1–6], including scene classification, object detection, semantic segmentation, change detection and 3D model reconstruction. These publicly available training datasets are valuable assets for AI researches and applications in Earth observation (EO) domain. Although many datasets are provided to the public in an open source manner, the representation and organization of these data are different, which lead to heterogeneity and difficulty for dataset sharing and interoperability [7–8].