Challenges and Solutions for using Machine Learning with Earth Observation Data

The application of artificial intelligence (AI) and machine learning (ML) is rapidly growing across Earth Observation (EO). State-of-the-art ML techniques can be used to analyse and exploit vast quantities of data, to produce greater insight into data, and enable better understanding of environmental issues. Over the past year, NEODAAS have been utilising their MAGEO GPU cluster to work with users on a range of projects including harmful algal bloom detection, tree monitoring, monitoring global mangroves, ocean oil-spill detection, road vehicle and ship exhaust tracking and underwater image maerl detection. This range of projects has provided opportunities and insights into common challenges in ML with EO data, and how best to overcome them.

To leverage the benefits of ML, there are several technical challenges that need addressing:

• Data access – Finding and accessing the most suitable data, how the data are structured, and what are appropriate data for ML approaches
• Problem framing and ML model selection methods – working out the most appropriate way to frame the problem to suit the available ML approaches.
• Access to computational resources – making sure GPUs are available and libraries are set up to utilise them.
• Data pipelines – how to ingest data into the model for training and then generate outputs.
• Output evaluations – including expert interpretations, evaluation metrics, error measurements and ML model generalisability

In this talk, we will summarise these challenges and discuss our solutions to overcome these challenges, with a focus on how the EO community can benefit from the challenges faced.