The Application of Machine Learning Techniques to EO Data Quality Control

Space Systems Consultant, Kevin Halsall, explains how Telespazio UK’s Ease QC initiative aims to employ Artificial Intelligence techniques in the field of Earth Observation (EO) data Quality Control (QC).

As one of the foremost companies in the field of EO data quality control in Europe, Telespazio UK is looking to employ Deep Learning to address the dramatic increase in data volumes over recent years and support organisations such as the European Space Agency (ESA) to help ensure that the critical activity of assuring the fitness for purpose of that data is able to keep pace.

To that end, as part of the Ease QC initiative, a number of development activities have been instigated to explore the potential use of Deep Learning techniques in this field:

  • A common toolset to support both the QC engineers and Machine Learning model developers, in order to enable the efficient definition of training datasets and an interface with which to view the results of the application of trained machine learning models to the data
  • A number of proof of concept machine learning models focused on both the detection of specific anomalies, as well as general anomaly detection within EO optical data
  • A highly configurable cloud based micro-service-based infrastructure supporting the deployment of the developed machine learning models to the data for assessment and collation of the results

Common Toolset

The ‘Quality Control Optical Learning Tool’ was developed to support the often effort-intensive activity of assembling suitable training data for machine learning model development. The tool’s aim is to closely integrate the activities of the QC engineers in the generation of the quality assessments, and the machine learning modellers’ focus on collating and curating a training dataset.

Quality Control information and associated metadata are stored in a database for each product in such a way to facilitate the filtering and assessment required for the assembly of training data. The output from the machine learning models is then also stored within the tool, enabling the QC engineers to accurately assess the results and inform the machine learning developers on each model’s effectiveness.

The philosophy behind the tool is that of closely embedding the machine learning process within that of the existing EO data QC process, rather than treat it as a subsequent or separate part of the chain.

Machine Learning Models

To date, several proof of concept models have been developed to determine the effectiveness of applying machine learning techniques to QC EO data assessments.  

Initially focused on Landsat optical data, a supervised model based on a convolutional neural network model was developed, targeting a specific known anomaly in the data that was difficult to detect through alternate methods. Following an iterative training and development process, the final model was shown to be very effective at detecting the selected anomaly over a large dataset.

Current development activities are now focusing on a semi-supervised model type that should theoretically be able to detect the existence of any type of anomalous feature in the data.