IPSTERS - IPSentinel Terrestrial Enhanced Recognition System
Published:
This project focused on the exploration of several machine learning (ML) techniques, covering different stages of a Land Use/Land Cover Classification (LULC) pipeline. These techniques aimed to minimise problems typically found in this kind of data, namely data ingestion, feature selection, data filtering and classification. The results shown here are the result of a joint effort. Manvel Khudinyan developed all active learning experiments, which I converted into a Command-line Interface, in order to facilitate its use by the remote sensing specialists at Direção Geral do Território.
Reports
Github repositories
- Remote Sensing (contains most of the code and experiments)
- Data Filtering (contains initial anomaly detection experiments)
- Active Learning (contains Manvel’s AL experiments and the CLI)
Introduction
The Copernicus programme is the European Union (EU) Earth Observation (EO) programme, headed by the European Space Agency, and the developer of the Sentinels EO satellites. The IPSentinel is the Portuguese infrastructure developed by Direção Geral do Território (DGT) and Instituto Português do Mar e da Atmosfera (IPMA) for storing and providing images of Sentinel satellites, covering the Portuguese territory and its search and rescue area. This free EO data has been used to inform environmental models, business strategies and political decisions. However, this ever growing volume of data requires big data workload that is overwhelming for Public Administration (PA) agencies. As a result, the use of IPSentinel data has not been widely adopted by the PA that would profit from it.
Often what these agencies need for their goals are digested data in the form of specific class maps. These value-added products are often called level-3 products and are fundamental for land-management and for the country’s international commitments such as the estimation of CO2 emissions. These products are mainly obtained by visual interpretation of high resolution satellite imagery, requiring significant allocation of human resources from the PA and taking a long time to produce, being one of the reasons for the low update rate and low resolution. In the case of COS (Portuguese Land cover-land use maps) they are produced every 5 years with 1ha resolution and EU CORINE maps at least every 6 years with 25ha.
Purpose and Objectives
The main goal of this project is to explore the applications and limitations of artificial intelligence (AI) algorithms with accelerated processing hardware capabilities, as a unit of the IPSentinel for the digestion of large volumes of remotely sensed data, to produce level-3 products for land applications with the least amount of human intervention. We propose exploring two artificial intelligence approaches, one applying active learning techniques and another based on fuzzy logic.
Specifically, this report focuses on the exploration of techniques to improve the quality of LULC map outputs by the model under development. As such, topics such as Dimensionality Reduction, Anomaly Detection and Classification methods will be discussed.