Introduction and Objectives
The aim of the Product Feature Extraction (PFA) project is to demonstrate that feature extraction and analysis of EO data makes it possible to exploit the large volumes of various types of satellite data more efficiently, and through this to foster the scientific analysis of mission wide data sets. This project addresses emerging methods and tools for data product features and information extraction, in view of possible implementations for enriching data description and easing the use of archived data. For this purpose, the PFA team will elaborate and implement a number of EO data exploitation scenarios whose usefulness will be demonstrated by means of real-life applications.
PFA will address three main scenarios:
- Content based image retrieval (CBIR) including a novel active learning component based on a relevance feedback loop and including an efficient Support Vector Machine (SVM) classifier as search engine;
- Content-based time series retrieval (CBTR) which would reuse the basic CBIR system and apply it to stacks of co-located data. The team decided to focus on queries based on bi-temporal change analysis only;
- Unsupervised classification by Kernel k–Means for identifying specific spatial, temporal or spatio-temporal cluster in the data on the basis of similarities of features.
By paying great attention to the need to extraction of features that can be obtained on a large amounts of EO data and can be as well effective on different kinds of EO data the team decided to extract primitive image features which are generated from summary statistics yield from equal-size patches of EO raster data. These have been approved to work for SAR, optical and also time series of SAR and optical images.
The first two scenarios (CBIR and CBTR) have been considered being the most important and should therefore find their way into the final demonstration system. The team also concluded that a fourth scenario would be very useful and will be easy to implement, namely letting users query for data products by submitting SQL-like queries that include feature values or value ranges. Of course this only makes sense if a feature has a physical meaning that is fully understood by users, such as chlorophyll concentration.
Four main EO applications will be supported:
- Algal bloom detection from full mission MERIS L1b data
- Urban area mapping with ASAR data
- Urban change detection with SAR data
- Burnt area detection with SAR and optical data
The active learning component of the CBIR and CBTR is made available to clients by a PFA web service which has direct access to feature databases and takes as input image patches that have been interactively labeled by users as being relevant or irrelevant. By doing so, the web service trains and manages user-owned classifiers. Once a user is satisfied with the classifier’s performance, all the data product instances which are relevant to the query can be downloaded from the EO data archives. The team used a RESTful architecture for the web service and used the JAX-RS standard to implement it. The actual service is provided by a Glassfish Jersey web server.
|M1||May 13||WebEx||System Requirements Review|
|M2||Nov 13||University of Trento||Preliminary Design Review|
|M3||Mar 14||ESRIN||Critical Design Review|
|M4||Dec 15||WebEx||Acceptance Review|
|M5||Feb 16||ESRIN||Final Presentation|
The PCA team partners are
- Array Systems Computing Inc. (Software Lead)
- Brockmann Consult GmbH (Project Lead)
- University of Trento (Science Lead)