CompEOD Work Plan


The project has a duration of 24 months and is structured into 3 phases. The project plan has a logical and continuous structure attained through the partitioning of each phase intro a series of activities well balanced in terms of theoretical research and experiments. The project aims at the development of a methodology for information retrieval from EO databases using compression data algorithms.



WP 1: Elaboration, implementation and validation of data compression algorithms for EO images - complete


Objectives: Define, develop and validate data compression algorithms for EO images. This analysis involves transforming the image into text, the generation of a dictionary to encode the text and the generation of the compressed ration vector. Several algorithms will be elaborated, implemented and validated, to construct a feature of the original input data.



TN 01 - Technical Note on Data compression algorithms for EO images:



1. Analysis, review and implementation of a set of methods to convert images into text.


Main challenge: to preserve dependencies between pixels in the EO images (including spatial relationships and spectral properties of the scene defined by the number of spectral bands).


2. Analyze, review and implement a set of algorithms for data compression  


Main challenge: obtain a good data representation (involves a tradeoff between data reduction and content description, with no information loss).


3. Definition of a methodology for feature extraction to describe image content based on compression ratio vectors.


Main challenge: general applicability regardless the type of analyzed data; to speed up classification process; ability to process a wide range of data; easy to deploy and fast working (involving as few resources as possible for data processing)




WP 2: Elaboration, implementation and validation of compression based similarity measures - complete 


Objectives: Extent the data compression algorithms to similarity measures that quantify the amount of information shared by a couple of images. Define, develop and validate compression based similarity measures such as NCD, FCD, aesthetic measures and then we perform a comparative study regarding the results considering image characteristics captured by them and the computational time.



TN 02 - Technical Note on Compression based similarity measures.

TN 03 - Technical Note on Comparative analysis towards identification of feature categories definition.




1. Analysis, review and implementation of a set of compression based similarity measures.


2. Evaluation and validation of the methodology proposed in WP1, including newly developed methods and algorithms. Defining new approaches for compression based analysis for information retrieval.


3. Semantic analysis for latent information discovery in EO images based on vector quantization.


4. Content based scene classification by combining heterogeneous data such as EO images and open access information, like text, maps or pictures, describing the same object. The main issue concerns the "leveling" of data, as it comes in very different formats, with different analysis requests. Moreover, each data source has particular constraints, requiring the definition of specific parameters.


  txt vs img

Analogy between the component elements of image and text. [C. Vaduva, Igarss2014]

Following a data leveling process, they will be both numerically represented as in the last column. 



WP 3: Elaboration, implementation and validation of information retrieval methods for EO image modeling using compression based similarity measures - complete


Objectives: Define, develop and validate information retrieval methods for EO image modeling using compression based similarity measures. There are two approaches developed in two activities: the EO image classification and EO image retrieval. The proposed methodologies will be applied on large collections of VHR EO images. Its generality will be proven by using different data sets: EO images acquired with different types of sensors, as well as multimedia data.



TN 04 - Technical Note on Scene classification using compression based similarity measures.

TN 05 - Technical Note on New approaches on content based EO image retrieval.





1. Tho parameter free methodologies for scene classification were proposed based on data compression algorithms. No pre-processing is required, the model is considered to be already included in the data. Feature selection is thus avoided, along with the human subjectivity when choosing a parameter.

general approach

General approach for compression based scene classification.



2. A study to evaluate and understand content based image retrieval using data compression algorithms and compression based similarity measures.


3. Evaluation and validation based on Earth Observation iamgery:


- Landsat 8 image, 30 m spatial resolution, 7 spectral bands, 3500x1500 pixels.

- WorldView 2 image, 2 m spatial resolution, 8 spectral bands, 2500x2500 pixels.



Bucharest area, WorldView2 image (left side), Scene classification using

dictionary distribution at patch level (center),Legend (right side).  [C. Vaduva, ACIVS 2015]







CompEOD_Scientific Report_2013-2014_brief