Using Exogenous Data for Image Analysis in the OTB Framework
Lahlou, Otmane; Ruiloba, Rosario

The goal of the study is to improve automatic image analysis by using available and complementary data. These exogenous data (maps, DEM, risks maps, etc.) are often used by image analyst in GIS to complete automatic image results. The purpose of the study is to show how some exogenous data can be integrated in classification chains to reduce human supervision and improve results. The exogenous data used are those coming from communities providing open data:
- Maps: from Open Street Map community
- DEM: from SRTM data
- Maps risks: from institutional like "Cartorisque" Web map service for example

Integration of exogenous data has been done on SVM classification method and used for urbain image classification and for risk mapping generation. SVM classification module can become unsupervised using OSM data as a priori information. OSM vectors and labels are used respectively as masks and classes for generating learning samples for the SVM model estimation. Results show that using OSM data for learning samples generation would be a solution to include SVM classification in automatic or semi-automatic mapping chains. Results based on radiometric characteristics for image classification (tested on Quickbird image) were equivalent with those based on by hand learning sample selection or automatically extracted from OSM. Best results were obtained for radiometric moments, other radiometric descriptors tested are redundant.

OSM Data has been included in two Monteverdi modules: SVM Classification module to make unsupervised the SVM-based classification and into OBIA((Object Based Image Analysis) module for automatic labeling of input segmented regions.

Concerning change detection, several available data seems useful for adding \textit{a priori} knowledge about potentially affected regions for rapid mapping of disaster affected zones. Several exogenous data are used here to improve flooded regions detection based on change detection classification:
- DEM: flooded regions are probably lower ones :SRTM data are used
- Maps: regions close to a river are more probably flooded :OSM data are used

Risk Maps: risk masks indicate potentially flooded zones :Cartorisque data from from frecnh national risk agency are used. The Monteverdi ( module for change detection is adapted to manage exogenous data, to: read OSM data, compute complete change descriptors with DEM, risk and other information, normalize and center learning samples... Results are compared to crisis products generated by SERTIT image analysts. The best automatic results are obtained by using a risk map (DEM multiplied by river distance map) and a water index.