COMBINING MULTI-SENSOR MEDIUM RESOLUTION SATELLITE IMAGERY FOR FOREST COVER CHANGE ASSESSMENT IN CENTRAL AFRICA

摘要

assess the evolution of forest cover changes and givernfigures about deforestation. Based on a systematicrnsample of image extracts, a processing chain had beenrndeveloped for producing deforestation estimates overrnthe years 1990-2000-2005. Whereas this monitoringrnexercise was based on Landsat images, limitations inrnLandsat image availability for the year 2010 overrnCentral Africa required alternative imagery. Given itsrnhigh revisit period and characteristics close to thernLandsat TM sensor, DMC images are considered in thisrnpaper to replace Landsat TM data gaps over CentralrnAfrica. However the classification module of thernexisting processing chain is based on Tasseled Capsrn(TC-based) analysis of Landsat TM imagery and needsrnto be adapted to such data in order to be sensorindependent.rnThis adaptation is described in this paper.rnA sub-sample of the available image extracts has beenrnused for the selection of the best object-based featuresrnthrough the analysis of existing Land-Cover maps forrn1990 and 2000. The processing chain has been adaptedrnfor the production of Land-Cover change maps for yearrn2010. The resulting maps from the two methods,rnoriginal TC-based classification and adapted Multi-rnSensor approach, have been compared and evaluated.rnThe overall accuracies of the 2010 Land-Coverrnclassification results are 90% for the TC-based approachrnand 92% for the Multi-Sensor approach. Whenrnconsidering only objects for which label is changingrnbetween 2000 and 2010, the accuracies of the 2010 LCrnclassifications are 45% and 72% for the TC-based andrnMulti-Sensor approaches respectively. These resultsrnshow that, even with lower radiometric quality of DMCrnimagery the performance of the automated classificationrnhas been improved thanks to an appropriate selection ofrnobject-based features. As similar adaptation will berndeveloped for other satellite imagery such as SPOT andrnRapid-Eye in order to be sensor-independent, the futurernadaptation to Sentinel-2 data will be very easy using thernsame approach.