0 avis
Comparison of the contributions of maximum likelihood, Svm and random forest to forest cover mapping
Archive ouverte
Edité par CCSD -
International audience. In this study we produce forest cover maps of the Pyrenees Mountains (Spain, Andorra, France) using supervised classification. The final prospect of this work is to carry out a change analysis and will be achieved obtaining a sufficient accuracy level for our classification results. The study presented aims to compare the quality of the classification results obtained with three different algorithms, maximum likelihood, Support Vector Machine approach (SVM), and Random Forest, with or without stratification. The classification process is based on a Normalized Difference Vegetation Index (NDVI) MODIS time series (MOD13Q01). We used images from 2000 to 2012. The annual NDVI profile is used as a temporal signature. A wide topographic, climatic and altitudinal variability is observed for the study area, conditioning the diversity of its vegetation. This context implies to take into account thirty different types of forest or natural vegetation. Mixed forest types or vegetation transitions are frequently observed in the study area. To observe how the three algorithms dealt with local appearances or small areas of specific forest types, a supervised classification was carried out for each one of them in two cases : with and without stratification. An Object-based image segmentation was therefore tested to provide a partitioning of the study area. This methodology uses a temporal series of vegetation and texture indices and an Object-Based Image Analysis (OBIA). An unsupervised analysis of several segmentations allowed selecting the best combination of input variables (seasonal/monthly vegetation and texture indices) and the best segmentation parameters. The accuracy of the classification results obtained for each of the 6 configurations studied was assessed calculating the Kappa Index, the overall accuracy and the thematic accuracy of each class (proportion of area with a correct classification). A comparison of these values was carried out. Our study is still being proceeded, but we consider it could highlight differences of temporal signature confusions, depending on the algorithm applied and on the presence or absence of stratification. If these observations are confirmed, it could be concluded that maximum likelihood, SVM and Random Forest are complementary approaches to apply a change analysis to specific locations or forest transitions.
Consulter en ligne
Suggestions
Du même auteur
