Evaluation of pixel based and object based classification methods for land cover mapping with high spatial resolution satellite imagery, in the Amazonas, Brazil
Abstract: In the state of Acre, Brazil, there is ongoing land use change, where inhabitants of this part of the Amazonian rainforest practice shifting agriculture. Practicing this type of agriculture is, according to the SKY Rainforest Rescue organization, damaging to forest ecosystems. This organization aims to educate people in how to maintain sustainable agriculture. By monitoring this shift in agricultural practices with the use of remotely sensed data, the organization can follow the development. In this thesis, an image with high spatial resolution from the SPOT-5 satellite is used to evaluate which classification method is most appropriate for monitoring land use change in this specific area. Three methods are tested; two pixels based and one object based. The pixel based methods are the Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel and the Maximum Likelihood Classifier (MLC), and the object based method is segmented with Multi Resolution Segmentation (MRS) and classified with the k-Nearest Neighbor (kNN). The parameters gamma and penalty parameter C in the SVM with an RBF kernel were estimated by a k-fold cross validation and grid search method; and for the MLC, an assumption that each class had an equal probability distribution was made. For the object based approach the first step was segmentation; for the MRS there are three parameters: scale, shape and compactness. The scale parameter was set by using an algorithm that was based on comparing local variance; shape and compactness were defined based on previous studies and visual evaluation of the segments. All three methods will produce two classified maps each; one where the feature space consists of the three original wavebands (green, red and NIR) and one where the feature space is of six dimensions that include the original three wavebands and three texture derivations, one from each original band. The texture is derived from the co-occurrence GLCM method, which can be used to calculate 14 different texture measures. The three most suitable texture derivations were the contrast texture measure from the green and NIR band, and an entropy texture derived from the red band. When combining these three texture derivations with the original bands, the classes were further separated. The original image was also lowered in resolution, from 2.5m to 25m in pixel size. However, this did not generate either higher or similar overall accuracy compared to any of the high spatial resolution classified images. The moderate spatial resolution classifications were only computed with the MLC and SVM due to the inefficiency of an object based image analysis method when used with moderate spatial resolution. Of these six classifications, only two exceeded the 85% threshold of an acceptable overall accuracy. These were the SVM (86.8%) and OB-kNN (86.2%), which included the texture analysis. None of those classifications with only the three original bands exceeded this threshold. In conclusion, the object based method is the most suitable approach for this dataset because: 1) the parameter optimization is less subjective, 2) computational time is relatively lower, 3) the classes in the image are more cohesive and 4) there is less need for post-classification filtering.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)