Performance Analysis of Object Oriented Remote Sensing Techniques for Forest Detection in Pakistan

Abstract

Effective monitoring of forestry in Pakistan is a challenging task for the environmentalists due to lack of automated procedure for land cover classification and forest identification. Currently, manual procedure of utilizing ground surveys and reports is used for forest detection, but such techniques are not reliable because of the possibility of human errors and deliberate alteration. Therefore, it is very important the remote sensing based effective monitoring techniques should be developed for the effective observation of land cover and green forests. Areas which are unapproachable or approachable at the cost of valuable resources can easily be monitor using remote sensing satellite imagery. In our experimental setup we have utilized 4999sqkm with 2.5m resolution multispectral SPOT-5 imagery of Abbottabad district of Pakistan. Classification of the land cover consisting of different regions of interest including the forest is challenging task in pixel based classification approach because of the heterogeneity in local pixel values of neighboring classes and similar spectral properties which restrict the class discrimination and yield undesired results. Visible effect of salt and pepper can also be seen in pixel based analysis. Object oriented classification approach (OOCA) overcomes these constraints and provides higher classification accuracy with more flexibility to extract different types of features. Homogeneous pixels that are consisting of similar spectral, spatial or texture characteristics are segmented to form object slices, which ideally corresponds to real world objects. While employing the OOCA with nearest neighbor classification based on training samples, user and producer accuracies of class forest is 100% with overall accuracy of 96% and kappa coefficient value of 0.96 was achieved. Support Vector Machine is applied on same training samples which were used for nearest neighbor algorithm and achieved improved results from that of NN classification. Overall accuracy of 98.93 % and kappa coefficient of 0.98 is achieved with user and producer accuracies of class forest as 100% each. Visual analysis shows that the classified images using OOCA are smoother due to the process of segmentation prior to classification. Based on these results, we are confident that the object oriented classification approach is favorite for identification of forest in Pakistan.