SATELLITE IMAGE BASED SUBPIXEL CHANGE DETECTION OF SNOW/ICE COVER OVER HIMALAYAS

Abstract

ABSTRACT The snow cover region acts as a natural source of water for power generation, agriculture growth, and smooth running of the hydrological cycle of the earth system. The mapping and monitoring of the snow/ice changes over different regions of the Earth surface is a challenging task due to regional rugged topography which leads towards shadow effects. Change detection is one of the most common approaches to extract maximum information and generate change maps from remotely sensed imagery. Traditional change detection methods follow the per pixel allocation in which only one class is assigned to a specific pixel. For accurate snow/ice mapping, there is a requirement of a sophisticated model that includes the consideration of terrain features, semi or fully automatic threshold methods, and detection of the small spurious changes. On the other hand, the satellite imagery acquired over rugged topography is generally affected by the topographic effects in form of the shadow. To overcome such limitations, this research work involves the (a) evaluation of topographically derived different classification algorithms over rugged terrain Himalayas using remote sensing data, (b) development of a subpixel-based change detection algorithm to identify and quantify the land-use and land-cover variations at the subpixel level and (c) analysis of spatial and temporal variability of snow/ice cover over the Himalayas using the proposed subpixel -based change detection algorithm From the experimental outcomes (statistical, graphical, and visual analysis), it is concluded that the Slope Match (SM) model is one of the best topographic correction (TC) techniques to overcome the topographic effects (shadow effects) with an accuracy of 92.89% as compared to Cosine correction (91.19%), C-correction (91.78%) and Variable Empirical Coefficient Algorithm (VECA) correction (91.70%). To evaluate the performance of TC on different classification models, three well-known classifiers i.e. K-Means Clustering (KMC) as unsupervised, Maximum Likelihood Classifier (MLC) as supervised, and LSM as subpixel classifier has been tested with TC and without TC using the AWiFS dataset. It is observed that LSM with TC achieved OA of 88.5% higher accuracy as compared to MLC (81%) and KMC (74%). It is also noteworthy that the optical data has good resolution but it is generally affected by the presence of clouds. Therefore, different classification models have also been demonstrated on the microwave (SCATSAT-1) dataset. From overall statistics, it is apparent that LSM (75.61 to 91.36%) achieved better accuracy as compared to KMC (71.79 to 85.47%) and Support Vector Machine (SVM) (72.07 to 85.71%