MACHINE LEARNING MODELS FOR SNOW, GLACIER TERRAIN AND HAZARD MAPPING

Abstract

In this thesis, a meta-classi cation ensemble approach is developed to improve the prediction performance of snow covered area. The methodology adopted in this case is based on neural network along with four state-of-the-art machine learning algorithms: Support Vector Machine (SVM), Arti cial Neural Networks (ANN), Spectral Angle Mapper (SAM), and k-means clustering. An AdaBoost Ensemble algorithm related to decision tree (DT) for snow cover mapping is also proposed. According to the available literature, these methods have been rarely used for snow cover mapping. Employing the above techniques, study is conducted for Raktavarn and Chaturangi Bamak glaciers, Uttarakhand, Himalaya using multi-spectral Landsat 7 Enhanced Thematic Mapper (ETM)+ image. The study also compares the results with those obtained from statistical combination methods (Majority rule and Belief functions) and accuracies of individual classi ers. Accuracy assessment is performed by computing the quantity and allocation disagreement, analysing statistic measures (accuracy, precision, speci city, Area Under the Curve (AUC), sensitivity) and Receiver Operating Characteristic (ROC) curves. A total of 225 combinations of parameters for individual classi ers are trained and tested on the dataset and results are compared with the proposed meta-classi er approach. It is observed that the proposed methodology produced the highest classi cation accuracy 95%, close to 94% that is produced by the proposed AdaBoost Ensemble algorithm. From the sets of observations, it is concluded that the ensemble of classi ers produced better results compared to individual classi ers. Extent of snow cover plays a vital role for better understanding of current and future climatic, ecological, and water cycle conditions. This study depicts a multilayer framework, employing hybrid classi cation approach based on integration of deep features and hand-crafted features, for meticulous mapping of snow cover area (SCA). It exploits diverse sensors, spatial, and temporal resolution satellite imagery. The pillars of the architecture are Convolutional Neural Network (CNN) for computing the deep spectral features and Random Forest (RF) classi er for peri forming the classi cation. Experiments are conducted for Khiroi region, Chamoli district, Uttarakhand for snow cover mapping using Sentinel 1/2 optical and Synthetic Aperture Radar (SAR) satellite imagery, along with Advanced Spaceborne Thermal Emission and Re ection Radiometer (ASTER) Global Digital Elevation Model (GDEM) derived topographic parameters. This framework with hybrid approach resulted in a classi cation accuracy of (97%) outperforming the state-of-theart algorithms, thereby providing an infrastructure for the future practical use of remote sensing resources for snow cover mapping in the Himalayan region. Glaciers present in the mountain regions constitute an important part of the earth system. Glacier components are very important input parameters for studies related to glacier mass balance, hydrological and climatological modelling and studies related to glacier hazard. The study proposes a weighted ensemble framework approach for glacier terrain mapping. The proposed methodology is based on the hybrid of classi er via weighted majority of vote and four states-of-the-art machine learning algorithms: Maximum Likelihood Classi cation (MLC), ANN, SVM, and RF. The weights are assigned to the classi ers using a priori recognition performance statistics. The study is conducted at the region near Khiroi, Uttarakhand, Himalaya employing multispectral Sentinel-2 satellite imagery, and derived topographic and texture parameters. It is observed that the proposed approach produced the highest classi cation accuracy (95.81%) with kappa coe cient (0.953). The results of the proposed hybrid techniques are compared with those of the individual classi ers. Accuracy assessment is performed by analysing several statistic measures and kappa coe cient. It is concluded that the ensemble of classi ers gives the better prediction of glacier terrain classes compared to the individual classi ers as it picks up the advantage of classi ers holding complementary information. Mapping of glaciers and their constant monitoring is essential to assess the impact of climate change on global coverage of glaciers. Even though studies mention innumerable methods for mapping clean glaciers but still to discover the extent of debris covered glaciers remains a challenge for researchers. The study proposes a novel hybrid Deep Learning (DL) framework approach for e cient and accurate mapping of debris covered glaciers. The framework comprises of integration of several CNNs architecture, in which di erent combinations of Sentinel-2 multispectral bands, topographic and texture parameters are passed as input for feature extraction. The output of an ensemble of these CNNs is hybrid with RF model for classi cation. The major pillars of the framework includes: 1) Technique for implementing topographic correction (pre-processing), 2) The proposed Hybrid of ii ensemble of CNNs and RF classi er, and 3) Procedures to determine whether a pixel predicted as snow is a cloud edge/shadow (post-processing). The proposed approach is implemented on the multispectral Sentinel-2 and Landsat 8 Operational Land Imager (OLI)/ Thermal infrared sensor (TIRS) data, and ASTER GDEM for the part of the region situated in Alaknanda basin, Uttarakhand, Himalaya. The proposed framework is observed to outperform (accuracy 96.79%) the current stateof- the-art machine learning algorithms such as ANN, SVM, and RF. The possibility of occurrence of snow avalanche depends both on the amount of snow which has been accumulated in the region along with the intensity of snowfall and other terrain parameters. In this study, an attempt has been made to compute avalanche prone areas in Alaknanda basin. For this, rule based classi cation approach is employed. Several parameters such as slope, surface curvature, elevation, aspect and peak ground acceleration (PGA) due to earthquake are used. First, snow accumulation of the region is computed which provides preliminary basis for occurrence of snow avalanche. It is observed that the regions which are detected as avalanche prone areas are very frequently a ected by snow avalanche occurrences. Avalanches are very frequent in regions where snow cover is present throughout the year. Damage caused by these avalanches is tremendous in the lives of humans and animals. Hence, it is important for researchers to understand reasons for the formation of avalanches and ways to control them. The study area is situated near Badrinath region, Alaknanda basin, Himalayas as shown in gure 11. For this study ve avalanche paths are detected, shed 1, shed 2, shed 3, shed 4 and shed 5. Both Normalized Di erence Vegetation Index (NDVI) and Tasseled Cap Transformation greenness method are applied in order to examine the amount of change in the vegetation cover due to occurrence of avalanche. The area is also classi ed using the supervised based classi cation method, to determine the amount of change in the vegetation area. Hence, it is observed that avalanche ow has a great impact on the vegetation of the area. It is determined that the regions in which the loss of vegetation is more, are subjected to more frequent occurrence of the avalanches. Avalanches can be activated due to an overload generated from excess snowfall, metamorphic transformations in ice-fall, rockfall, and snow pack. With their capacity to take bulk of snow at a fast pace, avalanches can cause a signi cant demolition to properties. In territories where snow avalanches constitute signi cant risk to individuals and properties, precautionary measures like snow wall, unloading avalanche potential snow packs by arti cial activation, and arti cial boundaries are setup to reduce their hindrance potential. The study proposes a deep learning iii based architecture that exploits both the spatio-temporal characteristics for snow avalanche classi cation.