Please use this identifier to cite or link to this item:
Authors: Kanungo, Debi Prasanna
Keywords: FUZZY
Issue Date: 2006
Abstract: Landslides are complex natural phenomena that constitute a serious natural hazard in the Himalayan region, causing damage to both property and life every year. Identification of landslide-prone areas is essential for safer strategic planning of future developmental activities. Therefore, landslide susceptibility zonation (LSZ) becomes important whereby an area may be divided into near-homogeneous domains that are ranked according to the degree of potential hazard due to landslides and mass movements. The most important task for LSZ studies is the determination of weights and ratings giving relative importance of factors and their categories respectively, for landslide occurrence. These weights and ratings can be determined by implementing different approaches, which at times are highly subjective in nature. Therefore, developing suitable approaches for determination of weights and ratings objectively and their implementation in a geographic information system (GIS) environment for LSZ mapping is extremely important. Further, most of the landslide-related studies culminate at providing LSZ maps only, therefore, procedures of landslide risk assessment (LRA) needs attention. The main objective of this research is to explore the potential of the advanced techniques - fuzzy set theory and artificial neural network (ANN) for determination of weights and ratings of causative factors and their categories, and to devise a fully objective approach for GIS based LSZ and LRA mapping. The study area covers a small region of about 254 km2 in Darjeeling Himalayas (Latitude 26° 56'-27° 8' N and Longitude 88°10'-88°25' E). It is a part of Darjeeling district of West Bengal State of India. The area lies within the Lesser- and SubHimalayan belts. Tea plantations form the most widespread land use. The area in the eastern part is dominated by thick forest. The following datasets have been used to generate various thematic data layers: a) Remote sensing images from IRS-1C LISS-1II multispectral and IRS-1D-PAN b) Survey of India (SOI) topographic maps at 1:50,000 and 1:25,000 scale c) Published geological map (Geological Survey of India) d) Extensive field data on landslides and land use land cover The LISS-III and PAN data have been precisely co-registered with the topographic map. The LISS-III data have been corrected for atmospheric path radiance. The various thematic data layers pertaining to causative factors of landslides form the input layers for LSZ mapping and have been generated using remote sensing-GIS tools. The digital elevation model (DEM) of the study area has been generated by digitization of contours on SOI topographic maps from which slope and aspect data layers have been derived. The lithology layer has been prepared by digitizing polygons from the coregistered geological map. Minor modifications in lithologic boundaries at some places have also been incorporated in the vector layer after field verification. The layer was rasterized. Lineaments have been interpreted from the PAN and LISS-III images. There is no major thrust/fault reported in the study area, but major lineaments have been identified. A lineament buffer layer was generated to deduce the influence of lineaments on the occurrence of landslide. A drainage layer has been prepared from the topographic maps and LISS-III image. The ordering of streams has been performed on the basis of Strahler's classification scheme. A drainage buffer layer with 25m buffer zone along 1st and 2nd order drainages (only) has been generated for further analysis. Eight dominant land use land cover classes, namely, thick forest, sparse forest, tea plantation, agriculture, barren, built up, water bodies and river sand in the area have been deciphered. The four spectral bands of LISS-III image, DEM and normalized difference vegetation index (NDVI) image have been integrated to prepare a land use land cover layer by a multi-source classification process using the most widely adopted maximum likelihood classifier. The field data and high spatial resolution PAN and PAN-sharpened LISS-III images have been used to produce a landslide distribution map. A total of 101 landslides have been identified, the majority of which have areal extent of 500 m2to 2000 m2. For LSZ mapping, four different approaches have been implemented to determine the weights and ratings. The most commonly used conventional weighting approach involved assignment of weights and ratings to the factors and their categories based on field knowledge. A landslide susceptibility index (LSI) map has been generated by integrating the weighted layers and the range of LSI values has been categorized into five landslide susceptible zones to prepare an LSZ map. In order to minimize subjectivity in the weight assignment process, at the first instance, ANN black box model has been attempted for LSZ mapping. A feed forward multi-layer ANN with one input layer, two hidden layers and one output layer has been designed. The input layer contained 6 neurons corresponding to 6 different causative factors. The LSZ map obtained from conventional weighting approach has been used as the reference map. The training and testing datasets consisted of 2500 mutually exclusive pixels corresponding to 500 pixels per landslide in susceptibility zone. The Levenberg-Marquardt back-propagation algorithm has been used to train the neural networks. A total of 39 neural network architectures were designed, trained and tested. The network architecture 6/13/7/1 has been found to be the most appropriate one. The connection weights obtained from this network have been captured and subsequently used to determine the network output of all the pixels in the dataset to prepare an LSZ map of the area. However, the major limitation of the ANN black box approach is that the weights and ratings of the factors and their categories remain hidden. In this research, an attempt has been made to open the ANN black box. A novel approach to derive weights for causative factors has been proposed, which has been referred to as ANN connection-weight method. The assignment of weights in this fashion may reduce the subjectivity. Moreover, to bring objectivity in the assignment of ratings, a new fuzzy set concept has been utilized. As a result, two unique ways of LSZ map preparation have been proposed here, namely fuzzy set based and combined neural and fuzzy approaches. In the fuzzy set based approach, ratings (r#) of each category of a given thematic layer have been determined using the cosine amplitude method. The integration of these values for various categories of thematic layers has been performed to compute LSI values in two different ways: (a) using arithmetic integration and (b) using fuzzy gamma operator. The range of LSI values thus determined has been categorized into five landslide susceptibility zones using success rate curves method to prepare the LSZ maps. The performances of the two methods have been examined and it has been found that the arithmetic integration approach has yielded better results than the fuzzy gamma operator in the present case. The combined neural and fuzzy approach has involved three main steps: (a) determination of weights of thematic layers through ANN connection-weight analysis, (b) determination of ratings for categories using cosine amplitude method and (c) integration of ratings and weights to generate the LSZ map. A feed forward back-propagation ANN with one input, two hidden and one output layers was considered. The data for the input neurons correspond to the normalized ratings (r#) of the categories. The output corresponds to the presence or absence of landslide at the pixel. 100 neural network architectures were designed, trained and tested. The adjusted weights of input-hidden, hidden-hidden and hidden-output connections for each network were captured and analyzed to obtain the weights for thematic layers corresponding to 6 factors. The integration of these weights for causative factors and the ratings for the categories (obtained from cosine amplitude fuzzy similarity method) has been performed to obtain distribution of LSI values across the area. The range of LSI values has been categorized into five different landslide susceptibility zones using success rate curves method to produce the LSZ map. A comparison of the LSZ maps produced from different approaches is very important. The LSZ maps have been compared and evaluated using three different approaches: a) landslide density analysis, b) error matrix analysis and c) difference image analysis. Landslide density is defined as the ratio of the percent existing landslide area to percent area of each landslide susceptibility zone, and is calculated on the basis of the number of pixels. It has been found that the LSZ Maps produced from conventional and ANN black box approaches have a similar trend of landslide densities for various susceptibility zones. This result is on expected lines, as the conventional weighting based LSZ map has been used as the reference map to generate ANN black box based LSZ map. The LSZ Map produced from combined neural and fuzzy approach has a much higher landslide density of VHS zone (>13) as compared to other LSZ maps. Based on the landslide density analysis, it is inferred that the LSZ map produced from combined neural and fuzzy approach is significantly better than those produced from other approaches (fuzzy, conventional and ANN black box approaches). Three error matrices for different LSZ map combinations have been generated to understand the distribution of number of pixels in different LSZ maps. There was a high degree of matching in the pixels of LSZ maps produced from the conventional weighting and ANN black box approaches. There was also a general correspondence in the LSZ maps produced from fuzzy and combined neural and fuzzy approaches. There was a lot of mismatch in number of pixels between LSZ Maps produced from the conventional weighting and combined neural and fuzzy approaches. This mismatch is inferred to be due to the differences in weights and ratings of both the approaches. Difference image analysis elucidates how pixels shift from one landslide susceptibility zone to another zone, based on the LSZ mapping approach adopted. A difference image of LSZ maps produced from conventional and ANN black box approaches showed a high degree of mutual correspondence and matching of landslide susceptibility zones throughout the area. A difference image of maps produced from fuzzy set based and combined neural and fuzzy approaches showed a high degree of spatial matching, with about 50% pixels having full matching, and 47% pixels exhibiting only one-zone difference. About 3.0% pixels showed two-zone difference and these mainly appeared to be related to a lithologic band in the northern part of the area. The difference image of maps produced from conventional weighting and combined neural and fuzzy approaches appeared to exhibit the widest spatial difference, with only 37.8% pixels fully matching, 46.4% pixels exhibiting one-zone difference, 14.6% pixels exhibiting two-zone difference, and 1.2% pixels showing three-zone difference. The most important was a two-zone difference band in the northern part of the difference image marking a lithologic layer. As lithology has the highest and significantly higher weight than other factors in combined neural and fuzzy approach, the importance of lithology has been brought out in the difference image. Further, the relative importance of drainage as observed in the field vis-a-vis lineament as deduced from spatial-domain filtering is deduced and discussed. As far as landslide risk assessment (LRA) is concerned, landslide risk is considered to be a function of landslide potential (LP) and the resource damage potential (RDP). In the present study, two different approaches namely, (I) LRA using danger pixels and (2) LRA using Fuzzy Concept have been developed and implemented to prepare LRA maps of the study area. A concept of danger pixel has been introduced for landslide risk assessment. Danger pixels are considered as those pixels which lie in VHS and HS zones in all the four LSZ maps produced from different approaches, i.e., the danger pixel map is an intersection map of all the four LSZ maps with (VHS + HS) zones combined. A resource map including all the existing land use land cover patterns and also the road network of the area has been prepared. The danger pixel map and the resource map have been integrated to generate the LRA map of the study area. The LRA map shows spatial distribution of different resource categories that appeared to be under real danger due to landslides. In the LRA using fuzzy concept approach, the LSZ map prepared using the combined neural and fuzzy approach; the best LSZ map of the area, has been used as an input to provide LP. Further, the resource map has been used as another input layer VI i to provide information on RDP. Linguistic rules have been framed for landslide susceptibility zones and resource categories and the fuzzy membership values representing the LP and RDP based on these linguistic rules have been assigned. Landslide risk values for different combinations of LP and RDP have been obtained by integrating LP and RDP layers and have been represented in the form of a LRA matrix. The range of landslide risk values has been segmented into five different landslide risk zones and the LRA map of the area has been prepared. It is observed that 2496 pixels (0.61% of total area) are under very high risk zone and 7204 pixels (1.77% of total area) are under high risk zone. The LRA Map has revealed that landslides pose very high risk to a few selected sites of habitation in Sonada. Darjeeling and northeastern part of Tiger hill, and high risk to a section of road from Sonada to Ghum. It is considered that the approaches developed in this study for objective LSZ and LRA mapping can be successfully implemented in other hilly regions that are susceptible to landslides.
Other Identifiers: Ph.D
Appears in Collections:DOCTORAL THESES (Earth Sci.)

Files in This Item:
File Description SizeFormat 

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.