REMOTE SENSING BASED GLACIER TERRAIN MAPPING IN PARTS OF CHENAB BASIN, HIMALAYA

Abstract

Himalayan glaciers constitute the lifeline of several thousands of people inhabiting the densely populated plains of the Indian Subcontinent. The fresh-water needs of these people for domestic, industrial and agricultural consumption and for hydropower generation make them heavily dependent on the perennial supply from Himalayan glacier-fed natural drainage. Therefore, glaciers profoundly influence the economic growth, industrial development, planning and management of water resources and tourism of the regions surrounding the Himalaya. However, glaciers in Himalaya, like other glaciated areas around the globe, have been experiencing severe stress resulting in increased rates of recession and deglaciation, as a consequence of global climatic warming. The dynamic and fragile nature of the Himalayan glaciers necessitates effective mapping and monitoring of their characteristics, such as their areal extents, mass balance, velocities, depth, runoff, etc. Accurate areal estimates of glaciers derived from their detailed mapping, constitute a key input to most of the glaciological studies such as the mass balance, estimation of glacier fluctuations, snow runoff modelling, glacier hazard prediction modelling, climatological modelling, etc. Remote sensing technology, due to its numerous advantages makes an obvious choice for detailed, complete and repetitive studies of the remote and inaccessible terrains such as the Himalaya, which are characterised by most unpredictable and severe weather conditions. However, remote sensing based mapping of glaciers, in general, is faced with several issues such as mapping of glacier under cloud cover, forest cover, mountain or cloud shadow, and debris cover. In addition to these, the status of remote sensing based mapping of glaciers in India is an immediate need, since even a complete inventory of Indian Himalayan glaciers as per global standards is still lacking and the manual digitisation of glacier remains the most widely used technique. In view of the abovementioned, the main aim of this research has been to explore the potential of satellite remote sensing for detailed and accurate mapping of the glaciers using prevalent, advanced and novel techniques. The study area selected for this work comprises of the Samudra Tapu glacier which covers an area of about 200 km2 and is part of the Chenab basin, Himalaya. The focus has been laid primarily on accurate mapping of various glacier terrain classes (snow, ice, mixed ice and debris (MID), supraglacial debris (SGD), periglacial debris (PGD), valley rock (VR), water (WT) and shadow (SHD) using remote sensing data. The data used for fulfilment of these objectives comprise of Survey of India (SOI) topographic maps (1963) (at 1:50,000 scale), remote sensing image data from various satellite sensors, ancillary data (DEM and its derivatives: slope and aspect and transformed spectral bands) and field data (photographs and in situ spectral reflectance curves). Data from IRS-P6 AWiFS and Terra ASTER sensors constitute the primary remote sensing data used for glacier terrain mapping. Data from IRS-P6 LISS-IV sensor has also been used in addtion to the above data for studying the geomorphology of the area. Besides these, a multi-temporal (1976-2006) and multi-sensor dataset has also been used to estimate the retreat and depletion of the test glacier. Before actual implementation of the image processing techniques, the remote sensing data have been pre-processed to remove the inherent geometric and radiometric defects. Preprocessing constitutes georeferencing and co-registration of the images, radiometric corrections and conversion of raw data to physical units of reflectance or brightness temperature, respectively for the optical reflective data and the thermal radiance data. Finally, the image data have been topographically corrected using the C-correction method. Detailed field investigations coupled with interpretations of remote sensing data have revealed the presence of various landforms of glacial erosional and depositional origin. Some glaciofluvial landforms and evidences of periglacial activity have also been found. Large-scale erosional and depositional landforms are clearly identified in digital 3-D perspective views generated by draping the satellite images over the DEM of the area. Further, comparison of the spectral reflectance curves of the major land cover classes measured in field with the image derived spectra have revealed a positive correlation verifying the quality of the radiometric corrections applied in the pre-processing stage. Identification of the abandoned lateral and terminal moraines in the area together with the presence of a moraine-dammed lake pointed towards the recessional history of the glacier. With this background a detailed study to estimate the retreat (from 1963 to 2006) of the Samudra Tapu glacier via comparison of glacier extents derived from topographic maps of 1963 and the multi-temporal remote sensing dataset (from 1976 to 2006). Results reveal that the glacier has receded by 969 m at an average rate of retreat 22.53 m/yr and the total glacier area has reduced by 14.44 km2 (15.29%) during the last 43 years (1963-2006). It has also been found that the rate of recession (42.37 m/yr) and overall depletion (at the rate of 0.80 km / yr) have increased markedly in recent years (2000-2006). Application of the prevalent techniques of glacier terrain mapping to AWiFS data has shown that some of the techniques (manual delineation, spectral ratios and indices) have limitations in detailed mapping of the glacier terrain. However, supervised classification owing to its superior capability may be more appropriate than the other techniques. Supervised classification of AWiFS DN and reflectance data (via Maximum Likelihood Classifier) has shown that classification of reflectance data leads to higher classification accuracies (increase of- 2.86% in overall accuracy (OA)) as compared to the DN data. This means that conversion of DN values to reflectances has improved the identification/classification of glacier terrain classes. It has also been found that in both the cases (DN and reflectance) a two band combination comprising of Green (B2) and NIR (B4) bands provides the highest values of overall and individual accuracies. However, some glacier terrain classes due to their inherent in mixed and spectrally similar nature (MID, SGD, PGD and SHD) have exhibited persistent lower individual accuracies. For upgrading the accuracies of these difficult classes, information from ancillary data (spectral ratios, indices, elevation, slope and aspect) has been incorporated together with spectral data (DN and/or reflectance data). Incorporating ancillary data along with AWiFS DN data has resulted in an enhancement of the classification accuracies (with an increase of ~ 5.63% in OA); however, in case of reflectance data the improvement has not been as significant. Further, use of spectral ratios and NDSI have led to significant improvement in the classification accuracies (OA, producer's accuracy (PA) and user's accuracy (UA)) when used along with DN data probably due to their additional capabilities of topographic normalization. However, the topographic attributes (DEM and its derivatives) have shown limited utility in this regard, which may be attributed to the fact that most of the glacier terrain classes are not > characterized by a single set of topographic attributes. Though, the use of ancillary data together with spectral data has improved the individual accuracies of the difficult classes to some extent, the accuracies of these classes have still been found to lie in the range of 60-75%. For further improving the segregation of difficult classes and increasing the classification accuracy, the potential of integrated optical (from AWiFS - four bands) and thermal (from ASTER - five TIR bands) for glacier terrain mapping has been explored. Results have indicated that application of integrated optical and thermal data for glacier terrain mapping leads to a remarkable improvement in the overall (with an increase of ~ 11.17% in OA) as well as individual accuracies of the glacier terrain classes, especially the mixed and spectrally similar classes (MID, SGD, PGD and SHD). A band combination (IB1, IB3, IB6, and IB8) comprising of two optical and two thermal bands gives the highest overall accuracy of 89.35 % and individual accuracies of the problem classes ranging from 80.99% to as high as 100.00%. Integrated optical and thermal data has also formed the basis of a synergistic iv approach devised for the mapping of debris-covered glacier boundaries. Accurate separation of supraglacial and periglacial debris using the integrated optical and thermal datasettogether with some essential DEM derived geomorphometric parameters have been used synergistically in an approach devised for mapping of the debris-covered glacier boundaries. Evaluation of the glacier-boundary delineated using proposed approach against the glacial trough width map and the visually interpreted boundary of the glacier clearly has shown that the devised approach successfully discerns glacier boundaries hidden beneath debris cover. Further, inter-comparison of all the per-pixel classifications performed using different datasets (as mentioned above) with respect to the propinquity of their areal estimates of the glacier terrain classes to those derived from the reference map has also shown that the use of integrated optical and thermal dataset show relatively closest match. Thus, integrated optical and thermal dataset have been recognized as the most appropriate dataset for glacier terrain mapping at per-pixel level. Use of per-pixel classifiers to classify images from moderate spatial resolution sensors AWiFS, essentially leads to erroneous areal estimates of classes since such images are usually fraught with mixed pixels (i.e., pixels containing more than one class). To address this problem, the AWiFS reflectance datasets which had been found to be successful at per-pixel level have been classified using some sub-pixel classification algorithms namely, Fuzzy MLC, Fuzzy cmeans (FCM) clustering, and Evidential Reasoning (ER). Accuracy assessment of outputs from the sub-pixel classification is quite complicated as it requires reference data at sub-pixel level (soft reference data). The soft reference data for the moderate resolution image (AWiFS), in this study, has been generated using a fine resolution image (ASTER VNIR) of the area. Fuzzy error matrix (FERM) based accuracy assessment and visual inter-comparison of the resultant fraction images with the corresponding reference fraction images has revealed that that MLC outperforms the other two classifiers, and they can be ranked in decreasing order of their performance as MLC, FCM and ER. Further, FERM based accuracy assessment together with areal estimate based evaluation of the fraction images from different datasets have unanimously shown that the sub-pixel classification of the integrated optical and thermal dataset does result in appreciably improved results which are much closer to the reference data. Finally, areal estimates based on evaluation of all the per- and sub-pixel classifications with the reference map have shown that while sub-pixel classification of majority of datasets (REF Case 9 and IOT Case 3) leads to improved areal estimates as compared to those obtained from per-pixel classification, some exceptions are also found (REF Case 6). Inter-comparison of areal estimates have also confirmed the inference that incorporation of integrated optical and thermal data for classification of the glacier terrain at per- as well as sub-pixel levels results in most accurate estimates of the glacier terrain classes. In summing up, it can be corroborated that the methodologies developed and proposed in this research have sufficiently demonstrated their usefulness for accurate and detailed mapping of mountain glaciers and may be successfully implemented for mapping of the glacier terrain at operational level.