Landsat data have been extensively utilized for the purpose of delineation and mapping glacial lakes due to their free availability, 30 m resolution and wide area coverage [37, 14,59, 55]. Landsat images of 30 m resolution and the Advanced Space-borne Thermal Emission & Reflection Radiometer – Digital Elevation Model (ASTER DEM, 30 m) downloaded from the web portal[1] were used in the present study (Table 2). The Landsat images hosted by the Global Land Survey (GLS) data system of the United States Geological Survey (USGS) were downloaded from their web portal[2] and used for the current study. Glacial lakes show minor changes in the post-monsoon season with limited cloud and snow cover [60]. Landsat scenes used in the study were from the post-monsoon season with less cloud coverage (<10%) or were cloud-free. The Landsat images from 1990 to 2018 were used to prepare glacial lake inventories for area change analysis. Validation of glacial lake boundaries for 1990 was performed using the Survey of India (SOI) toposheet 1979, Scale 1:50,000. The SOI toposheets alone are less reliable for change detection analysis [41]; hence, the Landsat TM (1990) scene has been used as a base image for glacial lake area change.
To validate the glacial lake outlines, Google Earth imagery of high resolution was used because of the comparatively coarser resolution of the Landsat data. The time series of meteorological data (e.g., temperature and precipitation) from 1980 to 2018 of the Pehalgam observatory was analysed to understand climate as a possible influencing factor for glacial lake changes in the study area.
Table 2 Satellite data used
Date of Pass
|
Satellite & Sensor
|
Bands & Wavelength (μm)
|
Path/Row
|
Spatial Resolution (m)
|
Repeat Cycle (Days)
|
Cloud Cover
|
14 Oct. 1990
|
Landsat TM
|
6
|
149/36
|
30
|
16
|
No
|
09 Oct. 2000
|
Landsat TM
|
6
|
149/36
|
30
|
16
|
18 Oct. 2010
|
Landsat ETM+
|
8
|
149/36
|
30
|
16
|
04 Oct. 2018
|
Landsat L8 (OLI)
|
9
|
149/36
|
30
|
16
|
Thematic Mapper TM. Enhanced Thematic Mapper ETM. Operational Land imager OLI.
Glacial lake mapping
A number of methods (processing satellite imagery) have been used for mapping glacial lakes over time. The two commonly used methods are automated and semiautomated methods, e.g., the normalized difference water index (NDWI) [1, 37, 25, 26, 70]. In the present study, we used the normalized difference water index NDWI – first proposed in 1996 by McFeeters – using a semiautomated method similar to [29] for delineation of glacial lake outlines (Figure 2; Equation 1).
*BNIR and Bblue are reflectance in the near infrared and blue bands, respectively
Pixel identification pertaining to the lakes was performed based on NDWI values in the range of -0.60 to -0.85 [24]. A few mountain-shadowed areas were mistaken as lakes due to the same spectral reflectance and topographic effects [62, 34]. Misclassification could occur due to near similarities between glacial lakes from frozen lakes and snow cover because of peculiar surface conditions. The pixels identified as mountain shadows were removed from ASTER DEM-derived values for slope, aspect and hill shade in NDWI images to overcome misclassification of lakes because of topographic effects (Figure 3).
Further correction was performed through a visual interpretation technique using ArcGIS 10.2 [34, 58, 46, 76, 60, 67]. Through this process, the first glacial lake layer in 2018 was prepared. The glacial lake outlines were overlaid with Google Earth imagery for validation and later crosschecked with the toposheet glacial lake inventory generated from SOI toposheet 1979 at a scale of 1:50,000. Corrections, if any, were taken on priority. Subsequently, glacial lake inventories for 2010, 2000 and 1990 were prepared to obtain the final database to observe glacial lake expansion in the study area.
Uncertainty analysis
In the present study, glacial lakes were identified, delineated, and mapped to observe changes in spatial extent using multidate/multisensor remote sensing data. Uncertainties in glacial lake mapping occur mainly due to image coregistration, area delineation and editing using manual interpretation. As a result, thorough consideration of errors is required to determine the accuracy and relevance of the findings. High-resolution satellite imagery would be the most precise way to assess the errors related to glacial lake outlines [41]. However, high-resolution satellite data were not available for the present research work. Therefore, we used Landsat satellite data in conjunction with high-resolution Google Earth imagery to maintain the accuracy of glacial lake boundaries. The primary errors of co-registration and lake outlines that could have resulted in various levels of accuracy were considered in this study. Most of the Landsat scenes have similar resolutions. The glacial lakes are clear on almost all the scenes utilized in the study with less snow and cloud cover, and the manually delineated lake boundaries were checked two times simultaneously by a single operator.
Initially, Landsat ETM+ images were merged with Pan images with high resolution to create a high-resolution pansharpened image by employing the methodology suggested by [41, 42]. All other images were coregistered with the pansharpened ETM+ image within 7.5 m for TM and OLI images, using it as the base map. Consequently, after the image coregistration, geometrical rectification of all images was carried out with the same projected coordinate system of WGS 1984 UTM Zone 43. The remote sensing uncertainty formula [75, 42] was used to estimate the terminus change uncertainty (U).
where a and b are the resolutions of the image and is the coregistration error of the images to the base image in equation (2). We estimated a terminus accuracy of 47.3 m for Landsat TM and ETM+ and 49.6 m for OLI images.
The uncertainties related to lake area have also been estimated through equation (3), as suggested by Yao et al., 2006.
Uarea=2UV (3)
where U and V are the glacier area uncertainty and pixel resolution, respectively.
In this way, the area uncertainties of the glacial lakes were found to be 0.003 km2 (0.3%) for TM and ETM+ and 0.0025 Km2 (0.25%) for OLI images. Thus, the overall uncertainty was estimated to be 0.005 km2 (0.55%), which are well or below the previously reported acceptable ranges (42,41, 82].