Mapping and monitoring glacial lakes in the high mountain ranges are essential due to the vulnerability of the downstream population to the glacial lakes outburst floods (GLOF). The changes in the number and size of glacial lakes are also linked to climate change and it is, therefore, important to map these variations in order to study the impact of climate change [1
]. The high mountain ranges of Hindu Kush, Karakoram and Himalaya (HKKH) contain a large number of glaciers and glacial lakes. The majority of the glaciers in the HKKH have shown negative mass balance in the past decades and the thinning and retreat is higher in the eastern Himalaya as compared to the Karakoram range [2
]. The retreating glaciers, as well as the warming climate, has increased the number and size of glacial lakes. The number of GLOF events have also increased in response to the glacier retreat and thinning. Glacial lakes are of different types e.g., moraine dammed lake, glacial erosion lake, supraglacial lake and ice-blocked lake [3
]. The outburst flood is often triggered by an avalanche depositing into the lake or by the failure of moraine or ice dammed lakes [4
]. The number of lakes as well as the potentially dangerous lakes in this region have been increasing [4
]. Many of the lakes are at a high elevation without any suitable access. A recent GLOF event in the Chitral district of the Hindu Kush range was caused by a newly formed glacial lake at an elevation of 4500 m [7
]. Lakes have also been formed due to the glacier surge in which the glacier blocks the river path and creates a glacier dammed lake [8
]. In the recent past, the surge of the Khordopin glacier and the Shishper glacier in the Karakoram range have created such lakes and the release of water from these lakes caused flooding in the downstream area [8
]. Therefore, it is essential to find automatic methods for mapping surface water using remote sensing imagery.
In the optical remote sensing domain, the freely available imagery from Sentinel-2, Landsat-8 and ASTER can potentially be utilized for the purpose of surface water mapping [12
]. Among these data sources, the Sentinel-2A,2B offers the highest spatial resolution of 10 m along with a shorter revisit time of five days. However, the clouds may prevent a clear image for a longer duration, which may restrict frequent monitoring of the glacial lakes. On the other hand, the cubesat constellation from Planet Labs provides daily global coverage with 3 m spatial resolution and is, therefore a very important data source for mapping and monitoring glacial lakes [10
]. The Planet Labs small satellite constellation consists of 130 + 3U cubesats with a majority of the satellites in Sun Synchronous orbit, capturing scenes in four bands i.e., blue (455–515 nm), green (500–590 nm), red (590–670 nm) and near-infrared (780–860 nm). The PlanetScope imagery has a scene footprint of approx.
km × 8.1 km. The satellites capture scenes after an approximately one-second interval, resulting in a small overlap between consecutive scenes. The main challenges involved in working with the PlanetScope imagery as compared to Sentinel-2 or Landsat-8 is that the PlanetScope imagery consists of only four bands, the imaging sensors in the small satellites have a lower dynamic range and the spectral responses of these satellites show variations even after the sensor calibration as shown in Figure 1
. Therefore, the focus of this study is on the exploration of 4-band PlanetScope imagery for surface water extraction and its accuracy assessment in the context of glacial lakes mapping and monitoring in the HKKH region. The PlanetScope imagery has previously been used in surface water mapping [18
], water velocity computation [19
], bathymetry [20
] and benthic habitat monitoring [22
]. Planet Labs offer limited free data access to researchers and academicians through their education and research program. The data used in this work have been acquired through this education and research program.
The extraction of water in multi-spectral satellite imagery is often implemented using the Normalized Difference Water Index (NDWI) [23
] or Modified Normalized Difference Water Index (MNDWI) [24
]. These indexes are based on the decrease in the reflectance of water in the NIR and SWIR wavelengths as compared to the visible spectrum. NDWI is computed using the normalized difference of the Green and the NIR bands. MNDWI is computed from the normalized difference of the Green and MIR [23
] or the SWIR bands and tends to perform better than NDWI in the built-up areas. Previous studies on the glacial lakes extraction have also exploited NDWI and MNDWI for water extraction in multi-spectral remote sensing satellite imagery [25
]. NDWI has also been used for the extraction of water in the PlanetScope imagery [18
]. Normalized differences using different band combinations have also been studied for water fraction mapping [27
]. The threshold used for classifying water from the rest of the pixels is often tuned based on the scene. The selection of the NDWI or MNDWI threshold is not a trivial task because lakes, even in close proximity, can have a varying spectral signature depending on the turbidity, composition and depth of lake water [28
]. The cast shadows pose as one of the main challenges in the mapping of the water pixels in the mountainous regions. In such mountainous terrain, the slope derived from the DEM has also been utilized to remove errors due to shadows [29
]. The pixel-based image classification techniques [30
] and object-based segmentation [32
] have also been used for mapping of surface water. The global surface water explorer [30
] is an automated system that can be used for global mapping of surface water. However, it is based on the Landsat series, so it has limitations in monitoring glacial lakes dynamics because of a longer revisit time. Due to difficulties in the automated mapping of the glacial lakes, there exists no continuously-updated glacial lake inventory for this region. The previous glacial lake inventories have been developed using manual digitization or require significant human intervention making frequent updates difficult [29
The recent dawn of deep learning has produced remarkable results in various disciplines and domains [37
]. The common deep learning architectures for image recognition consists of multiple convolution layers and are known as Convolutional Neural Networks (CNN). The convolutional filters are learned from the labeled data during the training process. CNN have been used for various classification, segmentation and object detection tasks in remote sensing imagery [40
]. For pixel-wise image classification or semantic segmentation, the input to the CNN is an image and the output is often an image of the same size with class labels for each pixel. This type of CNN are often called fully convolutional networks as they contain convolution layers throughout the network instead of the fully connected layers. The U-Net architecture [41
] originally presented for biomedical image segmentation has achieved state-of-the-art results for different image segmentation problems like building segmentation, roads extraction and other tasks related to remote sensing image segmentation. U-Net model has also been used by some of the top-performing architectures in the recent DeepGlobe challenge [42
], Spacenet challenges [43
] and IGARSS Data Fusion contest [44
] comprising of road extraction, building segmentation, landcover and semantic 3D reconstruction from satellite imagery. Due to the difficulties in the automated mapping of surface water, in this work, we explore the potential of CNN in automated extraction of surface water in 4-band PlanetScope imagery in the context of the glacial lake monitoring in HKKH.
The results presented above show that the deep learning-based model can be employed in practice for the automatic extraction of glacial lakes in PlanetScope imagery. The results show that even with the variations in the images from different PlanetScope satellites, the deep learning model was able to generalize and perform well on the unseen PlanetScope images of a different area. We have observed some false positives on pixels on the cloud edges during our evaluation as shown in Figure 16
. This can perhaps be solved by using the cloud masks provided in the PlanetScope supplementary data or to include clouds in the training data. False positives were also observed due to cast shadows in the images of Baltoro and Shishper glacier. We have also observed that some lakes with muddy brown color were not extracted by the U-Net, which shows that the training data need expansion to include all possible types of lakes that may occur. Another important point to consider here is that the larger lakes are sometimes easier to map than the relatively smaller lakes and the larger lakes have a higher weightage on the resulting metrics because it contains more pixels i.e., the evaluation metrics are influenced more by the larger lakes. Therefore, one may consider computing evaluation metrics for different lake sizes.
In the generation of the labeled data, the relative displacement of water bodies between the VHRS imagery and PlanetScope imagery was a major issue and required a lot of time to correct. Due to a lower resolution of the PlanetScope imagery, it was very difficult sometimes to determine this shift and as a result, even the labeled data will have some inaccuracies. When generating the labeled data, delineating the boundary of very shallow water bodies is also very difficult and the delineation is dependent on the interpretation of the digitizer. Some small lakes were missed in the original digitization process, but the trained U-Net was able to extract those lakes. Thus, we used the U-Net model to track errors in the digitization process and correct those errors and train the network again. The precise delineation of the supraglacial lakes is very challenging due to their smaller size, mixture with snow and debris and cast shadows. The changes in the extent of the supraglacial lakes within a short time period also makes it difficult to use and compare images of different dates.
The GLOF event in the Chitral district shows the importance of frequent mapping of the glacial lakes because new lakes can be formed, which may suddenly drain after some time. Such events have occurred in the past and it is imperative to monitor such situations using remote sensing datasets. The automation of the whole process is essential as this region spreads over a large area and it is not possible to visually inspect the whole area. It should be mentioned that the occurrence of GLOF depends on various factors. In this work, we only attempted to map the lake area in a dynamic scene without any consideration of the outburst probability. The damming of the glacial lakes in one of the important factors in designating a glacial lake as a potentially dangerous lake. Such classification of the lakes was also not studied in this work.
In the case of the Shishper glacier surge the volume of the lake changed over time and frequent mapping helps to determine the water quantity and compute simulations of water discharge and possible flood scenarios. It should be mentioned that, due to the presence of clouds, no PlanetScope scenes were available on the day the lake in Chitral and Shishper drained. Therefore, due to the limited availability of cloud-free images in parts of HKKH, any solution to monitor surface water from remote sensing satellites should also incorporate both optical and microwave satellite data in the future, especially due to the recent progress by the commercial companies to develop a constellation of Synthetic Aperture Radar (SAR) satellites [64
]. The inclusion of the microwave remote sensing data for glacial lakes mapping is inevitable.
The labeled data generated in this work covered only the HKKH region. Assessing the transferability of the model to a different study area is also interesting. Here, we map the glacial lakes in the Peruvian Andes using the PlanetScope imagery and the trained U-Net model as shown in Figure 17
. For a comparison, the glacial lake outlines of Cordillera Blanca (Peru) [65
], which have been derived from visual interpretation of satellite imagery are also shown as a reference. A visual analysis of the results shows that the trained U-Net model generalizes well to the other areas.