Next Article in Journal
Formation and Hazard Analysis of Landslide Damming Based on Multi-Source Remote Sensing Data
Previous Article in Journal
AOGC: Anchor-Free Oriented Object Detection Based on Gaussian Centerness
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 15
Issue 19
10.3390/rs15194687
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Supervised Geomorphic Mapping of Himalayan Rivers Based on Sentinel-2 Data

by
Zarka Mukhtar
1,*,
Simone Bizzi
2 and
Francesco Comiti
1
1
Faculty of Agricultural, Environmental and Food Sciences, Free University of Bozen-Bolzano, 39100 Bolzano, Italy
2
Department of Geosciences, University of Padova, 35122 Padova, Italy
*
Author to whom correspondence should be addressed.
Remote Sens. 2023, 15(19), 4687; https://doi.org/10.3390/rs15194687
Submission received: 17 July 2023 / Revised: 11 September 2023 / Accepted: 21 September 2023 / Published: 25 September 2023
Browse Figures
Review Reports Versions Notes

Abstract

:
The Himalayan region is a hotspot in terms of expected future hydrological and geomorphological variations induced by climate change on proglacial areas and the related implications for human societies established along the downstream rivers. Due to the remoteness of the proglacial zones in the Himalayas and the associated logistical problems in carrying out traditional field and UAV-based morphological monitoring activities, remote sensing here plays a crucial role to monitor past and current fluvial dynamics, which could be used to anticipate future changes; however, there has been, so far, limited research on morphological changes in Himalayan proglacial rivers. To address this gap, a morphological classification model was designed to classify recent changes in Himalayan proglacial rivers using the Google Earth Engine platform. The model is the first of its kind developed for the Himalayan region and uses multispectral S-2 satellite data to delineate submerged water channels, vegetated surfaces, and emerged, unvegetated sediment bars, and then to track their variations over time. The study focused on three training sites: Langtang-Khola (Nepal), Saltoro (Pakistan), and Nubra (Jammu and Kashmir) rivers, and one testing site, the Ganga-Bhagirathi River (India). A total of 900 polygons were used as training samples for the random forest classifier, which were further divided into 70% calibration and 30% validation datasets for the training sites, and a separate validation dataset was acquired from the testing site to assess the model performance. The model achieved high accuracy, with an average overall accuracy of 96% and a kappa index of 0.94, indicating the reliability of the S2 data for modeling proglacial geomorphic features in the Himalayan region. Therefore, this study provides a reliable tool to detect past and current morphological changes occurring in the Himalayan proglacial rivers, which will be of great value for both research and river management purposes.

Graphical Abstract

1. Introduction

Proglacial areas and their downstream river channels are fast-changing environments worldwide due to the current rapid global warming [1,2]. Proglacial rivers have been undergoing morphological changes over the last several decades [3,4,5] and are expected to markedly evolve in the future under the projected climatic scenarios [2], in parallel to changes in their hydrological regime [6]. The morphological characteristics (i.e., channel width, channel pattern, abundance and spatial distribution of riparian vegetation) are well known to significantly influence rivers’ ecosystems [7,8], as well as present potential hazards [9]. Therefore, frequent morphological monitoring of proglacial streams and of their downstream reaches is of the utmost importance to detect ongoing evolutionary trajectories [10] and to anticipate future changes in both aquatic and riparian ecosystems’ functioning and in the potential flood and geomorphic hazards for the riverine human communities; however, frequent morphological monitoring of the headwaters of glaciated catchments by traditional methods is challenging due their generally low accessibility and terrain complexity [11,12,13]. Such issues are extreme in the case of Himalayan river systems, where proglacial reaches can be very extended in length and width, featuring several low-flow channels (i.e., featuring a very high braiding index), and present quite difficult access along with harsh environmental conditions. Consequently, in such remote and vast areas, the use of UAV for morphological monitoring suffers from important limitations, and satellite-based remote sensing becomes extremely valuable.
In recent years, remote sensing techniques (RS) have emerged as powerful tools to study fluvial systems [14,15], and they have led to tremendous changes in how fluvial geomorphological analyses are conducted [16,17]. In any case, satellite-based RS may be a suitable approach to monitor morphological changes in rivers of sufficient width [18]; however, the limited channel width of most mountain rivers—including those draining the Himalayas—poses important limitations in the use of low-resolution satellite images, as those acquired, for example, by the former Landsat missions. Among the various satellite imageries, the Sentinel-2 satellite mission (hereafter referred to as S-2), launched by the European Space Agency (ESA) in 2015, offers great potential for river scientists [19]. With its multispectral sensors, S-2 provides 10–60 m resolution optical images of the Earth’s surface, covering a wide range of spectral bands. These freely available images have already been used to investigate several fluvial landforms and dynamics [20,21,22,23], such as the flow rate variability [24], sediment concentration [2], and, recently, the mean grain size on exposed sediment bars [25], mostly using supervised classification models. In addition, recent web-based platforms such as Google Earth Engine (GEE) have revolutionized remote sensing applications [26]. In fact, GEE holds the catalogues of freely accessible multivariate datasets and allows users to bring their own algorithms to the satellite, environmental, and geospatial datasets. GEE is a fundamental attempt to promote remote sensing applications and reconstructable code sharing for the monitoring of change in fluvial geomorphology [27].
Supervised classification models are a type of machine learning algorithm that can be trained using labeled data to accurately classify different features in the imagery [28,29]. By using a supervised classification model based on S-2 data, it is possible to identify and map different features in the proglacial river landscape, such as water [30], vegetation [31,32], and sediments [33]. This information can then be used to investigate changes in river dynamics over time, such as changes in channel morphology, sediment deposition and erosion, and vegetation growth [34,35]. The use of an S-2 based supervised classification model has the potential to provide valuable insight into proglacial river dynamics in the Himalayas by allowing researchers to monitor how such systems are evolving due to the ongoing global warming, thus enabling scientists to anticipate what their future geomorphological and ecological changes could be (e.g., channel width variations, riparian vegetation dynamics), which in turn may impact local and downstream human communities in terms of the flood hazards (e.g., increased bank erosion and/or augmented wood transport) and the stability of road and water intake infrastructures.
To the best of our knowledge, S-2-based supervised classification models that could be utilized for the morphological mapping and monitoring of proglacial Himalayan rivers have not been developed yet. The aim of the present work was to develop such a supervised classified model, utilizing the S-2 images and the new geospatial advancement of GEE. The goal was to accurately classify the areas within proglacial Himalayan fluvial corridors into three of the geomorphic macro-units defined in the Geomorphic Unit System (GUS) proposed by Belletti et al. (2017) [36], i.e., base-flow channels (submerged), emergent sediment units (unvegetated), and in-channel/riparian vegetation (emerged landforms covered with woody vegetation). The model was developed and trained in three rivers and then applied to a different river to properly evaluate its performance.

2. Study Areas and Available Satellite Images

2.1. Study Areas

The supervised classification model was developed and tested in four proglacial rivers (Figure 1) draining the Karakorum–Himalayan range: Saltoro, Nubra, Ganga-Bhagirathi, and Langtang-Khola Rivers, respectively located in Pakistan, the disputed Jammu–Kashmir region, India, and Nepal. The Nubra and Saltoro Rivers are two of the major tributaries of the Shyok River [37,38].
The Saltoro River (AOI-1) originates from the Saltoro Mountains, a subrange of the Karakoram located within the Kashmir region of Pakistan, and drains the Gyong and Sherpi Gang glaciers. In this work, an area of about 25 km2 of the Saltoro River was investigated, extending from 35°12′26″N to 76°31′12″E.
The Nubra River (AOI-2) originates from the longest Himalayan glacier, i.e., the Siachen glacier, roughly 5400 m a.s.l. in the Nubra valley. The total catchment area at the terminus of the Siachen Glacier is 2056 km2. The studied part of the Nubra River covers an area of about 112 km2, extending from 34°44′27″N to 77°33′10″E.
In Nepal, we selected the Langtang-Khola River (AOI-3) as the study area, which drains the Langtang range and is located in the Khatmandu region. The Langtang Glacier is the largest glacier in Langtang Valley with an area of 46.5 km2 and an elevation of 5579 m a.s.l. [39]. The Langtang River basin drains an area of about 354 km2 [40]. About 25 km2 of the proglacial segment of the Khola River was examined, from 28°12′90″N to 85°38′4″E.
Finally, one of the study areas was chosen in the Indian Himalayas, i.e., the Ganga-Bhagirathi River (AOI-4), which drains the Gangotri Glacier featuring a current area of about 122 km2 [41]. The total catchment area of the Ganga-Bhagirathi River basin is around 7811 km2. The investigated part in this study lies in the upper proglacial part of the Ganga-Bhagirathi River basin, extending from 31°1′79″N to 78°46′55″E and covering almost 25 km2.

2.2. Datasets Used

Two types of remotely sensed datasets were used in the study: Sentinel-2 (S2) and ALOS PALSAR DEM. A multispectral instrument aboard the S2 satellite provides optical satellite images with 10 m, 20 m, and 60 m resolution. S2 was launched in 2015, and measures 13 individual spectral bands varying from visible to short-wave infrared, with wavelengths from 493 nm to 1374 nm. The green (B3) and near-infrared (B8) bands of Sentinel-2 were used as input variables for training the algorithm. According to the normalized difference water index (NDWI), the reflectance factor of the water’s surface increases by the green band, and the near-infrared band increases the reflectance of vegetation and soil [42]. This band combination highlights open water features and allows a water body to stand out against soil and vegetation.
In order to develop a fluvial morphological classification model, the delineation of an extended fluvial corridor (i.e., the portion of main valley bottom shaped by current fluvial processes and presenting alluvial landforms, thus including active channels, islands, floodplains, and recent terraces, as well as active alluvial fans built by the tributaries) in each of the four AOI was first necessary. Such an extended fluvial corridor is bordered by much steeper landforms corresponding to hillslopes, moraines, and/or ancient terraces. The mapping of the fluvial corridors was performed manually within ArcGIS Pro software (version 3.1), integrating the evidence provided by digital terrain models and S2 satellite images. The freely available 12.5 m resolution Advanced Land Observing Satellite (ALOS) Phased Array Type L-Band synthetic Aperture Radar (PALSAR) DEM 2011, acquired by an ASF data search (alaska.edu), was processed to generate contour lines, which were crucial to manually delineate the border between the gentle-sloping alluvial landforms and the steeper colluvial and glacial landforms.
Remotesensing 15 04687 g001
Figure 1. Location and images (S-2 multispectral) of the 4 study areas: AOI-1 Saltoro River; AOI-2 Nubra River; AOI-3 Langtang-Khola River; AOI-4 Ganga-Bhagirathi River.
Figure 1. Location and images (S-2 multispectral) of the 4 study areas: AOI-1 Saltoro River; AOI-2 Nubra River; AOI-3 Langtang-Khola River; AOI-4 Ganga-Bhagirathi River.
Remotesensing 15 04687 g001

3. Methodological Approach and Workflow

The purpose of this study was to develop and test a model to delineate fluvial morphological macro-units in proglacial Himalayan rivers, following the definition provided in the Geomorphic Unit Systems (GUS) recently proposed by Belletti et al. (2017) [36]. Such macro-units, comprising the fluvial corridor identified as explained in Section 2.2, include the following: (i) submerged base-flow channels; (ii) emerged and unvegetated sediments (bars); (iii) and vegetated surfaces located within the channel (islands and banks) or at its margins (floodplains and recent terraces). Together, the submerged base-flow channels and the emerged unvegetated bars constitute the active channel (or just channel) of a river, which is shaped and maintained by the interplay between frequent water flows, ordinary flood events (recurrence interval less than about 2 yr), and bedload transport processes. The frequent flow and sediment-related disturbances occurring within the active channel inhibit the growth of stable, pluriannual vegetation inside it. Perennial vegetation can instead encroach and stabilize less frequently disturbed areas, either due to their higher elevation compared to the thalweg (islands and banks) or their distance from the fast flood flows occurring in the channel (floodplain and terraces). Vegetated surfaces are nonetheless inundated regularly during flood events but are subject to important geomorphic changes—excluding the case of progressive bank erosion—only during high-magnitude, infrequent events [43]. The capability to remotely monitor changes in the relative proportion of these three macro-units in proglacial rivers enables us to understand how water and sediment fluxes are being modified by the progressive deglaciation, as well as by variations in the magnitude and frequency of rainfall events.
The classification model was developed within the GEE platform, and it used a supervised classifier algorithm applied to the S-2 images, where the images were treated as a reference layer for calibrating and verifying the classification model.

3.1. Pre-Processing Platform

The GEE data catalogue helped us to access and process the image collection and filter between 2019 and 2021, given that, for the study areas, images were not available before 2019. The level-2 surface reflectance (SR) product was also available. The image collection was filtered to retain only those images with a maximum cloud coverage of 15%, resulting in 12, 3, 35, and 15 images for the Nubra, Langtang-Khola, Ganga-Bhagirathi, and Saltoro Rivers, respectively. Further filtering was performed by selecting one suitable image per year, considering the one with lowest coverage of cloud and mountain shadow. The 10 images selected for the analysis in the 4 AOI were acquired between the months of August and September, depending on the availability of images. The dates of acquisition are provided in Table 1. During the entire melt season of 6 months (May-October), July and August are considered as the peak melt period, and the water level fluctuates to a great extent. In such circumstances, there is a leeway of errors in flow dimensions [44].
To access the S2 collection, the cloud cover was filtered, and specific bands (B3 and B8) were selected, thus enabling the development of a code in GEE. Then, the code continued to develop an image classification model using a random forest classifier.

3.2. Model Development for Multitemporal Image Classification

The model used to classify the 3 morphological macro-units was developed on the GEE platform, as shown in Figure 2. The workflow was developed in 4 steps:
1. Selection of morphological macro-units;
2. Filtering of the training images and development of the training data;
3. Image classification;
4. Accuracy assessment.
Remotesensing 15 04687 g002
Figure 2. The workflow of the data processing. It is divided into two sections. The first step is to calibrate the model for the supervised river classification mapping, and the second step is to validate the performance of model.
Figure 2. The workflow of the data processing. It is divided into two sections. The first step is to calibrate the model for the supervised river classification mapping, and the second step is to validate the performance of model.
Remotesensing 15 04687 g002

3.2.1. Definition of Morphological Macro-Unit Classes

A morphological unit is a landform created and modified by a specific set of earth processes [10]. River morphological units are located within a river channel or on a flood plain shaped by erosion and deposition of sediments or by bedrock [36]. According to the Geomorphic Units survey and classification System (GUS) proposed by Belletti et al. (2017) [36], a macro-unit covers the course assemblage of the same type of spatial units, mainly water, sediment, and vegetation, including many more sub-units (e.g., backwater areas, ramps, small woody patches) depending on their level of characterization. Table 2 presents the 3 macro-units of interest in this study.
The macro-unit “Vegetated surface” includes all the areas where vegetation is present within the fluvial corridor, without any purpose of distinguishing it further into taxonomical (tree/shrub species) or structural (height) classes. One issue is that S-2’s resolution enables the detection of small-sized ponding water patches present within vegetation zones. These patches can disrupt the process of classification of morphological macro-units. Therefore, to distinguish between only water pixels and mixed water and vegetation pixels, different training polygons were created (Figure 3), grouping base-flow channel and dry channel as one feature and vegetation pixels and mixed vegetation plus water pixels as the other. The wet and dry channels were grouped in the same macro-unit category of the base-flow channels. The remaining macro-unit comprises emerged sediment areas without any substantial vegetation cover, mostly corresponding to what are known as sediment bars in fluvial geomorphology [36]. In such macro-units, however, it is possible that single individuals or small clusters of herbaceous or young woody plants (pioneering shrubs) are present; however, the negligible vegetation cover of the sediment bars—not reliably detectable through S-2 images—indicates the occurrence of active bedload transport processes on these morphological units, in addition to low-flow channels [9]. It is worth stressing again how the delineation of the 3 macro-units described above is key to quantify the spatial extension over which bedload transport processes take place within the fluvial corridor and tracking its changes through time (especially contrasting vegetated surfaces against low-flow channels and unvegetated bars) makes it possible to detect the relative variations of the hydrological and sediment supply induced by climate change in the Himalayan river basins.

3.2.2. Filtering Training Images and Developing Training Data

Three images from the Nubra, Saltoro, and Langtang-Khola Rivers, acquired in 2020 (which is the year when an image is available for all the selected locations), were selected for model training. Amongst the images available for each location, we selected those with acquisition dates closest in time. The months of August and September belong to the wet season, according to the monsoon period in the Himalayan Belt, which varies in intensity from east to west [45]. The images from the end of the rainy season selected for the collection of training data were taken on 17 August 2020 (Nubra, Figure 4), 29 September 2020 (Saltoro, Figure 5), and 25 August 2020 (Langtang-Khola, Figure 6). These images were used to take sample polygons for the calibration of the classification model.
We have developed a single classifier, merging all the calibrations datasets from different dates and sites. The one image assigned to the classifier is from the most complex and largest proglacial river location within the selected sites, i.e., the Nubra River, acquired on 17 August 2020 (Figure 4).
The training data are crucial to evaluate the accuracy of the remote sensing data. The training sample collection was accomplished through manual interpretation of the satellite images, where simple random sampling was carried out (Figure 3). The three macro-unit classes were easily detectable, and the manual delineation was intensively resource-demanding. The proportions of the sample size for each macro-unit class and image correspond to the idea of balance and the area of the study sites, respectively. There are therefore more sample polygons from the Nubra River, which is 4 times wider (500 polygons from 112 km2 area), than the other sites (200 polygons from 25 km2 area of each). Along with this, the pixel quantity also varied significantly because the polygon sizes are bigger for the wider rivers of Nubra and Saltoro than the rivers with a narrow channel (for example Langtang-Khola River). In total, 900 sample polygons were delineated from the training sites (Nubra, Saltoro, and Langtang-Khola Rivers) with at least 50 training samples per landcover class, and these were divided into calibration and validation datasets for each of the locations (Table 1). Keeping the balance of the validation dataset division on track, 60 polygons were taken separately to validate the performance of the model on the testing site of the Ganga Bhagirathi River.
Remotesensing 15 04687 g003
Figure 3. Sample polygons taken for 3 different macro-units: (a) Nubra River; (b) Ganga-Bhagirathi River.
Figure 3. Sample polygons taken for 3 different macro-units: (a) Nubra River; (b) Ganga-Bhagirathi River.
Remotesensing 15 04687 g003
The collection of the training data was performed within the GEE platform. The dataset (which includes 5452 pixels) was split into two subsets: 70% was used for the model’s calibration, while the remaining 30% was used for validation. The calibration samples were used to adjust the supervised classifier random forest (RF) algorithm (see Section 3.2.3), and the validation dataset was used in the accuracy assessment of the classified maps (see Section 3.2.4).

3.2.3. Image Classification

Random forest (RF) is a widely applied machine learning algorithm that combines multiple decision trees to create a more robust and accurate model [46]. It is a type of ensemble learning method that involves aggregating the predictions of several weaker models to create a stronger, more accurate one. In a random forest classifier, each decision tree is trained on a subset of the features and data points, and the final prediction is made by aggregating the predictions of all the trees [47]. RF fuels the predictive accuracy and monitors over-fitting by averaging the outcomes from each tree and using the majority of votes for prediction. In this study, RF was applied to obtain the morphological classification for each available year. For the other controlling factors, this study utilized default GEE parameters to train the algorithm. For example, n_estimator is the number of trees (30) we want the algorithm to create, minimum leaf population (1) specifies the sample amount that a node must hold after being split, and bag fraction (0.5) is a subsampling fraction randomly selected to propose the next tree in the expansion [48].
As described in Section 3.2.1, the classifier was trained on three images from 2020 and then used to classify the remaining 7 available images. We developed the model to produce the morphological macro-unit classification for the years 2019, 2020, and 2021 for the Nubra, Ganga-Bhagirathi, and Saltoro Rivers, and for 2020 for the Khola River. Out of a total of 10 images, 7 were classified with 70% of the training dataset and validated with 30% of the training dataset. The remaining 3 images of the Ganga-Bhagirathi River were classified with a designed classification model as a testing of the model’s performance in locations that were not included in the calibration of the model. The results were analyzed and validated through the validation dataset (Table 3), as well as a visual interpretation of the original images (Figure 4, Figure 5, Figure 6 and Figure 7).

3.2.4. Accuracy Assessment

To quantify the model’s performance, the accuracy assessment of the supervised macro-units’ classification was made through a confusion or error matrix. As mentioned above, the validation dataset (30% of the total sampled pixels from training sites and 231 pixels from testing site) was used to test the performance of the classifier. A 2D confusion matrix of the morphological macro-unit classes was calculated with the computed metrics of overall accuracy (OA) and kappa index in Table 4. The OA is the percentage of pixels correctly classified over the totality, and the kappa index measures the agreement between classification and the truth value, where 1 represents perfect agreement and 0 represents no agreement. Table 4 also presents the producer’s accuracy (PA), which measures the error of omission; this refers to the accuracy of the model in correctly excluding pixels that do not belong to a specific class or category. The user’s accuracy (UA) is also provided, which measures the error of commission; this refers to the accuracy of the model in correctly identifying pixels that belong to a specific class or category [49].
Remotesensing 15 04687 g004
Figure 4. Nubra River, Siachen Glacier, Jammu and Kashmir. All the sections present classified (A.1–A.3) and original S-2 (B.1–B.3) images from 2019, 2020, and 2021. Original S-2 images are the combination of green and near infrared bands highlighting water (blue color) and vegetation (red color).
Figure 4. Nubra River, Siachen Glacier, Jammu and Kashmir. All the sections present classified (A.1–A.3) and original S-2 (B.1–B.3) images from 2019, 2020, and 2021. Original S-2 images are the combination of green and near infrared bands highlighting water (blue color) and vegetation (red color).
Remotesensing 15 04687 g004
Remotesensing 15 04687 g005
Figure 5. Saltoro River, Gyong Glacier, Pakistan. All the sections present classified (A.1–A.3) and original S-2 (B.1–B.3) images from 2019, 2020, and 2021. Images from 2019 have two zoomed blocks in each classified map (a_3.1) and the S-2 original (b_3.1) image. Original S-2 images are the combination of green and near infrared bands highlighting water (blue color) and vegetation (red color).
Figure 5. Saltoro River, Gyong Glacier, Pakistan. All the sections present classified (A.1–A.3) and original S-2 (B.1–B.3) images from 2019, 2020, and 2021. Images from 2019 have two zoomed blocks in each classified map (a_3.1) and the S-2 original (b_3.1) image. Original S-2 images are the combination of green and near infrared bands highlighting water (blue color) and vegetation (red color).
Remotesensing 15 04687 g005
Remotesensing 15 04687 g006
Figure 6. Langtang-Khola River, Langtang Glacier, Nepal. (A,B) present classified map and original S-2 image, respectively, from 2020. The images have two zoomed blocks in each classified map (a.1,b.1), and S-2 original (a.2,b.2) represents misclassified water patches. Original S-2 image is the combination of green and near infrared bands highlighting water (blue color) and vegetation (red color).
Figure 6. Langtang-Khola River, Langtang Glacier, Nepal. (A,B) present classified map and original S-2 image, respectively, from 2020. The images have two zoomed blocks in each classified map (a.1,b.1), and S-2 original (a.2,b.2) represents misclassified water patches. Original S-2 image is the combination of green and near infrared bands highlighting water (blue color) and vegetation (red color).
Remotesensing 15 04687 g006
Remotesensing 15 04687 g007
Figure 7. Ganga-Bhagirathi River, Gangotri Glacier, India. All the sections present classified (A.1–A.3) and original S-2 (B.1–B.3) images from 2019, 2020, and 2021. Images from 2020 and 2021 have two zoomed blocks in each classified map (a.1,a.2) and S-2 original (b.1,b.2) images. Original S-2 images are the combination of green and near infrared bands highlighting water (blue color) and vegetation (red color).
Figure 7. Ganga-Bhagirathi River, Gangotri Glacier, India. All the sections present classified (A.1–A.3) and original S-2 (B.1–B.3) images from 2019, 2020, and 2021. Images from 2020 and 2021 have two zoomed blocks in each classified map (a.1,a.2) and S-2 original (b.1,b.2) images. Original S-2 images are the combination of green and near infrared bands highlighting water (blue color) and vegetation (red color).
Remotesensing 15 04687 g007
Table 4. Confusion matrix with three morphological macro-unit classes (numbers in the matrix are pixels of the validation datasets); vegetated surfaces (veg), base-flow channels (water), and emerged unvegetated sediment units (sed) for the years 2019, 2020, and 2021. MD: month and day; UA: user accuracy; PA: producer accuracy.
Table 4. Confusion matrix with three morphological macro-unit classes (numbers in the matrix are pixels of the validation datasets); vegetated surfaces (veg), base-flow channels (water), and emerged unvegetated sediment units (sed) for the years 2019, 2020, and 2021. MD: month and day; UA: user accuracy; PA: producer accuracy.
Error Matrix201920202021
Nubra RiverMD 09/20VegWaterSedUAMD 08/17VegWaterSedUAMD 08/25VegWaterSedUA
Veg450800.98Veg458001Veg458001
Water0263270.9Water029001Water028820.99
Sed011660.99Sed001671Sed001671
PA10.960.860.96PA1111PA110.980.99
Saltoro RiverMD 08/16VegWaterSedUAMD 08/25VegWaterSedUAMD 09/19VegWaterSedUA
Veg134001Veg134001Veg134001
Water083180.82Water010010.99Water068330.67
Sed001031Sed001031Sed08950.92
PA110.850.95PA110.990.99PA10.890.740.87
Langtang-Khola River MD 09/29VegWaterSedUA
Veg339001
Water04920.96
Sed00411
PA110.950.99
Ganga-Bhagirathi RiverMD 08/16VegWaterSedUAMD 08/25VegWaterSedUAMD 09/19VegWaterSedUA
Veg1171297.5Veg1173097.5Veg1181198.3
Water064691.4Water07001Water064691.4
Sed083380.4Sed00411Sed00411
PA187.680.492.6PA195.8198.6PA198.485.496.5

4. Results

4.1. Image Classification

Figure 3, Figure 4, Figure 5 and Figure 6 report the RF algorithm results and performance for the different AOI. A visual inspection of these figures shows how the model successfully separated, on average, the three morphological macro-units (see Figure 4). As expected, the highest degree of errors occurred in the border area between the macro-units, particularly when their size was smaller than a single pixel (100 m2). This is the case where the base flow channels are numerous, small in width, and separated by small patches of emerged sediments (e.g., Figure 5 zoomed-in areas and Figure 6(a.2,b.2), respectively).
We evaluated the macro-unit extraction accuracy by comparing the classified maps with the original images. In Figure 6(a.1,b.1), the algorithm provides master results by delineating very small (about 724 m2) unvegetated sediment bars. Similarly, Figure 7(a.2,b.2) shows the accuracy of the model’s performance by capturing even shallow river channels and emerged unvegetated sediment bars.
During the calibration of the model, two types of mixed pixels (“water and vegetation” and “water and sediments”) were grouped in polygons to train them as vegetated surfaces and submerged water channels, respectively. Due to some limitations (see Section 5) in a few areas of the rivers, the model could not correctly classify the pixels. Indeed, the analyzed river features quite a large number of mixed water and sediment pixels, which are challenging to distinguish. Figure 5(a_3.1,b_3.1) defends the previous statement and highlights the bad performance of the model on the borders of river channels and sediment bars. This means that the manual interpretation of the classified image with the original satellite image clearly represents that the banks of river braids are misclassified as part of sediment bars. Similarly, in Figure 6(a.2,b.2), the vegetated zone adjacent to the channel is misclassified as water.

4.2. Accuracy Assessment

The confusion matrices are presented in Table 3. From the table, it can be observed that the overall accuracy ranges between 87% and 100%. Most of the classified maps demonstrate the efficiency of the model in extracting morphological macro-units, with an overall accuracy and kappa coefficient of around 96% and 0.94, respectively. The lowest accuracy (87%) was obtained in the Saltoro River in 2021, which is nonetheless still a very good result [41]. The kappa coefficients range between 0.92 and 1, with the exception of 0.82 for the Saltoro River in 2021. As described above and in Table 3, a considerable amount of water pixels were misclassified as sediments in 2019 and 2021. On average, the classified images demonstrate 96% and 0.94 overall accuracy and kappa index, respectively.
The accuracy of the classification is reported in Table 4, which shows that small patches of the river are not classified as a part of base-flow channels, but as emerged sediments. For example, the Nubra River in 2019 had 27 misclassified pixels as sediments and the Saltoro River in 2019 and 2021 had 18 and 33 misclassified pixels, respectively. In the cases of Nubra 2019 and Saltoro 2021, few pixels of vegetation (1.7%) and sediments (7.8%), respectively, are categorized as base-flow channel. As mentioned earlier, these are the misclassified pixels with a mixed reflectance signature of water and sediment. The purpose of this study was to eliminate the mixed pixel confusion of sediments in low flow periods (Section 3.2.1) and mountain/cloud shadow effects, although, the accuracy obtained in the classifications shows that the uncertainty among the macro-units is negligible.
As already mentioned in Section 3, AOI-4 was selected to test the model performance in a different river where training was not performed. The model that was trained in the three selected sites of this study (A0I-1, AOI-2, and AOI-3) was applied to the Ganga-Bhagirathi River without taking any training samples from this location to calibrate the model. Figure 7 presents the highly satisfactory results of this test on the Ganga-Bhagirathi River, where the qualitative performance can be seen by comparing the classified image against the original image. It can be observed that both the low-flow channels and emerged sediment bars are most correctly classified. The overall accuracy for this river turned out to be 98.6%, based on 60 polygons delineated for such a purpose. An error matrix calculation carried out on the selected areas (Figure 7(a.1–b.1)) provides quite a high level of accuracy, up to 96.5%. Nonetheless, the visual inspection in the same sections indicated the presence of some misclassified pixels in all three morphological macro-units; however, the problematic areas correspond to the narrowest section of this AOI, where S2’s resolution may be the limiting factor for an excellent macro-unit delineation.

5. Discussion

5.1. Novelty and Impact of the Proposed Model

This study presents a supervised classification model for delineating fluvial macro-units (low-flow channels, emerged sediment bars, and vegetated surfaces) in proglacial rivers of the Himalaya, based on 10 m resolution Sentinel-2 images. Although supervised models able to classify water, vegetation, and bare sediments in fluvial corridors have been already proposed [50,51,52], to the best of our knowledge, this is the first machine learning, supervised, morphological macro-unit classification model developed for proglacial streams in the Himalayan region, which is a hotspot area where ongoing deglaciation is predicted to determine marked variations in terms of runoff, sediment supply and, thus, river dynamics [53]. The model was developed using different river morphological patterns, from wide braided rivers (Nubra and Saltoro) to relatively narrow channels exhibiting a transitional pattern (Langtang-Khola).
Remarkably, the accuracy assessment of the classified images (Section 4.2) demonstrated that the model can be applied to Himalayan proglacial rivers different from those on which the model was trained. The model was in fact tested on the relatively narrow (around 240 m) Ganga-Bhagirathi River in three different images (2019, 2020, and 2021). The results are highly satisfactory, with an average 96% OA. Nonetheless, the model’s performance is better in wider river channels, such as the Nubra and Saltoro Rivers, as narrow river channels present more shadowing effects from the hillslopes, which exacerbate the inherent uncertainties in water–sediment pixel classification. Therefore, the classification results can be used to accurately track fluvial morphological changes in the proglacial rivers of this region. The analysis can be re-run as new RS layers are automatically acquired by GEE, which will help to identify progressive or sudden variations in a river’s size and pattern following deglaciation or single disturbance events (landslides, GLOF). One key aspect of our model is that it is based on freely available satellite images, thus making the monitoring of river dynamics affordable for any research institute or management agency in the Himalayan countries.

5.2. Limitations of the Model

The supervised morphological macro-unit classification model was trained to classify pixels based on the spectral characteristics of the features present in the image; however, mixed pixels occur when a pixel contains more than one feature, such as a pixel containing both vegetation and sediment or water. This can result in classification errors, as the model may assign the pixel to only one class, while it actually contains multiple classes. Indeed, we verified (Figure 8) how many misclassification errors occur at the borders between different macro-units, especially at the margin of vegetated surfaces, due to the presence of an overhanging canopy on channel macro-units. Other errors were observed for small ponding areas corresponding to minor, hydrologically disconnected anabranches.
It is worth reporting here that Carbonneau et al. (2020) [19] compared automatically classified macro-units based on S-2 images against macro-units manually classified based on UAV images and observed that the intrinsic spatial error in edge pixels associated with the S2 image (resolution 10 m) ranges from 20 to 30%. Similarly, a later study by Bozzolan et al. (2023) [54] applied the model by Carbonneau et al. (2020) and compared its results to manually classified orthophotos (30 cm resolution) and planet images (3 m resolution). Their results confirmed an intrinsic spatial error associated with the S2 images of 20–30% for mixed pixels. Nonetheless, they argued that the high revisit frequency of S2 makes its images more informative in terms of ongoing geomorphic processes than higher resolution multiannual images, as long as the observed changes of the macro-geomorphic units are larger than the associated error of 20–30%. For instance, if the amount of vegetation removed after a flood in the active channel is two or three times lower than before the event, it can be considered a robust trajectory of active channel widening and vegetational removal. Indeed, in the near future, the use of higher resolution, freely available satellite images may reduce the problem associated with mixed pixels in S-2, as well as expand the number of proglacial rivers where our model can be successfully applied, as rivers smaller than those tested here could be addressed. Alternatively, to address the mixed pixel issue, several approaches can be used, such as sub-pixel classification [19,55] and spectral unmixing [56].
As it is evident by looking at Figure 8, the misclassification errors stemming from our model do not substantially affect its overall performance in terms of mapping geomorphic macro-units in proglacial river reaches, and most likely will be negligible once the model is used for investigating the river morphological evolutionary trajectories in terms of the relative extension of the active channel (i.e., including base-flow channels and emergent sediments) vs. vegetated landforms. The limitations of our model become significant when the spatial dimensions of the geomorphic macro-units are smaller than the pixel size (100 m2 in this case), as, for instance, when base-flow channels are numerous and small-sized (Section 4.1); however, we believe this limitation is negligible as long as the macro-unit size is >2–3 times the pixel size.
Remotesensing 15 04687 g008
Figure 8. Map showing the location of misclassification errors (colored in yellow) in a sub-reach of the Saltoro River reach where the model was developed (inset map).
Figure 8. Map showing the location of misclassification errors (colored in yellow) in a sub-reach of the Saltoro River reach where the model was developed (inset map).
Remotesensing 15 04687 g008
Regarding the temporal aspects, a restriction of our model lies in the rather scarce available S-2 images with little cloud cover for similar proglacial areas in the Himalayas. In fact, such high mountain regions (as in other mountain ranges) are often covered by clouds, making the use of Sentinel-2 products prone to a reduced frequency of a possible “snapshot” to capture the evolution of fluvial dynamics through time. Nonetheless, 2019 S-2 offers more frequent coverage in the region, and GEE significantly facilitates the retrieval of cloud-free satellite images as a cloud-processing platform.

6. Conclusions

A multi-spectral supervised classification model based on a random forest classifier was developed to map the fluvial geomorphic macro-units (i.e., submerged base-flow channels, emerged sediment bars, vegetated surfaces) present in proglacial fluvial corridors of the Himalayas. The model was built in GEE based on freely available S-2 images at 10 m resolution. The model was first trained in three proglacial rivers (located in Pakistan, Kashmir, and Nepal) with an average accuracy of 96%. The model was then applied to 60 new testing polygons within an Indian river, obtaining an overall accuracy around 95%. Such accuracies are highly satisfactory and indicate that our model can reliably identify the fluvial macro-units in proglacial Himalayan rivers. The most important limitation of the developed model for S-2 regards its application to relatively narrow river channels (about <30 m in width), as misclassification of mixed pixels (especially due to overhanging vegetation) may become important under such circumstances. Another major limitation is the rather infrequent availability of S-2 images until 2019. Overall, we can conclude that the proposed model based on medium resolution S-2 images performs in a reliable manner in proglacial river corridors of the Himalayas, providing fast and cheap detection of morphological changes in such important fluvial systems, which are undergoing rapid variations associated with deglaciation.

Author Contributions

Z.M.: data collection, data analysis, and main writing. F.C. and S.B.: conceptualization, coordination, support in writing, and text revision. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Data Availability Statement

GEE code and datasets are available at “https://github.com/ZarkaMukhtar/GEE-Code-Classification-of-Himalayan-Proglacial-Rivers.git” (accessed on 20 September 2023).

Acknowledgments

We are grateful to Elisa Bozzolan for her valuable support in the analysis and in reviewing the manuscript.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Heckmann, T.; Mccoll, S.; Morche, D. Retreating ice: Research in pro-glacial areas matters. Earth Surf. Process. Landf. 2016, 41, 271–276. [Google Scholar] [CrossRef]
  2. Zhang, C.; Liu, Y.; Chen, X.; Gao, Y. Estimation of Suspended Sediment Concentration in the Yangtze Main Stream Based on Sentinel-2 MSI Data. Remote Sens. 2022, 14, 4446. [Google Scholar] [CrossRef]
  3. Gurnell, A.M.; Edwards, P.P.; Petts, G.E.; Ward, J. A conceptual model for alpine proglacial river channel evolution under changing climatic conditions. Catena 2000, 38, 223–242. [Google Scholar] [CrossRef]
  4. Marren, P.M.; Toomath, S.C. Channel pattern of proglacial rivers: Topographic forcing due to glacier retreat. Earth Surf. Process. Landf. 2014, 39, 943–951. [Google Scholar] [CrossRef]
  5. Carrivick, J.L.; Rushmer, E.L. Inter- and Intra-Catchment Variations in Proglacial Geomorphology: An Example from Franz Josef Glacier and Fox Glacier, New Zealand. Arct. Antarct. Alp. Res. 2018, 41, 18–36. [Google Scholar] [CrossRef]
  6. Huss, M.; Hock, R. Global-scale hydrological response to future glacier mass loss. Nat. Clim. Change 2018, 8, 135–140. [Google Scholar] [CrossRef]
  7. Gurnell, A.M.; Rinaldi, M.; Buijse, A.D.; Brierley, G.; Piégay, H. Hydromorphological frameworks: Emerging trajectories. Aquat. Sci. 2015, 78, 135–138. [Google Scholar] [CrossRef]
  8. Owens, P.N.; Blake, W.H.; Millward, G.E. Extreme levels of fallout radionuclides and other contaminants in glacial sediment (cryoconite) and implications for downstream aquatic ecosystems. Sci. Rep. 2019, 9, 12531. [Google Scholar] [CrossRef]
  9. Rinaldi, M.; Surian, N.; Comiti, F.; Bussettini, M. A methodological framework for hydromorphological assessment, analysis and monitoring (IDRAIM) aimed at promoting integrated river management. Geomorphology 2015, 251, 122–136. [Google Scholar] [CrossRef]
  10. Brierley, G.; Fryirs, K. Assemblages of geomorphic units: A building block approach to analysis and interpretation of river character, behaviour, condition and recovery. Earth Surf. Process. Landf. 2021, 47, 92–108. [Google Scholar] [CrossRef]
  11. Goodwin, N.R.; Armston, J.; Stiller, I.; Muir, J. Assessing the repeatability of terrestrial laser scanning for monitoring gully topography: A case study from Aratula, Queensland, Australia. Geomorphology 2016, 262, 24–36. [Google Scholar] [CrossRef]
  12. Vericat, D.; Wheaton, J.M.; Brasington, J. Revisiting the Morphological Approach. In Gravel-Bed Rivers: Process Disasters; John Wiley & Sons Ltd.: Hoboken, NJ, USA, 2017; pp. 121–158. [Google Scholar] [CrossRef]
  13. Williams, R.D.; Brasington, J.; Vericat, D.; Hicks, D.M. Hyperscale terrain modelling of braided rivers: Fusing mobile terrestrial laser scanning and optical bathymetric mapping. Earth Surf. Process. Landf. 2014, 39, 167–183. [Google Scholar] [CrossRef]
  14. Entwistle, N.; Milan, D.J. Recent remote sensing applications for hydro and morphodynamic monitoring and modelling. Earth Surf. Process. Landf. 2018, 43, 2283–2291. [Google Scholar] [CrossRef]
  15. Shelar, R.; Shinde, S.; Pande, C.B.; Moharir, K.N.; Orimoloye, I.R.; Mishra, A.K.; Varade, A.M. Sub-watershed prioritization of Koyna river basin in India using multi criteria analytical hierarchical process, remote sensing and GIS techniques. Phys. Chem. Earth 2022, 128, 103219. [Google Scholar] [CrossRef]
  16. Marcus, W.A.; Fonstad, M.A. Remote sensing of rivers: The emergence of a subdiscipline in the river sciences. Earth Surf. Process. Landf. 2010, 35, 1867–1872. [Google Scholar] [CrossRef]
  17. Sudmanns, M.; Tiede, D.; Lang, S.; Bergstedt, H.; Trost, G.; Augustin, H.; Baraldi, A. Blaschke, TBig Earth data: Disruptive changes in Earth observation data management and analysis? Int. J. Digit. Earth 2020, 13, 832–850. [Google Scholar] [CrossRef]
  18. Henshaw, A.J.; Gurnell, A.M.; Bertoldi, W.; Drake, N.A. An assessment of the degree to which Landsat TM data can support the assessment of fluvial dynamics, as revealed by changes in vegetation extent and channel position, along a large river. Geomorphology 2013, 202, 74–85. [Google Scholar] [CrossRef]
  19. Carbonneau, P.E.; Belletti, B.; Micotti, M.; Lastoria, B.; Casaioli, M.; Mariani, S.; Marchetti, G.; Bizzi, S. UAV-based training for fully fuzzy classification of Sentinel-2 fluvial scenes. Earth Surf. Process. Landf. 2020, 45, 3120–3140. [Google Scholar] [CrossRef]
  20. Boothroyd, R.J.; Nones, M.; Guerrero, M. Deriving Planform Morphology and Vegetation Coverage from Remote Sensing to Support River Management Applications. Front. Environ. Sci. 2021, 9, 657354. [Google Scholar] [CrossRef]
  21. Boothroyd, R.J.; Williams, R.A.; Hoey, T.; Barrett, B.; Prasojo, O.A. Applications of Google Earth Engine in fluvial geomorphology for detecting river channel change. Wiley Interdiscip. Rev. Water 2020, 8, e21496. [Google Scholar] [CrossRef]
  22. Bioresita, F.; Puissant, A.; Stumpf, A.; Malet, J.P. Fusion of Sentinel-1 and Sentinel-2 image time series for permanent and temporary surface water mapping. Int. J. Remote Sens. 2019, 40, 9026–9049. [Google Scholar] [CrossRef]
  23. Cavallo, C.; Papa, M.N.; Negro, G.; Gargiulo, M.; Ruello, G.; Vezza, P. Exploiting Sentinel-2 dataset to assess flow intermittency in non-perennial rivers. Sci. Rep. 2022, 12, 21756. [Google Scholar] [CrossRef] [PubMed]
  24. Marchetti, G.; Bizzi, S.; Belletti, B.; Lastoria, B.; Comiti, F.; Carbonneau, P. Mapping riverbed sediment size from Sentinel-2 satellite data. Earth Surf. Process. Landf. 2022, 47, 2544–2559. [Google Scholar] [CrossRef]
  25. Popescu, M.C.; Sasu, L.M. Feature extraction, feature selection and machine learning for image classification: A case study. In Proceedings of the 2014 International Conference on Optimization of Electrical and Electronic Equipment, OPTIM, Bran, Romania, 22–24 May 2014; pp. 968–973. [Google Scholar] [CrossRef]
  26. Nemani, R.; Votava, P.; Michaelis, A.; Melton, F.; Milesi, C. Collaborative Supercomputing for Global Change Science. Eos Trans. Am. Geophys. Union 2011, 92, 109–110. [Google Scholar] [CrossRef]
  27. Kondolf, G.M.; Piégay, H. Tools in Fluvial Geomorphology: Problem Statement and Recent Practice. In Tools in Fluvial Geomorphology; John Wiley & Sons Ltd.: Hoboken, NJ, USA, 2005; pp. 1–22. [Google Scholar] [CrossRef]
  28. Piégay, H.; Arnaud, F.; Belletti, B.; Bertrand, M.; Bizzi, S.; Carbonneau, P.; Dufour, S.; Liébault, F.; Ruiz-Villanueva, V.; Slater, L. Remotely sensed rivers in the Anthropocene: State of the art and prospects. Earth Surf. Process. Landf. 2020, 45, 157–188. [Google Scholar] [CrossRef]
  29. Zhai, X.; Oliver, A.; Kolesnikov, A.; Beyer, L. S4L: Self-Supervised Semi-Supervised Learning. In Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Republic of Korea, 27 October–2 November 2019; pp. 1476–1485. [Google Scholar]
  30. Yang, H.; Wang, Z.; Zhao, H.; Guo, Y. Water Body Extraction Methods Study Based on RS and GIS. Procedia Environ. Sci. 2011, 10, 2619–2624. [Google Scholar] [CrossRef]
  31. Shorewala, S.; Ashfaque, A.; Sidharth, R.; Verma, U. Weed Density and Distribution Estimation for Precision Agriculture Using Semi-Supervised Learning. IEEE Access 2021, 9, 27971–27986. [Google Scholar] [CrossRef]
  32. Tooke, T.R.; Coops, N.C.; Goodwin, N.R.; Voogt, J.A. Extracting urban vegetation characteristics using spectral mixture analysis and decision tree classifications. Remote Sens. Environ. 2009, 113, 398–407. [Google Scholar] [CrossRef]
  33. Chen, Z.; Liu, X.; Yang, J.; Little, E.; Zhou, Y. Deep learning-based method for SEM image segmentation in mineral characterization, an example from Duvernay Shale samples in Western Canada Sedimentary Basin. Comput. Geosci. 2020, 138, 104450. [Google Scholar] [CrossRef]
  34. Camalan, S.; Cui, K.; Pauca, V.P.; Alqahtani, S.; Silman, M.; Chan, R.; Plemmons, R.J.; Dethier, E.N.; Fernandez, L.E.; Lutz, D.A. Change Detection of Amazonian Alluvial Gold Mining Using Deep Learning and Sentinel-2 Imagery. Remote Sens. 2022, 14, 1746. [Google Scholar] [CrossRef]
  35. Wieland, M.; Martinis, S. Large-scale surface water change observed by Sentinel-2 during the 2018 drought in Germany. Int. J. Remote Sens. 2020, 41, 4742–4756. [Google Scholar] [CrossRef]
  36. Belletti, B.; Rinaldi, M.; Bussettini, M.; Comiti, F.; Gurnell, A.M.; Mao, L.; Nardi, L.; Vezza, P. Characterising physical habitats and fluvial hydromorphology: A new system for the survey and classification of river geomorphic units. Geomorphology 2017, 283, 143–157. [Google Scholar] [CrossRef]
  37. Farooq Ahmed, M.; David Rogers, J. Regional level landslide inventory maps of the Shyok River watershed, Northern Pakistan. Bull. Eng. Geol. Environ. 2016, 75, 563–574. [Google Scholar] [CrossRef]
  38. Dubey, R.; Nand Jha, V. Soil Erosion Evaluation through Prioritization of Shyok-Nubra Watershed using Morphometric, Principal Component and Weighted Sum Approaches. 2022. Available online: https://www.researchsquare.com/article/rs-1922043/v1 (accessed on 20 September 2023).
  39. Ragettli, S.; Bolch, T.; Pellicciotti, F. Heterogeneous glacier thinning patterns over the last 40 years in Langtang Himal, Nepal. Cryosphere 2016, 10, 2075–2097. [Google Scholar] [CrossRef]
  40. Pradhananga, N.S.; Kayastha, R.B.; Bhattarai, B.C.; Adhikari, T.R.; Pradhan, S.C.; Devkota, L.P.; Shrestha, A.B.; Mool, P.K. Estimation of discharge from Langtang River basin, Rasuwa, Nepal, using a glacio-hydrological model. Ann. Glaciol. 2014, 55, 223–230. [Google Scholar] [CrossRef]
  41. Bhushan, S.; Syed, T.H.; Kulkarni, A.V.; Gantayat, P.; Agarwal, V. Quantifying Changes in the Gangotri Glacier of Central Himalaya: Evidence for Increasing Mass Loss and Decreasing Velocity. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2017, 10, 5295–5306. [Google Scholar] [CrossRef]
  42. Gao, B.C. NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 1966, 58, 257–266. [Google Scholar] [CrossRef]
  43. Ruiz-Villanueva, V.; Piégay, H.; Scorpio, V.; Bachmann, A.; Brousse, G.; Cavalli, M.; Comiti, F.; Crema, S.; Fernández, E.; Furdada, G.; et al. River widening in mountain and foothill areas during floods: Insights from a meta-analysis of 51 European Rivers. Sci. Total Environ. 2023, 903, 166103. [Google Scholar] [CrossRef]
  44. Singh, P.; Haritashya, U.K.; Ramasastri, K.S.; Kumar, N. Diurnal variations in discharge and suspended sediment concentration, including runoff-delaying characteristics, of the Gangotri Glacier in the Garhwal Himalayas. Hydrol. Process. 2005, 19, 1445–1457. [Google Scholar] [CrossRef]
  45. Clift, P.; Holbourn, A.; France-Lanord, C.; Zheng, H. Evolution of the Asian Monsoon. Eos 2020, 101. [Google Scholar] [CrossRef]
  46. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  47. Cutler, D.R.; Edwards, T.E.; Beard, K.H.; Cutler, A.; Hess, K.; Gibson, J.D.; Lawler, J.J. Random Forests for Classification in Ecology. Ecology 2007, 88, 2783–2792. [Google Scholar] [CrossRef] [PubMed]
  48. Yan, X.; Li, J.; Yang, D.; Li, J.; Ma, T.; Su, Y.; Shao, J.; Zhang, R. A Random Forest Algorithm for Landsat Image Chromatic Aberration Restoration Based on GEE Cloud Platform—A Case Study of Yucatán Peninsula, Mexico. Remote Sens. 2022, 14, 5154. [Google Scholar] [CrossRef]
  49. Congalton, R.G. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens. Environ. 1911, 37, 35–46. [Google Scholar] [CrossRef]
  50. Keuchel, J.; Naumann, S.; Heiler, M.; Siegmund, A. Automatic land cover analysis for Tenerife by supervised classification using remotely sensed data. Remote Sens. Environ. 2003, 86, 530–541. [Google Scholar] [CrossRef]
  51. Ma, L.; Li, M.; Ma, X.; Cheng, L.; Du, P.; Liu, Y. A review of supervised object-based land-cover image classification. ISPRS J. Photogramm. Remote Sens. 2017, 130, 277–293. [Google Scholar] [CrossRef]
  52. Zhang, W.; Tang, P.; Corpetti, T.; Zhao, L. WTS: A Weakly towards Strongly Supervised Learning Framework for Remote Sensing Land Cover Classification Using Segmentation Models. Remote Sens. 2021, 13, 394. [Google Scholar] [CrossRef]
  53. Smith, C.S.R. Observing the Seasonal Evolution of Supraglacial Ponds in High Mountain Asia: A Supervised Classification. Ph.D. Thesis, University of Cambridge, Cambridge, UK, 2022. [Google Scholar] [CrossRef]
  54. Bozzolan, E.; Brenna, A.; Surian, N.; Bizzi, S. Quantifying the impact of the spatiotemporal resolution of the Sentinel 2 images on the interpretation of the active channel trajectory. An application on a dynamic river section in Northern Italy. Water Resour. Res. 2023; under review. [Google Scholar]
  55. Thornton, M.W.; Atkinson, P.M.; Holland, D.A. Sub-pixel mapping of rural land cover objects from fine spatial resolution satellite sensor imagery using super-resolution pixel-swapping. Int. J. Remote Sens. 2007, 27, 473–491. [Google Scholar] [CrossRef]
  56. Zhang, J.; Rivard, B.; Sanchez-Azofeifa, A. Derivative spectral unmixing of hyperspectral data applied to mixtures of lichen and rock. IEEE Trans. Geosci. Remote Sens. 2004, 42, 1934–1940. [Google Scholar] [CrossRef]
Table 1. List of selected available satellite images, classification training, and validation datasets.
Table 1. List of selected available satellite images, classification training, and validation datasets.
Sentinel-2Sample PolygonsSample Pixels
Calibration
70%		Validation
30%		CalibrationValidation/Total Classified Pixels
Nubra20 September 20193601402273915
17 August 2020
25 August 2021
Saltoro16 August 201914060824338
25 August 2020
19 September 2021
Langtang-Khola29 September 202014060671431
Ganga-Bhagirathi20 August 2019
13 September 2020NA60NA231
18 September 2021
Table 2. Macro-unit classes * based on Google satellite images.
Table 2. Macro-unit classes * based on Google satellite images.
Proglacial Fluvial FeaturesDescription
Base-flow channelsWet (submerged) and dry (emerged) base-flow channels
Vegetated surfacesVegetated surfaces within the channel (islands and banks) or adjacent to it (floodplains and terraces)
Emerged sediments Emerged and unvegetated areas, mostly represented by sediment bars
* Macro-units mapped based on Geomorphic Unit Systems (GUS) recently proposed by Belletti et al. (2017) [3].
Table 3. List of selected available satellite images, classification training datasets, validation datasets, and overall accuracy.
Table 3. List of selected available satellite images, classification training datasets, validation datasets, and overall accuracy.
Sentinel-2Sample PixelsOverall Accuracy %Kappa Coefficient
Validation/Total Classified PixelsCorrectly Classified Pixels
Nubra20 September 2019915879960.94
17 August 20209151001
25 August 202191399.70.99
Saltoro16 August 201933832094.60.92
25 August 202033799.70.99
19 September 2021297870.82
Langtang-Khola29 September 2020431429990.99
Ganga-Bhagirathi20 August 201923121492.60.87
13 September 202022898.70.97
18 September 202122396.50.94
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Mukhtar, Z.; Bizzi, S.; Comiti, F. Supervised Geomorphic Mapping of Himalayan Rivers Based on Sentinel-2 Data. Remote Sens. 2023, 15, 4687. https://doi.org/10.3390/rs15194687

AMA Style

Mukhtar Z, Bizzi S, Comiti F. Supervised Geomorphic Mapping of Himalayan Rivers Based on Sentinel-2 Data. Remote Sensing. 2023; 15(19):4687. https://doi.org/10.3390/rs15194687

Chicago/Turabian Style

Mukhtar, Zarka, Simone Bizzi, and Francesco Comiti. 2023. "Supervised Geomorphic Mapping of Himalayan Rivers Based on Sentinel-2 Data" Remote Sensing 15, no. 19: 4687. https://doi.org/10.3390/rs15194687

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop