Introduction

Land Use/Land Cover (LULC) maps have emerged as one of the most critical sources of environmental information, employed in assessing the natural resources’ status in any river basin or watershed (Nie et al. 2011; Atzberger 2013; Roy et al. 2015). Land use and land cover information also find significant applications in hydrological modeling, watershed management, natural disaster management, and land management (Lillesand et al. 2008; Nie et al. 2011; Sierra-Soler et al. 2016). Satellite remote sensing offers a unique set of advantages like global coverage, high temporal frequency, synoptic view, and the ability to observe inaccessible areas (Campbell and Wynne 2011). The most important benefit is the availability of optical remote sensing datasets in the open-source domain. Renowned agencies like the United States Geological Survey (USGS) and the European Space Agency (ESA) have been providing optical satellite datasets, viz. Landsat and Sentinel series for free, which has popularized remote sensing for various applications (Saadat et al. 2011; Wondrade et al. 2014; Mishra et al. 2020; Thanh et al. 2020). Numerous studies by different researchers have demonstrated the efficacy of optical remote sensing data products for land cover classification (Gong et al. 2013; Gómez et al. 2016).

Generally, the land cover describes the physical land types viz. area covered by forests, impervious surfaces, agricultural lands, barren lands, and water bodies. On the other hand, as the term suggests, land use describes the utility of the land for various purposes by humans, viz. for development, conservation, or mixed uses. The entire ecosystem is strongly affected by variables responsible for land use and climatic changes (Shrestha et al. 2012). LULC is attributed as one of the most relevant flow contributors in a watershed (Kim et al. 2013; Himanshu et al. 2017). Therefore, for assessing water resources availability and its management in a watershed, LULC change analysis is of utmost importance (Malik and Bhat 2014).

Moreover, LULC changes cause a substantial effect in modifying the rainfall breakup into various hydrological components like surface runoff, infiltration, interception, and evaporation (Costa et al. 2003; Mao and Cherkauer 2009; Sajikumar and Remya 2015). There are two reasons for the variation in the LULC, one being the natural dynamics, and the other is attributed to human activities (Thenkabail et al. 2005; Bontemps et al. 2008; Singh et al. 2014; Zhang et al. 2019). These factors are responsible for deforestation, global warming, loss of biodiversity, increasing natural disasters, and global environmental change (DeFries et al. 2010; Owrangi et al. 2014; Mahmood et al. 2014; Barros et al. 2021).

Remote sensing data and techniques, along with GIS, provide an apt platform to develop and prepare LULC maps. Multi-spectral/temporal satellite data with medium/high spatial resolution have materialized as the most preferred data sources for deriving LULC maps (Güler et al. 2007). Conventionally, maps were prepared using available records and extensive field surveys, rendering them tedious, laborious, time-consuming, and expensive. Moreover, due to the dynamic nature of the environment, the output maps used to become outdated (Dash et al. 2015). In contrast, remote sensing data provides highly vital information in a very cost-effective and less time-consuming manner. High-resolution satellite data products are employed in large cities to estimate LULC changes. However, the too-high cost of such data sets limits their availability (Dwivedi et al. 2005). On the contrary, satellite data products with the medium resolution, specifically from the Landsat series, are among the most popular datasets worldwide for LULC mapping and change detection studies (Kumar et al. 2012; Wang et al. 2009; Odindi et al. 2012).

Numerous methods and techniques have been used for preparing satellite image-based land cover maps in the past (Li et al. 2014). These methods include unsupervised and supervised classification approaches, as well as parametric and non-parametric methods. In very recent times, non-parametric methods based on the machine learning approach have gained tremendous consideration for satellite image-based LULC classification and mapping. Researchers across the globe have been carrying out several studies on LULC mapping and modeling by employing various machine learning algorithms (Civco 1993; Pal 2005; Teluguntla et al. 2018; Talukdar et al. 2020a, b). Also, comparison-based studies have been carried out wherein multiple machine learning-based models, and conventional models were employed for image classification (Rogan et al. 2008; Camargo et al. 2019). The most favored and in-demand algorithms include Support Vector Machine (SVM), Random Forests (RF), and k-Nearest Neighbors (k-NN) (Huang et al. 2002; Franco-Lopez et al. 2001; Kennedy et al. 2015). Object-based classification is another preferred approach that has proved to be very useful for classifying fine resolution satellite images (Machala and Zejdová 2014). The most critical aspect in LULC mapping is classification accuracy, which plays a crucial role as a deciding factor in selecting the best among various classification methods. The significant elements affecting the classification accuracy are the type of satellite sensor, spatial resolution, training data sources, accuracy assessment data sources, the total number of classes, and the classification approach (Manandhar et al. 2009). Consequently, the most critical factor is choosing suitable classification algorithms for achieving acceptable classification accuracy (Lu and Weng 2007).

LULC mapping in mountainous terrain has always been challenging due to hills, valleys, plateaus, and mountains. This complexity in the landscape introduces effects like shadows and illumination issues due to aspect variation causing severe changes in the surface reflectance of various LULC types (Wang et al. 2020). Moreover, complex topography also poses challenges in water body identification. Advanced machine learning-based image classification techniques have been producing promising results, and therefore, their use to generate LULC maps in complex terrain must be explored.

Keeping the aforementioned into consideration, the following are the specific objectives of the present study:

  1. 1.

    To prepare Land Use/Land Cover maps for the Alaknanda River Basin (ARB) using three approaches.

  2. 2.

    To compare the classification results obtained from MLC, SVM, and RT algorithms based on accuracy assessment.

The inferences drawn from this study will provide practical insights into using machine learning techniques for performing LULC classification, especially in snow-covered mountainous regions. Also, the methodology presented can be replicated to classify complex topographic settings elsewhere.

Materials and methods

Study area

The study focusses on the Alaknanda River Basin (ARB), which lies between 78°33′ E to 80°15′ E longitude and 29°59′24″ N to 31°04′51″ N latitude in the Northwestern Himalayan region of Uttarakhand in India and encompasses a catchment area of 11,035.3 km2 (Fig. 1) with a total length of 183.5 km. Alaknanda originates from the Satopanth, and the Bhagirath Kharak glaciers flow downstream and meet the Bhagirathi River at Devprayag to finally become river Ganga or the Ganges. The River Alaknanda is joined by the Dhauliganga, Nandakini, Mandakini, Pindar and some other tributaries as shown in Fig. 1. The catchment area features a variety of climates, viz. subtropical, temperate, sub-alpine, and alpine, primarily due to the substantial variation in the altitude ranging from 446 to 7801 m. ARB has a unique topographic setting featuring high mountain peaks and glaciated valleys, especially in the northern part (Sharma and Mohanty 2018). The area is characterized by high relief, steep slopes, and high drainage density (Panwar et al. 2017). Furthermore, the topography is generally represented by north–south trending ridges and incised river valleys (Ghosh et al. 2019). Figure 1 shows clearly that the river valleys are very narrow in the upper part catchment, relatively narrow in the middle, and relatively wider in the lower reach.

Fig. 1
figure 1

Location map of the study area

Full size image

Data and sources

The satellite data-based LULC classification was carried out using freely available 30 m (spatial resolution) Landsat 8 OLI Level 1 images. The dataset was downloaded from the ‘EarthExplorer’ official website of the United States Geological Survey (USGS) (https://earthexplorer.usgs.gov/). A total of 2 multi-band images (Table 1) were downloaded for this purpose. Each L8/OLI image has data in 11 bands but, for classification purposes, data for bands 2 to 7 and 9 (details listed in Table 2) were downloaded.

Table 1 Details of satellite data downloaded for LULC classification
Full size table
Table 2 Landsat 8 OLI individual bands wavelength range and resolution
Full size table

Tools and techniques used

The Semi-Automatic Classification Plugin (SCP) version 6.4.0.2 in QGIS 3.14 software interface was used for pre-processing the downloaded L8/OLI images, and after that, ArcGIS 10.4 was used for classification and accuracy assessment. ENVI software was used for the estimation of the Normalized Difference Snow Index. For the classification in the ArcGIS interface, three different training algorithms (classifiers) from the spatial analyst toolbox were put into use, namely, MLC, SVM, and RT.

Methodology

This study focused on classifying the mountainous river basin and preparing LULC maps using three different classification algorithms. The resulting maps were compared on the basis of the accuracy assessment of each classified image. The flowchart in Fig. 2 depicts the sequence of steps followed as part of the methodology adopted in this study.

Fig. 2
figure 2

Methodology flowchart

Full size image

The downloaded satellite images were pre-processed, where the Digital Number (DN) values in the images were converted to the Top of Atmosphere (ToA) Reflectance. During this process, the Dark Object Subtraction (DOS) atmospheric corrections were also applied using the SCP tool in QGIS. All the pre-processed image bands were then stacked together and clipped using the vector boundary of the study area. Both the clipped band stacks were then mosaiced together to create a single image of the entire river basin. The image was classified into nine different classes, namely, Snow, Forest, Sparse vegetation (or barren areas), Built-up, River bed, Water, Agriculture, Cloud, and Shadow. Concepts of visual image interpretation were used to identify each class and consequently assigned to the image pixels for image classification.

For the image classification training, sample polygons were digitized using field data, Google Earth Imagery, and different band combinations of the satellite image. A total of seven band combinations (R-G-B) were used to select training samples, namely, Natural Color Composite (4–3-2), False Color Composite (FCC)-Urban (7–6-4), Color Infrared Composite-Vegetation (5–4-3), Agricultural Area Composite (6–5-2), Healthy Vegetation Composite (5–6-2) and Land Water Composite (5–6-4). The FCC of the two images (Fig. 3a) with a band combination of 7–5-3 is depicting the presence of snow/ice (cyan color) in the high elevation areas. Figure 3b shows the Natural Color Composite with band combination 5–4-3, wherein the snow-covered mountains are prominently visible. The same composite was used for the selection of training samples to classify snow. Also, to verify the classification of snow, the Normalized Difference Snow Index (NDSI) as shown in Fig. 3c was calculated using the following general equation: NDSI = (Green – SWIR)/(Green + SWIR). Based on the previous studies (Hall et al. 1998; Nijhawan et al. 2016) a threshold of 0.4 was selected to distinguish between snow and no snow areas. The process was carried out in ENVI software, using the Band Math tool. Even though the satellite images used in the analysis were acquired in April and May, the snow in the region is justified by the presence of glaciers in the high elevation zones of the Alaknanda River Basin. The higher elevation areas in the northern part of the study area remain snow-covered all round the year. Therefore, to appreciate and associate snow/ice in the study area, an elevation map of the study area using SRTM DEM having a spatial resolution of 30 m was prepared, as shown in Fig. 3d. This representation correlates with the classification of snow.

Fig. 3
figure 3

a FCC of the satellite images downloaded for classification b Clipped Natural Color Composite of ARB c Normalized Difference Snow Index map of the study area d Elevation maps of the study area

Full size image

A signature file with samples for all nine classes was developed to classify the satellite image. The classification was carried out using the Maximum Likelihood Classifier (MLC), Support Vector Machine Classifier (SVM), and Random Trees Classifier (RT). The spectral signature file containing samples for each class was used to generate classified images in each of the classification approaches. Finally, an accuracy assessment was carried out for all the three classified images separately. For the accuracy assessment, reference points shapefile was created using ground truth field data and points identified using Google Earth. A total of 946 points were identified as ground truth references to be compared with the corresponding pixels in the classified image. The reference points or the test pixels were chosen through random sampling, but it was made sure that they were distinct from the training pixels. The number of reference points for ‘Snow’, ‘Forest’, ‘Sparse vegetation’, ‘Built up’, ‘River bed’, ‘Agriculture’, ‘Clouds’, ‘Shadows’ and ‘Water’ were 131, 158, 46, 186, 18, 170, 41, 33 and 163, respectively. A pivot table of reference pixels class and classified pixels class was created in ArcGIS interface and then exported to Microsoft excel for the generation of error matrix and calculation of kappa coefficient for all three classified images. Additionally, the Producer’s and User’s accuracies for each class were also computed to access the accuracy of classification. The following Eqs. (1 and 2) were adopted for calculating producer’s accuracy and user’s accuracy:

$$Produce{r}^{^{\prime}}s\ accuracy\ \left(\%\right)=\left(\frac{{x}_{ii}}{{x}_{+i}}\right)*100$$
(1)
$$Use{r}^{^{\prime}}s\ accuracy\ \left(\%\right)=\left(\frac{{x}_{ii}}{{x}_{i+}}\right)*100$$
(2)

Equations (3 and 4) adopted for the overall accuracy and kappa coefficient calculation are:

$$Overall\;accuracy=\frac1N\sum_{i=1}^cn_i$$
(3)
$$Kappa\ coefficient=\left\{N\sum_{i=1}^cx_{ii}-\sum_{i=1}^c\left(x_{i+}\ast x_{+i}\right)\right\}\div\left\{N^2-\sum_{i=1}^c\left(x_{i+}\ast x_{+i}\right)\right\}$$
(4)

where N represents total pixels; c represents the total number of classes; xii = total number of pixels in row ‘i’ and column ‘i’; xi+  = total number of samples in a row ‘i’; x+i = total number of samples in column ‘i’ in the error matrix.

Since cloud detection is a significant challenge in any LULC map, the majority of LULC mapping is carried out using cloud-free satellite imageries, or the clouds are masked, and the masked portion of satellite imagery is obtained from the imagery of another year, preferably of the same month of the year. However, in this study, attempt has been made to classify the clouds and their shadows using the pixel values. The exact cloud and shadow pixel values were identified using the Landsat Quality Assessment Toolbox extension for ArcGIS. The Quality Assessment (QA) band was downloaded to use this tool effectively, wherein each pixel contains an integer value that represents bit packed combinations of surface, atmospheric, and sensor conditions. The integer values for cloud and shadow pixels were identified and extracted using the tool mentioned above in the ArcGIS interface, and then it was compared with the area classified as clouds by the classifiers. Using this approach, it was possible to classify the clouds with error-free accuracy using all three classification approaches.

Results and discussion

The LULC maps for the snow-fed Alaknanda River Basin of the Northwest Himalayan region were prepared using three different classification algorithms; namely, MLC, SVM, and RT are presented in Fig. 4. Since the river network in the area is relatively narrow, it is not prominently visible in the maps presented. Therefore, a zoomed-in portion of the southwestern part of the catchment has been shown in Fig. 5 to visually appreciate the difference in classified maps obtained by the three approaches. Furthermore, the accuracy assessment was conducted for each of these maps to compare and evaluate the efficiency of these algorithms. The accuracy assessment results for the LULC maps generated using ML, SVM, and RT classifiers are presented in Tables 3, 4, and 5, respectively.

Fig. 4
figure 4

LULC maps for the Alaknanda River Basin generated using (a) Maximum Likelihood Classifier, (b) Support Vector Machine, and (c) Random Trees

Full size image
Fig. 5
figure 5

LULC status of the southwestern portion of ARB (a) Maximum Likelihood Classifier (b)Support Vector Machine and (c) Random Trees

Full size image
Table 3 Error matrix showing the Kappa coefficient and overall accuracy for the LULC map generated using the Maximum Likelihood Classifier
Full size table
Table 4 Error matrix showing the Kappa coefficient and overall accuracy for the LULC map generated using the SVM classifier
Full size table
Table 5 Error matrix showing the Kappa coefficient and overall accuracy for the LULC map generated using the RT classifier
Full size table

Table 3 presents the accuracy assessment of the LULC map obtained using MLC. It shows that with a kappa coefficient of 0.79, the overall accuracy of the classification is 82%. The producer’s accuracy of ‘Forest’, ‘Snow’ and ‘Clouds’ is above 90%; ‘Agriculture’ and ‘Sparse vegetation’ were classified with a producer’s accuracy of above 80%; ‘Built up’ area with over 75%; ‘River bed’ and ‘shadows’ between 60 to 70% and ‘Water’ having least and unsatisfactory producer’s accuracy of 57.93%. As far as the user’s accuracy is concerned, except the ‘Built up’ class, which shows a very low accuracy of 40.22%, all the remaining classes show a user’s accuracy of greater than 70%.

Table 4 presents the accuracy assessment of the LULC map obtained using SVM. It shows a slightly higher kappa coefficient of 0.82 and an improved overall accuracy of 84%. The producer’s accuracies of ‘Snow’, ‘Forest’, ‘Sparse vegetation’, ‘River bed’, ‘Clouds’ and ‘Shadows’ are above 90%. Furthermore, SVM shows a remarkable improvement in the producer’s accuracy of ‘water’ as high as 86.59%. However, the producer’s accuracy of ‘Built up’ class is as low as 59.68%, which is not satisfactory. The SVM approach obtained a user’s accuracy of 42% and 68% for ‘Sparse vegetation’ and ‘River bed’, respectively. The user’s accuracy for the remaining classes is above 80%. Thus, SVM classification yielded much better results than the ML classification algorithm. The objective of SVM classification is to identify the optimal boundary between various classes or samples, also referred to as the support vectors. It is a binary classifier capable of identifying a single boundary between two separate classes (Maxwell et al. 2018).

Finally, Table 5 presents the accuracy assessment of the LULC map obtained using an RT classifier. This classification approach shows the highest values for the kappa coefficient (0.83) and overall accuracy (85%). The producer’s accuracies for ‘Snow’, ‘Forest’, ‘Sparse vegetation’, ‘Water’, ‘Cloud’ and ‘Shadow’ are above 90%. ‘Agriculture’ class shows a good producer’s accuracy of 82.35%, while for ‘Built up’ and ‘River bed’, it is 54.30% and 61.11%, respectively.

Visual comparison of the three LULC maps presented in Fig. 4 shows a similar pattern for the two machine learning approaches. However, the map classified using MLC looks different. Figure 4a shows that the Maximum Likelihood Classifier has classified a considerable area under as built up in the northeastern part of the basin. On the contrary, Fig. 4b and c shows that the same area has minimal built-up areas as classified by the machine learning algorithms. Another interesting visual comparison is for the sparse vegetation or barren areas class. MLC has classified a considerably large number of pixels in this class as compared to SVM and RT algorithms.

Similarly, comparing the share of area under each class using all three classification approaches reveals a mix of minor and noteworthy differences. To appreciate the percentage of the area classified under each LULC class, a quantitative comparison of the classification results was performed and is presented in the form of bar charts in Fig. 6. The bar charts clearly indicate the variation of classification output generated using MLC, SVM, and RT algorithms.

Fig. 6
figure 6

Percentage of the area under different classes for each classification approach

Full size image

In this study, particular emphasis is directed towards classifying water using medium resolution satellite data in complex mountainous terrain. Since the Alaknanda River Basin features narrow river stretches and steep valleys, it becomes challenging to identify water pixels in optical imagery. Therefore, multiple locations along the river course were carefully selected to prepare the training samples, and classification results indicate that MLC failed to capture water bodies, but both the machine learning approaches performed well. For a better visual representation of water body classification, a subset of the study area from the southwestern part of the catchment is presented in Fig. 5. It is visually evident that MLC has not been able to capture the river, especially in the southwest part of the map (Fig. 5a) while, SVM and RT have detected the river quite well (Fig. 5b, c).

In light of the classification results, the above discussion strongly suggests that machine learning algorithms (SVM and RT) have outperformed the conventional MLC technique to generate LULC maps for the snow-covered mountainous Alaknanda River Basin. The results are in line with other studies wherein several machine learning models like SVM, random forest (RF), radial basis function (RBF), decision tree (DT), and naïve bayes (NB) have performed better in comparison to the conventional classification approaches (Ma et al. 2019; Shih et al. 2019; Talukdar et al. 2020a, b). Researchers have concluded explicitly that SVM and RF (Mountrakis et al. 2011; Ma et al. 2017; Carranza et al. 2019) are the best ML-based image classification techniques.

Summary and conclusions

It has been reported that Machine Learning is capable of generating a classification with higher accuracy for the remotely sensed satellite data in comparison to the parametric approaches like Maximum Likelihood (Maxwell et al. 2018). In the current study, an attempt has been made to compare three different classification approaches, namely ML, SVM, and RT, to prepare LULC maps for the snow-fed Alaknanda River Basin of the Northwest Himalayan region. The following conclusions are drawn from this study:

  • The highest values of the Kappa coefficient and overall accuracy were obtained using the RT approach, followed by SVM and ML. This shows that advanced machine learning-based classifiers performed better than the parametric Maximum Likelihood Classifier in the Alaknanda River Basin.

  • Snow has been classified with a very high accuracy of 99.24% in all the classification approaches. Forest has also been classified with high accuracy in all three methods, but SVM and RT are better than ML. Similarly, Sparse vegetation has been most accurately classified by RT, followed by SVM and ML.

  • Classification of Built-up shows flipped results where ML obtained the highest accuracy of 75.81. On the other hand, both SVM and RT performed poorly, with accuracies of less than 60%. Also, Agriculture was best classified by the ML classifier, followed by RT and SVM.

  • The SVM classifier outperformed the ‘River bed’ classification. In all three approaches, the clouds were very accurately classified. ML and RT showed a cent percent accuracy.

  • The shadow pixels were classified with an accuracy of 70% by ML classifier, but the machine learning-based approaches classified these with 100% accuracy.

  • Finally, the water pixels in the satellite images were most accurately classified by the RT classifier followed by SVM and were poorly classified by the ML classifier.

Finally, it can be concluded that for the snow-fed Alaknanda River Basin, the advanced machine learning-based parametric classifiers have performed better than the Maximum Likelihood Classifier. The results cannot be generalized, but the accuracy assessment shows that Machine Learning based classifiers have outperformed in accurately classifying the majority of the classes. Especially for the watershed LULC mapping, accurate classification of water bodies is of utmost importance. This study suggests that the RT classifier is the best of three for accurately classifying water bodies. This particular inference signifies a robust utility of advanced machine learning algorithms for performing LULC classification.

Furthermore, the methodology can be replicated in snow-fed mountainous regions elsewhere. Change detection studies can also be conducted using RT, SVM, or even other machine learning-based algorithms using freely available medium resolution satellite data products. However, it can be argued that microwave remote sensing data is most preferred to classify water, but in highly complex and undulated terrain, the occurrence of foreshortening and shadow effect introduces errors in microwave data. Therefore, the study shows that the Machine learning-based classification approaches improve water detection capability and LULC mapping functionality using satellite-based remote sensing data. Moreover, researchers in the remote sensing domain have increasingly interested in exploring and employing advanced machine learning algorithms for image classification (Rodriguez-Galiano et al. 2012; Yeom et al. 2013; Jamali 2020). The field is developing rapidly, and new algorithms and implementations are becoming available continuously. The application of machine learning algorithms in LULC classification can result in high-quality results, as the classification results of this research shows.