ESTIMATION OF LANDCOVER TYPES OVER HIMALAYAN REGION WITH THE CLASSIFICATION OF OPTICAL AND MICROWAVE-BASED IMAGE FUSION DATASET

Himalayas play a significant role in terms of climate influence, the origin of rivers, hydropower generation, tourism, and forest wealth. The monitoring of the rugged terrain Himalayas via remote sensing is one of the efficient solutions to meet future requirements. In remote sensing, the sensors can be categorized as optical and microwave. The optical-based sensor provides multispectral or hyperspectral information at a very fine spatial resolution but is limited to daytime images without any penetration through the clouds. Whereas, the microwave works more effectively due to day/night image acquisition and cloud penetration capabilities. Therefore, the image fusion of multi-sensors (optical and microwave) datasets is important to extract crucial information about the Earth surface, especially over the Himalayas. However, the main aim of the article is to retrieve the different landcover types using various classifiers i.e., Linear Spectral Mixing (LSM), Random Forest Classifier (RFC), and Support Vector Machine (SVM) on the fused dataset. The dataset has been acquired over a part of Indian Himalayan terrain i.e., Uttarakhand State, India using microwave-based ISRO’s Scatterometer Satellite (SCATSAT-1) and optical-based NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS). The results show the effectiveness of the RFC classifier in the mapping of land surface features as compared to other classification algorithms i.e., LSM and SVM. This study not only highlights the potential of the RFC classifier in the extraction of information but also, shows the significance of fusion of optical and microwave datasets in the extraction of important Earth surface features.


INTRODUCTION
The Himalayan region is one of the important sources of water for south Asian countries but due to the climate variability, the occurrence of flash floods and avalanches is also increased (Gurung et al., 2011). Remote sensing via spaceborne sensors plays an important role to provide the Earth surface imagery on daily basis. Therefore, the continuous monitoring of the Himalayas is essential to mitigate the impact of natural hazards (Ballesteros-Cánovas et al., 2018). Since past few decades, many improvements have been made in the spaceborne sensor in terms of spectral and spatial resolution to improve the pictorial representation and extract the crucial parameters from the Earth surface (Amro et al., 2011;Kahraman and Ertürk, 2017;Snehmani et al., 2017;Vivone et al., 2015).
The optical-based sensors acquire the Earth information in a huge range of wavelengths but are generally affected in the presence of differential illuminations and atmospheric conditions at a time of acquisitions (Rahman et al., 2010). Moreover, the optical sensors are not capable enough to penetrate through the clouds and thus, the applicability is limited (Tsai et al., 2019). On the other hand, microwave sensors depend on their source of illumination at microwave frequencies and thus, not much affected by differential illuminations and penetrate through the clouds (Ulaby and Long, 2015). Moreover, microwave-based sensors can also penetrate the surface to a certain extent depending upon the operating frequency and dielectric constant of the object . The microwave sensors provide useful information which may not be fetched from the multispectral or hyperspectral sensors but optical sensors are more competent in terms of reflective characteristics of the target (Rahman et al., 2010;Xiao et al., 2020).
Image fusion is a process of combining multiple images to improve visual perception and deliver more meaningful and suitable information for computer processing (Kulkarni and Rege, 2020). The fused dataset can be proven as useful in many scientific applications as crop yield estimation, forest cover mapping, management of natural resources and disaster management (Ghassemian, 2016). There are many optical and microwave sensors available with different technical features and characteristics. Here, we have considered Moderate Resolution Imaging Spectroradiometer (MODIS) as optical-based dataset and Scatterometer Satellite (SCATSAT-1) as microwave dataset because both datasets are freely available and provide the global imagery with more than 90% of coverage on daily basis. Moreover, many products are available for both datasets that can be utilized to enhance its applications ranges. In previous literature, different image fusion algorithms have been tested to fuse the optical and microwave imagery and nearest-neighbour based fusion (NNF) is found to be more accurate to compute the classified or change maps (Singh et al., 2021a). Therefore, the NNF approach has been utilized in the present work. The classification of the fused dataset is one of the critical tasks in the accurate estimation of land use/cover type using the fused dataset. The classification algorithms are generally categorized into three categories based on their training procedure (Sood et al., 2018) such as (a) unsupervised that does not require inputs training sample from the user and deliver the outcomes based on similarity, (b) supervised that uses the labelled dataset to classify the datasets, (c) semi-supervised takes the middle ground and labels the unlabeled data points using knowledge learned from a limited number of labelled training samples (Lu and Weng, 2007;Myint et al., 2011). The efficacy of any supervised classification algorithm is strongly dependent on the quality of samples used to train the classifier (Tuia et al., 2016). However, three well-defined supervised classifiers i.e., Linear Spectral Mixing (LSM) (Choodarathnakara et al., 2012;Du et al., 2014), Random Forest (RF) (Du et al., 2015;, and Support Vector Machine (SVM) (Pal andFoody, 2012, 2010) have been evaluated in the present work to estimate the different landcover types over the Himalayas. In the present article, the main aim is to estimate the land-cover types using fusion datasets and evaluate the applicability of well-defined classification algorithms. To perform the image fusion for optical and microwave datasets, the NNF approach has been performed as previous studies confirmed the effectiveness of fusion of SCATSAT-1 and optical data (Singh et al., 2022a). Afterwards, different classification algorithms have been implemented such as LSM, RF and SVM. At last, the qualitative analysis has been performed to compute the effectiveness of each algorithm. The outcomes of the study allow the effective utilization of fused datasets in many applications.

Study Area
The Indian State, Uttarakhand located in the north-western part of the country, has been selected to perform the methodology. It covers an area of about 55,483 sq. km lies between latitude 28°42′40″N to 31°27′43″N and longitude 77°35′02″E to 81°02′23″N as shown in Figure 1. The Uttarakhand lies within the Himalayas and is rich in natural resources such as the origin of many rivers, snow-clad mountain peaks, dense forests, and the exitance of many glaciers. Moreover, this state is also well known as Devbhumi (land of Gods) due to the existence of many Hindu pilgrimage's sites. This is one of the major reasons behind the attraction of tourism in the state. On the other hand, it is more susceptible to high erosion, rainstorms, earthquakes, flash floods, avalanches, and glacial lake outburst flood (GLOF) (Gautam et al., 2013;Mool et al., 2001;Raza et al., 2012;You et al., 2017). This state is the witness of numerous natural hazards especially in the past decades due to the change in climate variability. Recently, flash floods were reported due to the collapse of a hanging glacier (Shugar et al., 2021). This natural disaster resulted in numerous deaths, property damage and also, impacted the hydropower plants. Therefore, it is essential to monitor the snow/ice cover changes in this region to forecast or mitigate the impact of natural hazards Sood et al., 2020).

Satellite Dataset
There are two different datasets were utilized in this study i.e., ISRO's SCATSAT-1 as microwave data and NASA's MODIS as optical data. The SCATSAT-1 offers the daily-based backscattered coefficient (sigma-nought) at a frequency of 13.5 GHz in two different polarizations i.e., HH and VV. The SCATSAT-1 data is available in four-level of data products to enhance its applicability. On the other hand, the multispectral MODIS dataset offers information in 36 different spectral bands for different applications. However, we have utilized the first seven-band to merge with the microwave dataset. The band's details are shown in

METHODOLOGY
The methodology of the proposed work is shown in Figure 2. It is generally divided into four steps: (a) pre-processing, (b) image fusion, (c) classification, and (d) accuracy assessment.

Pre-processing
The SCATSAT-1 level 4 products involved the various types of corrections as explained in different studies (Mankad et al., 2019;Misra et al., 2019). The coded values are converted into dB according to equation (1).
Where = slope of product parameters; 0 = image coordinates; = offset value that normally considered as 0.001

Image fusion
To overcome the limitation of optical and microwave datasets and improve the extraction of land surface parameters, the image fusion approaches are very successful (Snehmani et al., 2017). Particularly, with the availability of optical and microwave datasets on daily basis at the global level (Kulkarni and Rege, 2020). They provide complementary information to deliver enhanced data products that can be utilized in numerous applications. Previous studies based on image fusion have been tested in numerous applications such as forest classification, land cover mapping, urban mapping and snow cover applications (Amro et al., 2011;Byun et al., 2015;Mishra and Susaki, 2014;Singh et al., 2020). The fusion of different sensors generally faces a problem of misalignment but scatterometers are sensitive towards the water contents within the snow which may helps in accurate detection of snow parameters (Singh et al., 2022b).
To perform the fusion on SCATSAT-1 and multispectral MODIS, the NNF approach has been followed. It is based on a nearest-neighbour approach that involved resampling, computation of the difference factor and linear mixing (Singh et al., 2021a;Sun et al., 2014). Here, both the datasets have been resampled at 500 m resolution and then, the difference factor is computed to estimate the similarity between low-resolution pixels and nearest-neighbour nine pixels. The value of the difference factor is calculated as a summation of the shortest geodesic distance that represents the difference of pixels in microwave data with respect to neighbour super-pixels in the optical dataset. It reconstructs the cloud containment pixels in the multispectral optical dataset (Singh et al., 2021b).

Classification
Classification is a standard procedure to generate thematic maps that correspond to the different class categories. Here, three different classifiers i.e., LSM, RFC and SVM have been implemented on the fused dataset. The LSM is a supervised classifier is one of the unique approaches to solve the mixed problem by assigning more than one class category to a particular pixel. It is very helpful to improve the outcomes of the coarse resolution datasets such as MODIS and SCATSAT-1. Whereas, the RFC as supervised is one of the widely used classification algorithms in solving many remote sensing problems. It comprises many trees and provides the means of averaging the prediction of various decision trees to avoid the problem of overfitting. It generally involves the selection of trees, optimum nodes, input spectral bands and exit criteria (Du et al., 2015;Pal, 2005). SVM is a supervised classifier that handles complex datasets to achieve better results (Vapnik, 2013). It involves structure-risk minimization to minimize the error rate in the training sampling procedure and improves the performance in the classification process as compared to other classification methods (Liu et al., 2018). On the other hand, support vectors also increase the complexity level of the algorithm. To compute the SVM, it is essential to select the various associated parameters such as degree polynomial, kernel function, threshold value and bias. The detailed information regarding the hyperspectral parameters and training dataset can be found in previous studies (Singh et al., 2022a;Sood et al., 2020a).

Accuracy Assessment
Accuracy assessment is essential to be applied on each thematic map to compute the efficacy of each algorithm. It plays a vital role in any classification project to compare the thematic image with the reference dataset. The major components include producer's accuracy (PA), user's accuracy (UA), overall accuracy (OA), and Kappa-coefficient (Kc).

RESULT ANALYSIS AND DISCUSSION
In the present work, two dairy-based datasets i.e., ISRO's SCATSAT-1 (microwave) and NASA's MODIS (optical) have been acquired on 4th May 2020 as shown in Figures 3(a) and 3(b), respectively. To integrate both datasets, the NNF-based fusion approach has been followed. It combines both optical and microwave datasets based on the nearest-neighbourhood method. The outcome of the NNF-based image fusion has been shown in Figure 3(c). It is noteworthy that microwave data can easily penetrate through the cloud whereas optical data is impacted in the presence of the clouds. But with the merging of both microwave and optical datasets, the impact of cloud has been reduced in the optical dataset which can be visualized in Figure  3(a)-(c).
Afterwards, three well-defined classifiers i.e., LSM, RFC and SVM have were implemented on the fused dataset. It is essential to check the performance of different classifier algorithms on fused datasets so that better outcomes can be attained. Figure 3 (d)-(f) has shown the comparison of thematic or classified maps generated via LSM, RFC and SVM classification algorithms. From the visual analysis, the outcomes are very close to each other. Therefore, the qualitative analysis (accuracy assessment) has also been computed as shown in Table 2. From the qualitative analysis, it has been observed that all the classifiers performed well enough (with OA of 93.99 -94.15%).
But RFC performed marginally better (94.11 %) as compared to other classification algorithms i.e., LSM (94.11%) and SVM (93.99%). It has also been observed that in the snow category, the PA has been impacted in LSM (78.72%), RFC (88.871%) and SVM (80.59) which may be the reason behind such results.
The integration of SCATSAT-1 and MODIS dataset on a rugged terrain Himalayan region has a different degree of utilization as compared to other optical image fusion algorithms where the removal of cloud is a major concern. The image fusion and classification of fused dataset have an advantage in many remote sensing applications. It is expected that with the proposed work, many applications can be explored by merging the optical properties and scattering mechanisms of the object.

CONCLUSION
Based on the experimental outcomes, it is concluded that the image fusion of microwave-based SCATSAT-1 and opticalbased MODIS offers better imagery as compared to the original images. This study has been evaluated the three classification algorithms on the fused dataset on a rugged terrain Himalayas region (Uttarakhand state, India). All the classification algorithms delivered satisfactory outcomes but RFC provided marginally better accuracy as compared to LSM and SVM. Future studies may incorporate the deep learning methods for the automatic extraction earth surface features. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2022 XXIV ISPRS Congress (2022 edition), 6-11 June 2022, Nice, France