Fusion Methods and Multi-Classiers for Improving Land Cover Estimation By Remote Sensing Analysis

Adopting a low spatial resolution remote sensing imagery to get an accurate estimation of land-use and land-cover (LU/LC) is a very dicult task to perform. Image fusion plays a big role to map the LU/LC. Therefore, This study aims to nd out a rening method for the LU/LC estimating by adopting these steps; (1) apply a three pan-sharpening fusion approaches to combine panchromatic (PAN) imagery has high spatial resolution with multispectral (MS) imagery has low spatial resolution, (2) employing ve pixel-based classier approaches on MS and fused images; articial neural net (ANN), support vector machine (SVM), parallelepiped (PP), Mahalanobis distance (Mah) and spectral angle mapper (SAM), (3) Make a statistical comparison between classication results. The Landsat-8 image was adopted for this research. There are twenty LU/LC thematic maps were created in this study. A suitable and reliable LU/LC method was presented based on the obtained results. The validations of the results were performed by adopting a confusion matrix. A comparison made between the classication results of MS and all fused images levels. It proved that mapping the LU/LC produced by Gram-Schmidt Pan-sharpening (GS) and classied by SVM method has the most accurate result among all other MS and fused images that classied by the other classiers, it has an overall accuracy about (99.85%) and a kappa coecient of about (0.98). However, the SAM algorithm has the lowest accuracy compared to all other adopted methods, with overall accuracy of 53.41% and the kappa coecient of about 0.48. The proposed procedure is useful in the industry and academic side for estimating purposes. In addition, it is also a good tool for analysts and researchers, who could interest to extend the technique to employ different datasets and regions.


Introduction
An accurate thematic map of LU/LC plays a big role in different remote sensing applications such as; change detection, environment managing and monitoring, LU/LC detection, hazard prediction, urban area expansion, forest monitoring and other (Sang et al., 2014;Khatami et al., 2016;Dibs et al, 2017;Karar et al., 2020). Image fusion plays a big role to re ne and improve the estimation of LU/LC.
In other hand, remote sensing is a powerful tool and very useful for mapping the LU/LC from using a suitable satellite images with a good selecting of classi cation method. However, image classi cation approaches consider as the best method to monitor, manage and estimate the LU/LC (Dibs, 2013;Sang et al., 2014). To perform classi cation, it needs to involve different stages such as selection training and testing samples, atmospheric correction, radiometric correction, geometric correction, objects extraction, classi er method selection, post-classi cation process, and performing results validation (Singh et al., 2014;Dixon et al., 2015;Hayder et al., 2018;Dibs, 2018).
The Selection of a reliable classi er technique is very critical to obtain an accurate LU/LC thematic map (Dixon et al., 2015;Li et al., 2017). For LU/LC estimating there are large numbers of techniques and methodologies to apply, some of these classi ers under pixel-based and other under object-based, these algorithms such as the SVM, ANN, SAM, PP, Decision Trees (DT) as discussed by (Chasmer et  Additionally, many improved techniques have been applied to improve LU/LC mapping such as the image pan-sharpening technique (Ghosh et al., 2014;Cavur et al., 2019). These approaches can be divided into many categories; component substitution techniques, multi-resolution dataset analysis (Li et al., 2017).
Firstly, there are two basic types of fusion pan-sharpening methods; regarding color, statistical and numerical algorithms (Ma et al., 2019). The most commonly adopted method is regarding to component substitution. Intensity hue saturation spectral sharpening (IHS) method is one of the commonly employed methods of the IHS group (Li et al., 2017). IHS works based on color space transformation (Paidamwoyo et al., 2020). However, the GS method is a new generation of pan-sharpening approaches of deep learning, it has been adopted widely in previous years, it relies on the applications of color transform and it converts low-resolution multi-spectral band to a new color system that differences in both spatial and spectral information and details (Paidamwoyo et al., 2020). The principal component analysis (PCA) method is another one to use, and it works based on a statistical method, therefore, PCA is included under the group of statistical methods (Cavur et al., 2019). The Brovey method is a multiplicative approach, it is modi ed by normalization of the results (Elatawneh et al., 2014). Many studies discuss imagery pansharpening between PAN and MS images (Sang et al, 2014;Khatami et al., 2016;Li et al., 2018;Hayder et al., 2020). The spatial resolution will enhance when, replacing the PAN imagery that has high-spatial resolution by the MS image that has high-spectral resolution without saving all spectral information Azarang and Kehtarnava, 2020). The purposes behind using imagery pan-sharpening method are; (1) upsurging of spatial resolution, (2) advancing of geometric accurateness, (3) improving topography presentation, (4) re ning of classi cation precision (Ma et al., 2019).
There are several pan-sharpening methods that have been adopted using remotely sensed data throughout the world. However, some key unanswered questions: (a) does incorporate PAN imagery will support the LU/LC mapping? (b) What is the best pansharpening fusion method between the Landsat MS and PAN data? (c) What is the best algorithm to classify to produce LU/LC Landsat data? To address all these questions, the current research focuses on investigating the pansharpening of PAN an MS Landsat images and examine it with different pixel-based classi cation approaches to propose an improved procedure for estimating LU/LC. The layout of this article will start with the used materials and methods section, and then go through collecting the truth dataset, performing an image noise removal (geometric and radiometric noise) for all images. The next step will be conducting the image fusion levels. Then, apply different classi cation methods on the Multi-spectral and fused images. The Discussion section will show the discuss in deep all the outcomes of these processing and analysis to get the most accurate methodology to map the LU/LC. The outcomes of this study will help to provide a big contribution to industry and academic elds. Analysts and researchers can improve, develop and extend the present method to work and apply on different dataset sources and regions.

Material And Methods
For this research different processing, analyzing and integrating methods were adopted to nd an appropriate procedure for generating the LU/LC map. In this study, after downloading the used datasets from Landsat-8 satellite PAN and MS imagery of 2018, the noise removed is starting to perform in order to remove and reduce images radiometric and geometric errors. The next step, it was satellite images resampling by using bilinear approach. Then the Landsat-8 PAN and MS images become ready for further processing and analyzing stages. After that, the research has two different procedures to perform estimating the LU/LC map of the study area. Firstly, the Landsat-8 MS images go classi ed with a ve pixel-based classi er approaches (PP, ANN, SVM, SAM, and Mah), with selecting training and testing sites to conduct each classi cation algorithm. Secondly, the Landsat-8 PAN and MS satellite images were analyzing and integrating together with a three different kinds of image fusion levels (Gram-Schmidt Pansharpening, intensity hue saturation spectral sharpening and Brovey pan-sharpening method). Each level of images fusion go through different the ve pixel-based classi cation methods with using the same collected training and testing sites of the PAN and MS satellite images to examine which methodology will provide the most accurate result to produce LU/LC map. As indicated above the MS and fused satellite images will be classi ed twenty times for the purpose of this study. After that, a confusion matrix will apply on the results of the twenty classi ed images to validate their accuracy. Then, the outcomes of all the previous stages will examine from making a statistical comparison between them. Figure 1 indicates the adopted method for this study.

Study Area Description
In this research, the Baghdad city in Iraq was selected as the study area. Baghdad city is a very famous city in Iraq, and it is considered as the second-largest city in the Arab world after Cairo city. It has a location along the Tigris River. In the eighth century, Baghdad city has a golden history, it became the Abbasid caliphate capital city that time. Baghdad has a signi cant in both commercial and cultural elds in the Arab world. It has a population of about 6,719,500 person regarding to the estimate of 2018, this population value makes this city as one of the biggest cities in there public of Iraq. It is located in 44° 27' 54.37'' Easting and 33° 23' 03.98'' Northing. The area of Baghdad city is around 204.2 km². The altitude ranges of Baghdad city in between (32-38) m above the mean sea level (MSL). Baghdad city is almost covered by urban areas. Figure 2 indicates the location of an interesting area of this study (Hayder et al., 2020).

Satellite images and Truth Dataset
The analyzed satellite imagery for this study was obtained from the Landsat-8 sensor. This sensor is launched into space on 11/2/2013. Landsat sensor has carried two different sensors, the operational land imager (OLI) and Thermal Infrared Sensor (TIRS). Landsat satellite data has (11) bands, some of them have a spatial resolution of about 30m for each band of (1 to 7 & 9). However, the PAN channel (8 band), it has a spatial resolution of about 15m. In addition, the thermal bands (10 and 11) have a spatial resolution of 100m. Table 1 describes the speci cations of Landsat sensor bands. The dataset for this research was freely downloaded from the Glovis website (https://glovis.usgs.gov/app) with path=168 and row=37. The satellite image was captured on 20/2/2018 and it has level processing 1T standard correction, UTM projection with zone 38 N, and datum a WGS 84. The processed image has no cloud.
The ground truth data must be observed to apply the pixel-based classi cations. The truth data usually collect from using different methods such as collecting GPS references in eldwork and/or higher resolution remotely sensed imagery (Lu, 2011;Hayder and Suhad, 2019;Hayder et al., 2020). However, in this research, the authors used Google Earth Pro to collect the training and testing samples in the image by visual interpretation process for each class. The Google Earth Pro image has a very high spatial resolution, and that will help to discover the located features in the study area (Hayder et al., 2020). For this study, ve classes were selected to be used in estimating of LU/LC map, and they are; urban area, water body, soil area, roads and vegetation, respectively. Selecting these ve classes was made based on the regular features that distribute in the study area as indicate in Figure 3. However, urban area de nes all build-up areas such as building areas and/or housing area or any other kind of buildings (commercial, education, plaza and so on). In other side, water body class de nes all water bodies located in the research area such as (rivers, marshes, and lake) and soil class represents all the barren lands area. However, the roads class represents the main roads located in the Baghdad city that covered by asphalt layer. The last class was the vegetation class, and it represents evergreen lands and the area that covered with vegetation, whether natural or cultivated by humans. Randomly training and testing sites procedure were adopted for all the ve class. The training and testing sites were equally distributed overall in the image of the research area to ensure get an accurate classi cation outcome. For every single class, there are more than 250 pixels were collected. Figure 3 indicates the ground truth datasets that collected to involve in image classi cation processes.

Image Noise Removal
The geometric correction (GC) for any satellite images is required before performing any processing and analyzing on satellite images Hayder and Suhad, 2019;Aysar et al., 2020). A good selection of ground control points (GCPs) location should be done (Hayder and Suhad, 2019). In this study, the geometric correction of Landsat MS and PAN images was performed using ten GCPs, which regularly distributed throughout the image portions. These GCPs were collected using the Google Earth image as mention in section (2.2). The rst polynomial transformation and the nearest neighbor were adopted to obtain a root mean square error (RMSE), and it was about 1.32 pixels. The next correction for Landsat images was performed a radiometric correction. It is an essential algorithm for image preprocessing to remove the effects of sun illumination (Bello and Parviz, 2013;Sang et al., 2014;Hayder and Suhad, 2019;Hashim et al., 2020 a & b;Aysar et al., 2020) . The Dark Object Subtraction (DOS) was adopted to remove the radiance errors of MS and Pan. Figure 4 (a & b) indicates the corrected MS and PAN images after removing all kind of noise.

Image Fusion Levels
After conducting image layer stacking and sub-setting, the MS and PAN images will integrate together with using of a three different fusion pan-sharpening spectral methods. The rst image fusion level was performed using the IHS approach. The IHS spectral sharpening method is usually adopted in imagery fusion to use the MS image complementary nature (Jain et al., 2019;Saha et al., 2019). For the spectral IHS sharpening, each of R, G and B bands of the MS data were converted to this component (Zhong et al., 2016).The PAN imagery histogram matched to the MS data intensity component (Jain et al., 2019;Hayder et al., 2020). Then, the intensity component was replaced by the PAN data. Then, the inverse transformation was conducted in order to get the MS image that has a high resolution. The pixel size of the outcome RGB imagery will have the same as the input PAN image of high-resolution. Figure 5 describes the steps of the adopted method. The second pan-sharpening level was applied with using the Brovey method. This method adopts a mathematical combination to make integration between the high and low-resolution bands (Liu, 2018) For this method, each MS band will multiply by a ratio of the band of high resolution, then divided by the MS band. The outputs of IHS processing will automatically resample the three MS bands to the PAN pixel size (Paidamwoyo et al., 2020). The result of RGB imagery will have the pixel size of the input highresolution data (Bovolo, 2010). The Brovey method equation is de ned below: where (DN) represents as a particular band digital number and (bi) is the MS image particular band (Ma et al., 2019). The third applied image fusion level was performed using Gram-Schmidt spectral sharpening (GS) algorithm. The GS sharpening method enhances the MS band's spatial resolution by integrating high and low image resolutions (Ma et al., 2019). The GS transformation conducts to the simulated high-resolution PAN band with the MS low-resolution bands. The simulated PAN high-resolution image band is adopted at the rst. Then the PAN data will be replaced with the GS band (Paidamwoyo et al., 2020). The last step is inversed transformation will apply to generate the spectral sharpened MS band (Yuan et al., 2018). Figure 6 (a, b & c) shows the fused images after employed the IHS, Brovey and GS spectral sharpening algorithms.

Classi cations Of Multi-spectral And Pan-sharpening Images
Many classi cation approaches have been adopted and applied for mapping the LU/LC (Yifang and Alexander, 2013). However, performing image classi cation needs to collect the training and testing samples for each class to guide all the processes of classi cations and accuracy assessments of the output results (Rwanga and Ndambuki, 2017). All the training sites will comprise to the corresponding group of the region of interests (ROIs). However, the candidates' sample groups from the same class may be spectrally (Yifang and Alexander, 2013). Therefore, wide candidate pixels should be sampled (Paidamwoyo et al., 2020) There are many different supervised classi ers adopted in different remote sensing applications include PP, minimum distance, ANN, Mah, spectral information divergence, SVM, binary encoding and SAM methods (Zoleikani et al., 2017). In this study, ve pixel-based classi er methods (PP, ANN, SVM, SAM, and Mah) were adopted to classify the MS data and the three fusion methods that have a good e ciency when apply on data has low spatial resolution (Taubenböck et al., 2012). The weights of the ANN method were used as uniform distribution. Values of about 0.001 and 100 were employed for learning rate for the output layer and hidden layer, respectively. So, the stopping criteria on (0.001) were xed. However, applying the SVM approach was employed based on the default parameters, because authors for this research want to examine different classi ers, they not focus on different parameters of the SVM method. The applied SVM parameters in this research, it included of using a radial basis function as a kernel type, for gamma in kernel function. In other hand, the penalty parameter and pyramid levels were 0.167, 100.00 and 0.00, respectively. The adopted classi cation probability threshold value was zero. The ve supervised pixel-based classi cation techniques were evaluated for this research using the confusion matrix (Paidamwoyo et al., 2020;Zoleikani et al., 2017). Both of overall accuracy and the kappa coe cient are widely used for quality assessment of classi cation results (Li et al., 2012;Pushparaj et al., 2017). In this study, these observation methods and their equations are presented below: where (n) is a total number of pixels, (n ij ) equal to the classi ed pixels total number, (n i ) is instances number, label (i) that has been classi ed in the label (j).

Landsat-8 OLI Multi-spectral image Classi cation
The rst stage of image classi cation was performed by applying the supervised classi ers to classify the multispectral image of Landsat OLI by SVM, ANN, PP, Mah and SAM methods of Baghdad city and then produce the LU/LC thematic maps. The results of these classi ers were ve thematic maps. The image processing steps were performed using the Envi 5.3 environment. The confusion matrix method was adopted to validate the classi cation results (Li et al., 2010;Li et al., 2018;Azarang and Kehtarnava, 2020). The validation of all classi cations indicates that the SVM has the highest accuracy compared to all other adopted methods with overall accuracy (93.25) and kappa coe cient (0.92), and in both sides the statistically and visually. Figure 7 shows the thematic map of LU/LC. The IHS pan-sharpening imagery generates from integrating the MS imagery that has low-spatialresolution and PAN imagery that has a high-spatial-resolution to re ne and enhance the LU/LC mapping of Baghdad city, and also to obtain the highest accuracy procedure of estimating the LU/LC. The pansharpening fused image classi ed by employed methods of; Mah, ANN, SVM, SAM, and PP. The confusion matrix was applied in order to perform results evaluation of the ve output results. A statistical comparison was made between the results of all above classi ers to get an accurate result. Statistically, the SVM method illustrates the highest overall accuracy about (98.56%) and kappa coe cient about (0.96) for IHS fused images as shown in Table 3. Figure 8 shows the results of LU/LC classi cations of the ve classi ers on HIS fused image.

Brovey sharpening Fused Image Classi cation
Another fusion method is called the Brovey sharpening approach was adopted for this study to improve and enhance the estimating and mapping of LU/LC. Several types of classi cation algorithms were applied on the fused image by the Brovey sharpening approach; SVM, ANN, PP, Mah and SAM methods to map LU/LC. The confusion matrix function once again was adopted to assess the result of the ve classi ers. Statistically, the SVM shows the highest OA about (98.7%) with a kappa coe cient of (0.97). Then, the obtained results were compared the classi cation results between the only Landsat MS image and the image fused by Brovey pan-sharpening method in order to assess the role of involving PAN data for LU/LC mapping and to examine if the image fusion will improve and enhance the accuracy results of LU/LC classi cation. Figure 9 and Table 4 reveal the LU/LC map produced by integrating PAN and MS data and classi ed with using several types of supervised pixel base classi ers.

Gram-Schmidt Sharpening Fused Image Classi cation
The third spectral pan-sharpening method applied for this research was the GS sharpening algorithm. The fused image was classi ed by applying also the same classi cation approaches for previous steps: (SVM, SAM, Mah, PP and ANN) in order to estimate the LU/LC of Baghdad city. All the results of the fused image classi cations were evaluated using the confusion matrix technique. A statistical comparison was performed to all the results of the ve classi ers in order to determine which methodology has the most accurate result. The comparison shows that using the SVM approach to classify the fused image has the highest OA about (99.85%) with a kappa coe cient of (0.98). Figures 10 and Table 5 are illustrated the integrating of the PAN and MS images. The GS spectral pan-sharpening method with the SVM classi cation method reveals high improvement for image classi cation to generate the LU/LC maps.

Discussion
Figures 11 and 12 are representing the results of all the applied classi cation methods; ANN, SVM, Mah, PP and SAM in this research regarding to the level of overall accuracy and kappa coe cient that applied on the MS and the three fused pan-sharpening image by each of (IHS, Brovey and GS) sharpening algorithms. The comparison was made for this research regarding to the twenty created LU/LC thematic maps and the results of all the overall accuracies and kappa coe cients of all the classi ers approaches from the MS and the three image fusion levels. One of the most di culties tasks that facing this research for images fusion was the images has different spatial resolutions. The MS image has a low-spatial resolution of about 30 m and for PAN image has a high-spatial resolution of about 15 m, and the image fusion provides superior spatial details and information (Xing et al., 2018). Different researches deal with imagery fusion between PAN and MS images, it conducts from combining the PAN image that has features with high-frequency with the spectral information of MS image that has features with lowfrequency (Azarang and Kehtarnava, 2020). Replacing the MS image high-frequency features with the PAN image high-frequency features, will enhance the spatial resolution with loss of some spectral information (Azarang and Kehtarnava, 2020). Therefore, for this research as indicated previously in Figure 1. Image resampling was made by using a bilinear approach and the resampling process was performed with Envi software. So, by resampling the spatial resolution of MS image from using image fusion with PAN image, it can be obtained a good results and also enhance the LU/LC estimation map. Figure 12 indicates that the all-accurate assessment values were obtained for all the adopted approaches form each classi ed image of the MS data and the three spectral pan-sharpening fused images. The SVM method was provided the best performance when applied on data of MS Landsat and PAN when they integrating together using the GS pan-sharpening technique. The classi cation outputs reveal that the accuracy obtained from adopting the SVM approach provides the highest results, the overall accuracy of about (99.85%) with a kappa coe cient of about (0.98) from image classi cation. However, the SAM classi cation of the fused image using IHS spectral pan-sharpening method shows the lowest accuracy overall images classi cations by representing an overall accuracy of (53.41%) with the kappa coe cient about (0.48). The research aims to investigate and nd out the possibility of using the PAN data to improve the estimation accuracy of the LU/LC thematic map. Therefore, based on all the results of this study, it is found that the optimal methodology to obtain the highest results for generating the LU/LC thematic map for Baghdad city is by performing image integration of MS and PAN data using GS spectral pan-sharpening method and classify with the SVM method. Figure 13 illustrates the LU/LC estimated map of Baghdad city, this thematic map has ve different classes (urbanization area, vegetation area, water bodies, soil area and roads).

Conclusion
This study investigates and analyzes the use of Landsat-8 OLI both of MS and PAN datasets in order to nd the best and an accurate method for LU/LC estimating in the area of Baghdad city, Iraq by performing a statistical comparison between many classi cation approaches (SVM, SAM, Mah, PP and ANN) were applied on MS images and other three pan-sharpening fused images by IHS, Brovey and GS methods. The re ned producer was proposed for LU/LC mapping regarding to the obtained results. The results validation was conducted by applying the confusion matrix. The obtained overall accuracy and kappa coe cient from applying the SVM classi er on the fused imagery by the GS spectral sharpening algorithm shows the highest accurate result over all other classi ers and the use of IHS and Brovey spectral sharpening fusion methods. The SVM approach achieves the highest results among all classi cation methods with different levels of image classi cations; (1) with MS image, it is provided OA about 93.25% and kappa coe cient 0.92; (2) with classi cation of integrating IHS pan-sharpening spectral with MS image, SVM provides OA of 98.56% and kappa coe cient about 0.96; (3) with classi cation of integrating Brovey sharpening spectral, it provides OA about 98.7% with a kappa coe cient of 0.97, and (4) with Gram-Schmidt Sharpening, SVM achieves OA about 99.85% with a kappa coe cient of 0.98. However, the SAM algorithm has the lowest accuracy compared to all other adopted methods, with OA 53.41% and the kappa coe cient about 0.48. Therefore, the outcome results con rm that the image fusion using the GS spectral algorithm and SVM classi er was determined as the best technique to estimate the thematic map of LU/LC for this study. In future work, object-based approaches and methods should be examined and compared to the results of classi cation methods of pixel base. In addition, it should be trying to use satellite imagery has high spatial and spectral resolution; such as QuickBird, worldview-3, SPOT series, and IKONOS satellite systems.

Declarations
Con icts of Interest: The authors declare no con ict of interest.