Optimum supervised classification algorithm identification by investigating PlanetScope and Skysat multispectral satellite data of Covid lockdown

This research identifies the optimum supervised classification algorithm based on modeling Covid 19 lockdown situations all around the World. The deadly Covid 19 viruses suddenly stopped the fast-moving world and all the commercial and noncommercial activities were stalled for an uncertain period during 2020-2021. In this work, object-based image classification approaches have been used to compare pre-Covid and post-Covid (at the time lockdown) images of the study area. These study areas are Washington DC, USA, Sao Paulo, Brazil, Cairo, Egypt, Afghanistan/Iran border, and Beijing, China. All the study areas possess different geographical conditions but have a similar situation of Covid 19 lockdowns. Six supervised image classification techniques are known as Parallelepiped classification (PPC), Minimum distance classification (MDC), Mahalanobis distance classification (MaDC), Maximum likelihood classification (MLC), Spectral angle mapper classification (SAMC) and Spectral information divergence classification (SIDC) are used to classify the satellite data of the study area. Thus based on classification results and statistical features, it has been observed that PPChas obtained the least significant results. In contrast, the most reliable results and highest classification accuracies are obtained through MDC, MaDC, and MLCclassification algorithms.

river and canal mapping ( EL-Arafy, 2021 ), etc. Image classificationbased change estimation is widely used in remote sensing applications. Change estimation is broadly calculated using two different methodologies, i.e ., pixel-based change estimation Ramola et al., 2020 ) and object-based change estimation ( Amitrano et al., 2021 ). Object-based change estimation classifies the data in a small set of objects which are further investigated to determine the changes developed in the study area. Pahlefia et al. (2021) used 50 Sentinel 2A images and a Google Earth engine to investigate tropical Savannah grassland using image classification-based principle component analysis (PCA) and P P C algorithm. Their classification has obtained an overall accuracy of 82.86%. Lv and Wang (2020) applied image classification techniques on hyperspectral data to accurately classify the ground truth data. They have compared classified images based on supervised, semi-supervised, and unsupervised image classificationbased algorithms. They have tested several classification algorithms like SV M, P P C , MDC, and MLC algorithms. C N N and deep learning-based deep relief networks ( DRN) are also used for semisupervised image classification. Finally, k-mean classification and Isodata algorithms are used for the unsupervised image classification. Li et al. (2018) used MaDC classification approach to test hyperspectral remote sensing data. They have developed a new classification approach having a combination of MaDC with kernel features. They have observed that their progressive system is performing significantly well when compared with the conventional MaDC approach with improved classification results and classification accuracy. ( Verbovšek and Popit, 2018 ) classified various "litho-geomorphological units" in two carbonate plateaus, alluvial deposits, steep carbonate cliffs, quaternary deposits, slope deposits, flysch, etc., using a supervised image classification based MLC approach. Their investigation suggests that MLC is an appropriate approach to identifying various lithological compositions. Thus image classification methods not only present classification results of the exterior land use/land cover but also have the potential to distinguish the chemical composition of the various "lithogeomorphological units". ( Gopinath et al., 2020 ) investigated the land use/land cover of the Kozhikode district, Kerala, India, using the SAMC and Support vector machine ( SV M) classification algorithms. Here SV M presents better results than the SAMC approach with overall classification accuracy of 85.6% and Kappa coefficient ( K p) of 0.89. Mustapha et al. (2020) used convolution neural network ( CN N ) and SV M based approch to model crowd intensity on the pedestrian bridge using machine learning-based approaches. They have used single-class motion speed classification for multiclass motion, which received a classification accuracy of 91%, with load estimation regression reaching 9%. Thus, it can be understood that various events can be classified through image classification techniques. The classification accuracy and K p further determine the level of the classified data.
This study is organized into six different sections. Section 2 presents background information on various supervised classification algorithms used in this work. Section 3 illustrates details of the study area and background information on the Skysat and PlanetScope satellite systems. Section 4 presents the classification results and feature quantification of the case studies. Section 5 offers a discussion on the performed work. Finally, Section 6 presents a conclusive remark on the research work.

Background information on various image classification algorithms
Image classification is the technique in which pixels or digital numbers ( DN) of the image are categorized into multiple classes (objects) based on different image classification algorithms. It can be understood as mapping the image DN and symbols ( Hong et al., 2020 ). The general methodology of the image classification is expressed by Eq. (1) ( Hong et al., 2020 ).
where B = numbers of bands, C = numbers of classes, here if f (. ) is a function assigning a pixel vector Y to a "single class" in the "set of class". In remote sensing applications, data classification involves the computer program and studying the relationship between the image data, classified data, and the associated parameters ( Shakya et al., 2018 ). Depending upon the type of interaction between image DN and the computer interpreter classification process is divided into two different categories, i.e. , "supervised image learning ( SIL )" and "unsupervised image learning ( UIL ) ( Schmarje et al., 2021 )." In SIL the investigator categorizes the image DNs by selecting an appropriate SIL algorithm for classifying various land-use/landcover ( LU/LC) . The investigation area under these algorithms is known as cover type. These algorithms compute the reflectance value of each DN, then compare the obtained data with each classification category. Finally, the DN value is assigned to the classification category having the maximum possibility of getting selected for the individual class. There are three different steps for classifying remote sensing-based satellite data. These steps are expressed as training the satellite data, classifying the data using an appropriate image classification approach, and finally, accuracy assessment of the performed category. These are further explained in steps 1-3.
Step 1: The satellite data investigator selects a Region of interest ( ROI) from the whole study area to develop a numerical description of the spectral attributes of the LU/LC. Step 2: Appropriate image classification scheme is selected, and the classification of the image data is performed based on the ROI developed in step 1.
Step 3: Assessment of classification accuracy of the classified image.
In unsupervised image classification-based algorithms, image DN are assigned to a class without any prior information about the class. This process of transferring DN to an unknown class is known as "clustering" ( Kim et al., 2020 ). Here, the user predicts the number of clusters that need to be generated to classify the image. These cluster maps are compared with the primary reference data to identify a specific "class" or "cluster." Some prominent unsupervised image classification algorithms are Isodata classification, k-mean classification, decision tree classification, fuzzy classifiers, and clustering criteria classification algorithms Uzma et al., 2022 ).

Parallelepiped classification ( PPC )
This classification technique follows a simple detection principle in which the "highest" and the "lowest" brightness value in each spectral dimension is identified. Higher and lower values are determined through histogram signature plotting these spectral components ( Yua et al., 2022 ). This classification creates a box corresponding to each class to observe where an unknown pixel is assigned. Fig. 1 a represents the lower and higher spectral component selection. Fig. 1b illustrates the class formation for twodimensional parallelepipeds. Finally, Fig. 1 c depicts the inseparability region developed by the intersection of the two parallelepipeds. Advantages (a) This methodology has computational efficiency. The pixel classification under this methodology is easy to perform ( Mittal et al., 2021 ). (b) The classification of the image DNs under this methodology is relatively faster than other image classification algorithms ( Mafanya et al., 2022 ).

Shortcomings
The image DNs are appropriately classified under this approach as DNs get wrongly included in other classes. This classification algorithm is faster because classified DNs get wrongly involved in several other classes. Thus a DNs of an individual class gets distributed in several different classes. Due to this, the computation time gets reduced, but overall classification accuracy decreases significantly ( Mishra et al., 2011 ).

Minimum distance classification ( MDC )
This image classification method depends upon the number of the DNs used for classifying data. It relies on the quantity of training DNs used to train the data. If a classified class has not obtained sufficient numbers of DNs then an inaccurate covariance matrix may be formed, resulting in poor classification outcomes. This method also depends upon the mean vector from each classified class and does not use the covariance information of the covariance matrix. The data under investigation is used to calculate the mean of the class, and later, each DN is used to classify according to the class of the nearest mean value ( Zhang and Weng, 2016 ). Fig. 1 d represents the DN allotment methodology for the MDC approach. Here the classification data is categorized into four classes, i.e., Class A , Class B , Class C, and Class D. A cyan color triangle needs to be allotted to any class. Thus the mean distance of the cyan triangle from the individual class is computed, and identified Therefore it is observed that the cyan triangle has a minimum distance from the Class C. Thus the cyan triangle is assigned to the Class C. In this manner, different DNs are classified through MDC approach. This methodology of identifying the class means and later allotting each class means organized to the class of the nearest mean value is mathematically expressed by Eq. (2) . If C DN represents a column vector with the brightness value corresponding to each image pixel, then 1 ≤ DN ≤ M, and the image classification class is represented by the I C i .
If I C i is the mean of the training data corresponding to M class obtained from the training dataset. Thus the discernment function of the MDC classifier is expressed by Eq. (3) .
Advantages (a) This classification approach provides good accuracy as most of the DNs gets classified. The unclassified DNs under this approach remain low in number ( Pikies and Ali, 2021 ). (b) This classification approach allows fast classification and excitation of the image DNs in different classes ( Wang et al., 2018 ).

Shortcomings
The main problem while classifying the image data through this classification algorithm is the commission error which negatively affects the overall classification accuracy result ( Pacheco et al., 2021 ).

Mahalanobis distance classification ( MaDC )
The Mahalanobis distance classification between two multivariable quantities x and y is explained in Eq. (4) ( Chang et al., 2020 ).
It is assumed that if the covariance matrix is k.I, where I be a unit matrix, and k is a constant. Then MaDC is expressed as the scaled version of the Euclidean distance. Euclidean distance can be understood as the distance between two points, says m and n. The distance between m and n can be expressed in many forms. Let the coordinate corresponding to m, and n are ( x 1 , y 1 ) and ( x 2 , y 2 ) .
Distance for the points lying on the real axis.
Distance calculation for the higher degree.
The distance calculation for the higher degree has two different cartesian coordinates. The MaDC technique reduces the Euclidean distance according to the extent of variation by the covariance matrix . MaDC is also considered the opposite of the MLC to some extent as the initial point of the MaDC is the discriminant function of the MLC. When the sign is reversed, the expression becomes the square measure of the quadratic entry, and a constant term is expressed by Eq. (8) ( Chang et al., 2020 ).
The square distance term d ( x, m i ) 2 is used to classify the unknown image pixel, similar to the MDC algorithm. Considering the covariance generated by each class is equal, then the average class's expression is expressed by Eq. (9) ( Chang et al., 2020 ).
This squared distance is known as Mahalanobis distance.

Advantages
The most crucial advantage of MaDC classifier approach is that it can extract multi-variate outliers, combining two or more variables ( Sun et al., 2021 ). Shortcomings (a) One of the significant issues related to MaDC approach is the reliability of the obtained classification results ( Siddappa and Kampalappa, 2020 ). (b) The second major issue related to this classification approach is the "inverse of the correlation matrix," which is needed for the calculation. This becomes exceptionally tedious when the variables are highly correlated ( Zeng et al., 2021 ).

Maximum likelihood classification (MLC)
This is one of the most common image classification algorithms. It is exclusively used in remote sensing due to its superior ability to produce efficient results. This algorithm uses a probabilistic distribution model based on the Bayes theorem ( Aljemely et al., 2021 ). The main principle of this algorithm is to assign a DN to the object having "maximum probability," i.e., "(maximum likelihood)." The likelihood of a feature vector X that is to be included in a class C i having the conditional probability is expressed as P (C I /X ) . The conditional probability P (X/C I ) of the class C I . The given vector X needs to be calculated to determine the MLC classification. Finally, the direct estimation of a class's conditional probability is impossible. Instead, it is computed indirectly in terms of a feature vector X that belongs to the class C i . During the likelihood classification P (X/C I ) is assumed to be a multi-variate Gaussian distribution model. The Gaussian distribution model is simple and easy to handle. The expression of the MLC consists of conditional density function P (C I /X ) , mean vector ( μ i ) and covariance matrix i is expressed by Eq. (10) ( Verma et al., 2020 ).
While using the Gaussian distribution model, some assumptions are considered, which are expressed as follows: (a) Each classification class is assumed to multi-variate normal distribution. (b) This suggests that each class's mean value has the highest occurrence value. (c) An exponential decline is seen in the likelihood function as the features' value shifts from the mean vector. (d) The rate of the Gaussian decline is associated with the class variance. It is observed that the rate of decrease is inversely proportional to the variance.

Advantages
(a) This is considered one of the most accurate classification methods ( Lin et al., 2020 ). (b) The unclassified DN problem is undoubtedly reduced as most DNs get classified in this algorithm ( Ha et al., 2020 ). (c) The overlapping DN problem is this algorithm is also reduced as it consistently reduces the overlapping pixels between classes and assignment of the pixels ( Hu et al., 2020 ). (d) In this algorithm, when a DN is assigned to a class, the algorithm is already aware that the DN is allocated to the correct category. This assumption gets correct, producing good classification accuracy ( Zhao et al., 2019 ).

Shortcomings
(a) The DN, which is not assigned to any class, is set to be misclas-  . (d) This algorithm does not work on the data extracted from the spectrometer .
The working methodology of the MLC approach can be under in Fig. 2 a. If the DN falls under the range of ( μ ± σ ) then it belongs to the Class A . Similarly, when the scope of DN is ( μ ± 2 σ ) , it belongs to Class B . Finally, DN falling in the range of ( μ ± 3 σ ) will be classified under the C lass C . Thus the DN, which do not lie in these ranges, are treated as unclassified.

Spectral angle mapper classification ( SAMC )
In sizeable remote sensing datasets, computation of the covariance matrix, the inverse of the covariance matrix, and the distance of each DN is a complicated task. Even if a situation is considered when computation of ( x − μ) T −1 I ( x − μ) is possible even then, the computation of the inverse of the expression is tedious as there is an excellent possibility for the expression to become unstable. This can produce abrupt classification results. Thus in such a situation, SAMC algorithm comes to rescue the investigator. In this algorithm, 2 D features space is described for the bands, say B 1 and B 2 . These bands represent two different objects having vector notations as V 1 and V 2 . Then the Euclidian distance between the vector point is expressed by the distance d. Finally, a spectral angle is formed between the vectors, represented by Eq. (11) ( Tu et al., 2018 ). The spectral distance and spectral angle relationship between vectors V 1 and V 2 along with distance " d " is presented in Fig. 2 Advantages (a) The vector value is not affected by the minor change developed in the vector V 1 and V 2 ( Tu et al., 2018 ). (b) The classification computation is relatively easy compared to the MaDC algorithm, which uses MLC approach ( Tu et al., 2018 ).

Shortcomings
This approach is widely used for hyperspectral remote sensing applications. Hyperspectral remote sensing has several limitations as a high-performance computer system must process the data and obtain interrupted results. Super sensitive detectors and ample storage space are other requirements for this classification algorithm ( Ahmad et al., 2019 ).

Spectral information divergence classification ( SIDC )
SIDC is used to obtain the spectral similarity between the DN of an image to capture the spectral correlation. Let con- The information theory is used to express the information of p and q with Eq. (12) ( Khajehrayeni and Ghassemian, 2020 ).
The relative entropy of y can be calculated with the assistance of x and is expressed by Eq. (13) ( Khajehrayeni and Ghassemian, 2020 ).
Thus this technique of image classification is used for classifying remote sensing images. This technique is also helpful for organizing hyperspectral remote sensing-based data. Advantages (a) This method improves the performance based on "spectralspatial" classification between the two DN of the image ( Sharma and Verma, 2022 ). (b) It is an easy and fast approach to measuring spectral similarity between DN of the images ( Ye et al., 2019 ). (c) This approach uses only the direction of the image spectra, and their length is ignored. Thus this method is insensitive to unknown gains, and the illumination effects are treated equally .

Shortcomings
SID technique is widely used to classify hyperspectral images, one of the problems noticed while performing image classification is light intensity. It means as the intensity of the light is varied, output classification results get affected. Thus while performing real-time image classification through this algorithm, one should take care of this illumination issue ( Wang et al., 2018 ).
Several unsupervised image classification algorithms like kmean ( Haroun et al., 2021 ), Isodata ( Verma et al., 2017 ), etc. , can be used for classifying the image data according to one need and application. Besides these image classification algorithms, several other image classification algorithms employing supervised image classification includes Binary encoding classification ( BEC ) (Khan et al., 2018), Neural network-based classification ( NNBC ) ( Amjad and Geiger, 2020 ), and Support vector machine classification ( SVMC ) ( Maulik and Chakraborty, 2017 ). These algorithms are used in several "image processing" and "remote sensing applications." These algorithms also provide superior accuracy with susceptible image classification parameters.
The pre-Covidand the post-Covid lockdown satellite data of the investigating site are analyzed through supervised classificationbased techniques. The classification of these sites is performed based on several parameters like data training, class, Confusion matrix ( Cm ) or Error matrix ( Em ), User accuracy ( Ua ), Producer accuracy ( Pa ), Omission error ( Oe ), Commission error ( Ce ), Region of interest ( ROI ), Kappa coefficient ( Kp ), Overall accuracy ( Oa ), Accuracy assessment ( Aa ), etc. ( Shakya et al., 2018 ).
The training data set's purpose is to familiarize the classification algorithm with different scenarios. This training data may be DN of the image, latitude, and longitude coordinate to prepare the data. A class is expressed as the object formed in the embodiment based on which classification is performed in an image. These classes may be specific locations in the image. Fig. 3 presents an example of a supervised classified image and computation of various image classification features.
In the test image presented in Fig. 3 a, five different classes are created, i.e., grass, water, house, soil, and road. Thus an image can be classified into several categories, as presented in Fig. 3 b. Therefore, several classes are created based on the required number of applications. Fig. 3 c illustrates an arbitrary confusion matrix through which pixels are arranged in five categories, i.e. , grass, water, house, soil, and road.
Commission error ( Ce ) is expressed as the ratio of incorrectly classified pixels of one class to the total number of classified pixels in that particular class. The Ce for the arbitrarily created confusion matrix represented in Fig. 3 c for the five different classes is expressed by Eqs. ( 15 − 19 ) ( Chen et al., 2018 ).
Omission error ( Oe ) is expressed as the ratio of DN, which is accidentally omitted from the classified class to the total numbers of DN of the individual category. The Oe for the arbitrarily created confusion matrix represented in Fig. 3 c for the five different classes is expressed by Eqs. ( 20 − 24 ) ( Chen et al., 2018 ).
Oe ( W ater ) = ( 10 + 13 User Accuracy ( U a ) is a classification accuracy from the point of "map designer," not from the end of the "map user." This assessment provides details of the data present in a classified class. It is expressed with the assistance of Eq. (25) ( Chen et al., 2018 ). The Ue for the arbitrarily created confusion matrix represented in Fig. 3 c for the five different classes is expressed by Eqs. ( 26 − 30 ).

User accuracy (
Producer Accuracy ( Pa ) is obtained by dividing the accurately classified DN (major diagonal) of each class by the reference DN (column total). This presents how correctly the reference DN are classified. It is expressed by Eq. (31) ( Chen et al., 2018 ). The Pa for the arbitrarily created confusion matrix represented in Fig. 3 c for the five different classes are defined by Eqs. ( 32 − 36 ) ( Chen et al., 2018 ).
Overall accuracy ( Oa ) is the ratio of correctly classified DN to the total numbers of DN present in an image. The expression of Oa is represented by Eq. (37) ( Chen et al., 2018 ). The Oa of arbitrarily created confusion matrix represented in Fig. 3 is a parameter that indicates how efficiently the classification of the "study area" is performed. Suppose the completed sort is done close to reality. In that case reality, K p will obtain the value close to "one," if K p value is close to "zero," then it is assumed that the classification is not performed accurately, finally a negative K p value present that image classification is worst and can be compared with random classification because the range of the K p is [ −1 , 1 ] . K p is expressed by Eq. (39) ( Chen et al., 2018 ).
where probablity (e ) is known as chance arrangement probability and probablity (O ) is known as the observed agreement among various class raters.
The pre-Covid and post-Covid lockdown satellite data is obtained from "Planet" (Inc., The leading provider of global daily Earth data, 2010). The information is in the RGB format having three separate layers of the "RE D ," "GRE E N," and "BLUE " band. These bands are fused using the layer stacking approach available in ENV I 5 . 2 ( https://www.l3harrisgeospatial.com/ Software-Technology/ENVI ). Image classification and feature extraction are also performed using ENV I 5 . 2 . pre-Covid is considered a typical situation where everyone is busy in their day-today activities. People are going to offices, children are going to school, and newspaper vendors are distributing newspapers, i.e., everything is normal, just like a typical day.
On the contrary, in the post-Covid lockdown situation, change in the geographical surrounding can be observed. No activity is seen from the people, no vehicles are found on the road, and no work related to the infrastructure development, closed malls, shops, schools, etc., are located in the investigation area. Thus a lockdown can be considered the complete shutdown of the regular daily routine.
The investigation performed in this research is spread across four different continents, i.e., North America, South America, Africa, and Asia. Fig. 4 a presents the global location of the study areas, which shows a uniform selection of the individual case study across the globe. The Covid 19 lockdowns have adversely affected the World. But some advantages of the lockdown also exist, like improvement in the air quality index all across the World. Due to the lockdown, the movement of vehicles across the World stopped suddenly, reducing the pollution level; thus, air quality improved. The water quality also gets improved from the Covid 19 lockdown due to minimum industrial activities. Therefore, besides the maximum dark side, lockdown also has some positive sides. Fig. 4 a-j presents pre and post-optical imagery of Covid 19 lockdown situation, representing the change developed in land use-landcover The study areas represent the LU/LC geography, latitude and longitude locations, coordinates of the study area, a 3D model of the study area, and finally, the satellite data. The pre Cov id and post Cov id lockdown images of the study area present a variety of geographical variations. In case study 1, Tyson food factory in Washington DC, USA, it can be observed that in the pre-Covid image, food processing activities are performed in full swing, and the rooftops appear dark in color due to kitchen smoke and heat. Still, in the post-Covid lockdown image, the rooftops appear light due to no kitchen activity. In case 2 pre-Covid image represents a typical football stadium, whereas when we observe the post-Covid lockdown image new temporary setups can be identified which are created for the sake of keeping the infected Covid 19 patients. In case 3, an urban area of Cairo is presented, which shows heavy traffic in the pre-Covid image, but no activity in the metropolitan area is reported during post-Covid lockdown. The case 4 border area of Iran/Afghanistan is explored where it can be observed  that the pre-Covid image has the normal flow of traffic across the border. Still, in the post-Covid lockdown image, a large queue of vehicles is presented across both ends of the border. Finally, in case 5 urban city of China is explored for the lockdown situation where in the pre-Covid image, a regular daily routine of the city is presented, whereas in the post-Covid lockdown image roads have no traffic, and the town seems to be empty. Thus in this research work, different types of cases related to Covid lockdown are investigated to identify the optimum supervised classification algorithm. Planet provides daily data sets acquired through Skysat and PlanetScope constellations. It is an automated cloud-based satellite imagery platform. The PlanetScope provides data in RGB , multispectral and Near-infrared ( NIR ) band format. This data has a spatial resolution of 3 . 7 m . The constellation which captures this data is Planet Dove. At the same time, Skysat operates in RGB , NIR, and PAN bands. It provides satellite data in 0 . 5 m spatial resolution. It provides data in three different formats, i.e., planet base maps, planet monitoring, and planet archives ( https://www.planet.com/ products/basemap/ ).  Table 1 ( https://www.planet.com/products/ planet-imagery/ ).
The flowchart of the proposed methodology for comparing the pre-Covid and post-Covid lockdown satellite data for the case studies is presented in the Fig. 5 .

Case study 1, Washington DC, North America, USA
The geographical location of the USA is presented in Fig. 6   The pre-Covid and post-Covid lockdown multispectral images are the study area shown in Fig. 6 d-e, respectively, obtained from Planet.
The pre-Covid image was obtained on 25 June 2015, the post-Covid lockdown image was acquired on 30 April 2020. These satellite images belong to the Tyson meatpacking plant near Pasco, Walla Walla County ( https://www.planet.com/gallery/#!/post/ meat-processing-plant-shut-down ). The pre-Covid image represents that the meat processing operations are in full swing due to which rooftops appear dark in color due to heat transfer due to the announcement of sudden lockdown. In the post-Covid lockdown image, it can be observed that meat processing activities are performed, and the meat plant remains shut down ( https://www. planet.com/gallery/#!/post/meat-processing-plant-shut-down ).
The visual representation of the changes developed in the pre-Covid and post-Covid lockdown image dataset of the investigating site is presented in Fig. 7 a-l. Here classification of the images is performed based on six different classes, i.e., agriculture, meat unit, rooftop, roads, cars, and ground. The total size of the image pixels while creating a particular class in both images is approximately 1500 pixels. Thus, using the methodology discussed in Section 2 , images are classified based on different supervised classification techniques.
The assessment of the image classification is performed based on Ua, Pa, Oe, Ce , and Kp . These classification features are tabulated in Table 2 . Here, the individual class has produced a different visual result reflected in the classification parameters.
The detailed classification result showed that the most efficient classification scheme for analyzing pre-Covid image of this case study is MaDC which produces the highest value of the overall Oa . Similarly, the Kp obtained for this case is highest for MaDC . It is important to note that the training and testing DN numbers are the same for all classification schemes. Concerning pre-Covid image, minimum classification accuracy is obtained through PPC algorithm. The visual appearance of the classified images through PPC classification technique does not provide any helpful information. The maximum classification accuracy is obtained using the MaDC classification scheme. Corresponding to the post-Covid lockdown classified image MDC classified image has produced maximum classification accuracy with the highest Kp , and PPC algorithm has presented the minimum classification accuracy and minimum Kp .

Case study 2, Sao Paulo, Brazil
The geographical location of Brazil is presented in Fig. 8 a, obtained using Google Earth ( Google, 2022 ). The latitude and longitude of Sao Paulo are 23 . 53 • S and 46 . 62 • W and presented in the Fig. 8 b. The 3D model of the investigating site is illustrated in Fig. 8 c. The pre-Covid and post-Covid lockdown images of the study area are shown in Fig. 8 d and e, respectively.
The post-Covid lockdown image was obtained on 22 March 2020, and the pre-Covid image was acquired before 22 March 2020 as no exact date is provided by the Planet (Inc., Planet Labs, 2020). Because of increased Covid 19 cases, Brazilian football clubs joined hands with their government to fight against Covid 19 and to help the people of their country. "Maracana" and "Pacaembu," two famous soccer grounds situated in Sao Paulo, were converted into "emergency headquarters" to receive Covid 19 patients and provide health services to them (Inc., Planet Labs, 2020). This was a new starting in which football fields were converted into emergency health care centers. Soon after this initiative, news from all across the globe began to come, and sports stadiums were announced to be used to keep infected people separate from noninfected persons. The image classification results of the pre-Covid and post-Covid lockdown images of the "Marcana" football stadium are presented in Fig. 9 a-l. Here in the pre-image, no structure is visible, whereas, in the post-image, a new block is visible, which is the newly created emergency headquarters for treating people. The top cover over the stadium is a temporary arrangement to keep the people safe from atmospheric disturbance. Here, six different classes, i.e., houses, apartments, trees, roads, stadiums, and pools, are designed to classify the satellite data. A particular category of the class is created by taking approximately 1500 pixels. Thus, pre-Covid and post-Covid lockdown images are classified into six types using different classification schemes. Table 3 presents the features obtained after performing image classification of the pre-Covid and post-Covid lockdown images.
Classification is performed based on common feature values followed by an overall Oa assessment and K p evaluation.
This study results that the maximum classification accuracy is obtained for the MLC scheme, whereas the minimum classification accuracy is obtained for the P P C scheme corresponding to the pre-Covid study area. For the post-Covid lockdown study area, MLC have produced maximum classification accuracy, and P P C have obtained minimum classification accuracy. The K p is likewise altered by the change in the classification scheme.

Case study 3, Cairo, Egypt
The geographical location of Egypt is presented in Fig. 10 a, obtained through Google Earth ( Google, 2022 ). The latitude and longitude coordinates of the study area Cairo Egypt are 30 . 03 • N and 31 . 23 • E as illustrated in Fig. 10 b. The 3D model of the investigating site is presented in Fig. 10 c. The pre-Covid and post-Covid lockdown images of the study area are illustrated in Fig. 10 d and e, respectively. The pre-Covid image was obtained before 12 February 2020, and the post-Covid lockdown image was acquired on 12 February 2020 (Inc., Planet Labs, 2020). These satellite images represent the change developed in the local street of Cairo due to the lockdown, which the Egyptian government imposed as a preventive measure to protect its citizens from getting infected by the deadly virus (Inc., Planet Labs 2020). The pre-image consists of a heavy rush of vehicles on the streets, but the post image is free from heavy traffic. This study area is classified into six classes: buildings, roads, cars, trees, water, and pool. All the classes in both images are created by approximately taking 1500 pixels.
The image classification is performed based on six classification schemes and standard features. The visual representation of the changes developed in the pre-Covid and post-Covid images of Cairo's study area is presented in Fig. 11 a-l. The pre-Covid image represents everyday daily life in Cairo, whereas, in the post-Covid lockdown image, no traffic is seen on the roads.    The classification features of the pre-Covid and post-Covid lockdown images of Cairo, Egypt, are tabulated in Table 4 . This classification is performed based on standard object-based image classification features.
This case study compares the pre-and post-images based on similar conditions. Here it can be observed that the pre-Covid study area MLC has obtained maximum classification accuracy, whereas the minimum classification accuracy is obtained for the PPC scheme. Similarly, for the post-Covid lockdown image, MLC have produced maximum classification accuracy, and PPC have achieved minimum classification accuracy. The Kp value is also changed concerning different classification schemes.

Case study 4, Iran Afghanistan border
In this case study, the geographical location of the Iran/Afghanistan border is presented in Fig. 12 a, obtained from Google Earth ( Google, 2022 ). The latitude and longitude coordinates of the Dogharoon border terminal are 35 . 54 • N and 69 . 16 • S, illustrated in Fig. 12 b. The 3D model of the investigating site is presented in Fig. 12 c. The pre-Covid image of the investigating site is represented in Fig. 12 d. The post-Covid lockdown image of the investigating site is presented in Fig. 12 e.
The pre-Covid image was obtained on 21 February 2020, and the post-Covid lockdown image was snapped on 26 April 2020 (Inc., Planet Labs, 2020). Due to the Cov id lockdown situation, trucks packed with export items are stopped on either side of the roads. Due to international lockdown conditions, commercial vehicles can be seen standing on the roadside and not crossing the border. According to the news sources of BBC, officials of both countries were strictly following Covid 19 protocols and checking the persons across both international borders for Covid 19 symptoms. The pre and post-lockdown Covid 19 images are classified into six different objects based on six other supervised image classification approaches. The comparison is performed based on the conventional classification parameters. These classes include rooftop (RT) white, RT blue, ground, trees, cars, and road. While performing image classification between the two images, a class is created using approximately 1200 pixels. Here the pixels used for defining a class are less than those used in the above cases because these images have fewer numbers the objects that can be classified. A vast geographical area is covered by ground and sand itself.
The visual representation of the changes developed in the pre-Covid and post-Covid lockdown images of the Iran, Afghanistan border is presented in Fig. 13 a-l.
The classification features of the pre-Covid and post-Covid images of the Iran Afghanistan border are tabulated in Table 5 . The features are computed based on standard image classification features, overall accuracy, and K p.
The image classification of the study area suggests that for the pre-Covid image MLC classification scheme has obtained maximum classification accuracy and maximum K p. The minimum classification accuracy is accepted for the SIDC classification algorithm. Similarly, for the post-Covid image, MLC classification scheme again has obtained maximum classification accuracy and K p. Here SIDC classification algorithm has received minimum classification results.

Case study 5, Beijing China
The geographical location of China is presented in Fig. 14 a, obtained from Google Earth ( Google, 2022 ). The latitude and longitude positions of the study area are illustrated in Fig. 14 b. The 3D model of the study area is presented in Fig. 14 c. The pre-Covid and post-Covid lockdown image of the study area is present in Fig. 14 d and e, respectively. The pre-and post-Covid lockdown images were     obtained on 23 February 2020 and 12 April 2020, respectively (Inc., Planet Labs, 2020). In the pre-Covid image, a typical day of the heavy traffic in Beijing is reflected where traffic density is high with people following their routine work schedules. In the post-Covid lockdown image, it can be observed that due to the sudden lockdown, the traffic density gets significantly reduced as several lockdown restrictions get enforced on the city. The image classification results of the pre-Covid and post-Covid lockdown situation are presented in Fig. 15 a-l. The image classification is performed based on six different classes. These classes are RT, trees, rivers, cars, ground, and roads. Here the classification is performed in an urban area. Thus, approximately 20 0 0 pixels are used for more precise visualization of classes to define an image class.
The classification feature of the pre-Covid and post-Covid lockdown images of Beijing, China, are tabulated in Table 6 based on standard classification techniques and parameters. The overall classification accuracy and K p values are the determining parameters that suggest how good classification of the study area is performed.
The image classification of the study area suggests that for the pre-Covid image MLC classification scheme has obtained maximum classification accuracy and maximum K p. The minimum classification accuracy is obtained for the P P C method. Similarly, for the post-Covid lockdown image also MLC scheme has obtained maximum classification accuracy and K p. The minimum classification accuracy is obtained for the P P C classification technique. Thus, the change developed can be visually identified through the image classification approach. One of the advantages of the image classification approach is that even a tiny change in a particular object can be identified and analyzed through these approaches.

Discussions
This research presents the efficiency of the various supervised image classification-based approaches' in classifying remote sensing data based on distinct classification classes and features. Image classification has a prominent role in remote sensing applications. Image classification techniques are used to model several realistic situations like flood mapping ( Manavalan, 2017 ), urban expansion assessment ( Sahana et al., 2018 ), landslide mapping ( Mondini, et al., 2021 ), agriculture land modeling ( Zadbagher et al., 2018 ), volcanic eruption modeling ( Kereszturi et al., 2018 ), etc . It can be observed from these cases that these situations are enormously complicated to model when it comes to reaching exactly the ground location and collecting the data. Thus in these complicated scenarios, image classification techniques are used. In image classification techniques, data collected with the assistance of satellites, drones, and UAVs are analyzed, presenting details of the study area ( Bhatnagar et al., 2021 ). While mapping these investigation areas, image classification approaches play an extremely prominent role. The accuracy level of the performed image classification is determined by parameters like Kp, Oa, etc . (Foody, 2020). Thus it become quite important to identify an optimum image classification algorithm having the maximum efficiency in modeling the satellite data. Therefore this research identifies and presents the implementation of the most suitable supervised image classification algorithm for satellite data.
Six different case studies are investigated in this work based on pre-Covid and post-Covid lockdown satellite data obtained from Planet using supervised image classification techniques. Classification techniques can present the different visual appearances of a study area. The image classification depends on the number of objects used to define a study area. Six supervised classification techniques used in this work are P P C , MDC , MaDC , MLC , SAMC, and SIDC, respectively. It is observed from this work that P P C has produced the most abrupt results for image classification. This technique has not provided any useful information from an image classification point of view. The pre-and post-Covid lockdown classified images become utterly different in visual appearances through these techniques. Therefore, the Oa and K p values re-  ported for this classification scheme are minimum, whereas MDC, MaDC, and MLC are three classification techniques that have documented reasonable high Oa and K p. It can be observed concerning case 1 from Fig. 16 a that MaDC have obtained maximum classification accuracy from the pre-Covid image, whereas through MDC algorithm post-Covid lockdown image has obtained maximum classification accuracy. The K p coefficient has received the highest value for the MaDC algorithm for pre-Covid image, whereas it has acquired the lowest value through P P C algorithm represented by Fig. 16 b. Similarly, concerning case 2, through MLC technique, the highest classification accuracy and K p obtained for both pre-and post-Covid lockdown images are represented by Fig. 16 c, d, respectively. MLC approach has again produced the highest classification accuracy and K p for both pre-and post-Covid lockdown images illustrated by Fig. 16 e-f respectively concerning case 3. MLC technique has again produced the highest classification accuracy, and K p for both preand post-Covid lockdown images concerning case 4 represented by Fig. 16 g, h respectively. Finally, the classification results of case 5 also report that MLC technique is among one of the best classification techniques as it produced maximum classification accuracy and K p.
Thus it can be concluded that MaDC and MLC image classification techniques have produced maximum Oa . The K p value seems to increase in the same order as Oa. A relationship between Oa and K p to obtain maximum R 2 provides information about how well these image classification accuracies are associated with the classification techniques. The classified images arranged in descending order of classification accuracy for the pre Cov id image follow the pattern expressed by Eqs. (40 -44 ). The curve fitting between the Oa and K p for the pre-Covid images following the order expressed by Eqs. (40 -44 ) is presented in Fig. 17 a-e. Here it has been observed that the value of R 2 is approximately about 0 . 99 in most of the classified images. This suggests a great fitting between Oa and K p.
The curve fitting between the Oa and K p for the post Cov id lockdown images following the order expressed by Eqs. (45 -49 ) is presented in Fig. 18 a-e. Here it has been observed that the value of R 2 is approximately close to unity in most of the classified images. This suggests a great fitting between Oa and K p.  Thus through this research work, some novel findings are obtained concerning image classification, expressed as follows.
• Among the various supervised image classification algorithms discussed in this study M LC, M aDC and M DC are the most prominent image classification algorithms for classifying satellite data producing the highest classification accuracy and K p. • Satellite image classification accuracies are directly related to the K p. • It has been observed from this work that good classification also depends on the pixel selections for creating a class. Suppose fewer numbers of image pixels are used to create a classification class. In that case, the quality of the classified image will be compromised, and possibilities of omission error and commission error increase, affecting all the classification parameters negatively. • Pre-Covid and Post Covid 19 lockdown is a new scenario investigated in this research work for change detection and identification of the optimum supervised classification algorithm.
Thus, image classification-based changed detection techniques have tremendous potential to model different situations. The visual representation of the developed changes and presentation of an image's statistical classification features can be used for several applications. Further, image classification techniques also provide information about the accuracy assessment, indicating how well the image classification is achieved.

Conclusions
This research compares supervised object-based image classification techniques to investigate Pre-Covid and post-Covid lockdown situations of the Washington DC USA, Sao Paulo, Brazil, Cairo, Egypt, Afghanistan/Iran border, and Beijing, China. Six different image classification techniques known as P P C , MDC , MaDC , M LC, SAM C and SIDC are used to investigate the satellite data obtained from Planet. Image classification through supervised image classification techniques requires the use of training data. This training data is used to train the algorithm by extracting data from the existing region of interest, including representative pixels of the desired class. Later post classification accuracy assessment is performed based on which overall classification accuracy, K p, and other related classification parameters are obtained. Thus based on classification results and statistical features achieved on ENV I 5 . 2 , it has been observed that P P C classification has produced the least significant results. In contrast, the most reliable results, along with the highest classification accuracies and K p are obtained by MLC, M DC, and M aDC techniques. Specifically, M LC has produced the highest classification accuracies in most numbers of cases. The reason for MLC producing the highest classification accuracies is that this technique follows a pretty sophisticated procedure for performing image classification. Here the statistics for each class in each band are typically distributed, thus calculating the probability of a pixel belonging to a specific category. Therefore, a detailed procedure is adopted for classifying the individual pixel into a particular class, producing the highest classification accuracies. It is also observed that the K p is directly related to the overall classification accuracy. Finally, image classification-based algorithms can successfully identify and quantify the amount of change developed in an investigating class or area. Thus, an image classification algorithm can be directly used as per the investigation's requirement to model any geological situation.

Declaration of Competing Interests
All the authors declare no conflict of interest.