Granite Exposure Mapping Through Sentinel‐2 Visible and Short Wave Infrared Bands

Nonmetallic minerals like granite and limestone have calcite and biotitic ingredients as their major part which exhibit wonderful absorption features in the visible and short wave range of the electromagnetic spectrum. This research puts emphasis on delineating granite and limestone deposits of the Mardan district through the latest multispectral Landsat‐9 and Sentinel‐2 sensors of which the latter provided 94% mapping accuracy in delineating granites (biotitic bearing minerals) and limestone (calcite‐bearing minerals). The Image processing techniques of minimum noise fraction, which is double cascaded principal components analysis and pixel purity index algorithms proved helpful to bring significant improvements in classification results and in the reduction of noise and data size. The outcomes of the research study show that supervised machine learning algorithms are impactful to map such minerals provided that the data must be well organized and limited in size. The results achieved were verified through validation steps like, (a) Independent reference data of high‐resolution Google Earth maps and (b) Ground survey reports. Arc GIS 10.2 and Envi 5.3 software suite were used to create (a) ground truth points at random for accuracy assessment (b) portraying study area maps (c) Image Processing and Preprocessing tools and (d) implementation of machine learning algorithms. Access to the data and software suite is being provided for open research work.


Introduction
Pakistan holds territory that has been blessed by nature with monumental deposits of minerals and precious stones.The conventional minerals mapping and extraction technique has many drawbacks that need to be resolved by adopting mineralogy and geology to new modern lines.Remote Sensing (RS) has come up as a technological alternative for modern mineral mapping over the past many decades.Present research work focuses on delineating and mapping minerals' deposits with RS techniques and assessing multispectral satellite sensors for their mineral mapping capabilities.RS is a systematic and sophisticated technique for figuring out minerals using lithological, structural, and alteration mapping (Joana et al., 2020).Surface reading and spectral signature analysis a key factors in identifying minerals on the Earth's surface (Beiranvand et al., 2015).The spectral signature of a mineral depicts its reflection and absorption properties against various electromagnetic (EM) signals (Chellamuthu et al., 2021).
The hydrothermal process is the major cause of minerals forming.Extremely hot boiling water and a mixture of gases is the main source of hydrothermal alterations.They manifest on the surface as feature indicators such as surface color, surface cracks, rock patterns, and many more.Remote sensing photographs and imageries highlight surface features and rock patterns to predict minerals on the Earth's surface.Knowing the fact that absorption and reflection features are specific to each mineral, it is possible to extract information from the spectra of natural rocks, detect the different minerals, and estimate their relative proportion.This is what remote sensing exploits in mapping minerals and rock types.total of 5070 discovered mineral types are internationally recognized and accepted by the International Mineralogical Association (IMA) (Uddin et al., 2020).
Way back in time, minerals were delineated and dug out using traditional methods which demanded enormous manpower as well as time (Van & Van, 2016).Remote sensing optical sensors are mainly used to discover minerals without having to make any physical contact.The spectral range of 0.4-2.5-μmprovides abundant information about many important earth-surface minerals (Sheikhrahimi et al., 2019).The range of 2.0-to 2.5-μm of Short Wave Infrared (SWIR) is of utmost importance and shows greater reflectance to those minerals that bear carbonates, hydroxyl particles, and sulfates as a common ingredient (Gupta & Venkatesan, 2020).Literature reveals many multispectral and hyperspectral satellite sensors have been used so far for mapping mineral deposits successfully.RS data has been used to map rocks, minerals, hydrothermal alteration, lithological units, and surface textures (Krutz et al., 2019;Loizzo et al., 2022;Pour et al., 2019;Samani et al., 2020;Shirmard et al., 2020).Mapping minerals using multispectral data is an easy job as compared to mapping with hyperspectral data sets which are most often comprised of hundreds of bands with relatively short interval bandwidths (5-10 nm) (Carvalho et al., 2019;Irons et al., 2012;Peyghambari & Zhang, 2021).Earlier Landsat (Barsi et al., 2019;Clevers et al., 2017;Masoumi F. et al., 2017) satellites used 4 bands of EM spectrum which, later on, kept increasing to 11 bands in the latest Landsat series.The latest of its launches is Landsat-9.It was turned on fully functional on 16 February 2022.Landsat carries two different instruments (a): The Operational Land Imager (OLI) and (b): The Thermal Infrared Sensor (TIRS).Each image product contains nine spectral bands with a spatial resolution of 30 m for bands 1-7 and 9, and 15 m for band 8, and two thermal bands can capture images with 100 m resolution.The other multispectral imageries are Aster and Sentinel 2 which are more suited for mapping non-metallic minerals.Landsat 9 and Sentinel 2 imageries are used to delineate mineral deposits in this research article.
Most often, when the mineral lineaments are minute and unclear, the space-born sensors stay too far away to record their distinguishing and clear images.In this sort of situation, close-range remote sensing provides a practical solution to acquire close images.Unmanned Aerial System (UAS) offers a better but more expensive alternative to space-borne remote sensing for studying geological lineaments and structures with controlled passive radiation sources.They are expensive as customized equipment like sensors, Light detection, and ranging (Lidar) (Kurza et al., 2012), and drones are required to record the data of some locations on the other hand, most of the space-born satellite data is freely available for research and study purposes.Rocks' characteristics and finer details can be observed closely if the data is acquired from lower altitudes and with much finer spatial resolution.Carbonates like calcites, dolomites, and limestone deposits have been mapped with greater accuracy and finer details which would otherwise be difficult if opted for space-borne and satellite remote sensing.This technology is a promising one and highly recommendable for studying rock structures and steep clip walls.A literature study, covering the latest remote sensing data (2021 and onward) to figure out recent minerals mapping classification results against various satellites, is accumulated in (Table 1) which is meant for the purpose of comparing various sensors' mapping capabilities.

Data Sets and Geology of Study Area
Mardan district of Khyber Pakhtunkhwa (KP) Pakistan gets a reputation in the country for its rich mineral deposits.It is located in the vicinity of the historical Peshawar city of KP province.Huge deposits of Granite, limestone, Soapstone, Iron ores, marble, and many other precious minerals have been discovered there in a survey campaign held by a government organization (KP Minerals' Department) and given on open lease to many firms for extraction and processing.There exist monumental deposits worth millions of dollars which can be viewed at coordinates 34°12′11″N & 72°20′07″E and its vicinity.A survey team, from Lab, National Center in Big Data and Cloud Computing (NCBC) UET, Peshawar, visited the site and gathered field samples and the coordinates of the study area through the Global Positioning System (GPS) on 05 June 2022.The next day, RS data of the study area was downloaded at the lab starting at 02:10:10 and ending at 02:51:42 local time.Courtesy to the United States Geological Survey (USGS) for providing valuable databases of remote sensing data.Cloud cover was kept under 10%.The location of the surveyed area is highlighted on a globe map and shown in Figure 1.Like ASTER sensors Both the granule record data with bands that cover SWIR, VNIR, and Thermal Infrared (TIR) bands of the EM spectrum (Eslami et al., 2015;Masoumi et al., 2017).RS data can be viewed in many false color combinations.The Sentinel 2 imagery can be viewed as natural when Band 4, Band 3, and Band 2 are selected as red, green, and blue respectively.This combination provides natural scene visualization for readers.Similarly, Band 7, Band 5, and Band 3 when set as red, green, and blue respectively are a useful combination for highlighting geological structure and minerals exposure in an acquired scene.Figure 1 is Sentinel 2 scene displayed as B7, B5, B3: R, G, B for mineral characterization exposure.

Methodology and Treatments
RS granule underwent many treatments before it could be classified for various class members.It consists of many stages which include image pre-processing, Image processing, Image Classification, Image Cleaning, and data dimensionality reduction.The order they follow is depicted in the methodology diagram as shown in Figure 2 and discussed thoroughly in upcoming sections.

Remote Sensing Granule
Multispectral satellite sensors are the most appropriate sensors used to record the reflectance of surface minerals through visible and short-wave infrared bands.Non-metallic minerals, like granite and limestone, possess biotite and calcite (CaCo 3 ) ingredients which offer greater reflection and absorption characteristics in visible and shortwave infrared bands, thus making them the most appropriate choice for multispectral sensors.The reflectance signature of the targeted minerals in this range of electromagnetic (EM) spectrum clearly distinguishes them on the spectral curve, when compared with curves provided by the USGS spectral library, from other metallic minerals like iron ore, copper, etc.The most relevant-to-minerals bands (Highlighted in Tables 2 and 3) from both granules were picked from their band's list.We chose bands 2, 3, 4, 5, 6, 7, 8, 8a, 11, and 12 for Sentinel 2 and bands 2, 3, 4, 5, 6, 7, 8, and 9 for Landsat 9 sensors.The chosen bands were stacked using the layer stacking toolbox in Envi 5.3 and resampled to a spatial resolution of 10 m for Sentinel 2, and 30 m for Landsat 9 to improve visual quality (Shirmard et al., 2020).

Pre-Processing
The image pre-processing steps are applied in order to achieve a couple of goals.First, to wipe off atmospheric noise, and second, to translate radiance data into reflectance images.Both these steps are considered prerequisites to processing RS data for practical applications.RS data is recorded by the sensors in the form of an array of values called Digital Numbers (DN) where each pixel inside gets its specific DN value.Every single pixel inside the image is more likely to hold information about multiple objects on the ground surface.Thus, the DN value of every single pixel is the average value which includes fractions of all the objects included inside the pixel.However, every single DN value needs to be converted into an Absolute Reflectance (AR) value, as the reflectance imagery is the fundamental requirement for spectral mapping and identification purposes.This task is achieved following two consecutive steps.First, DN values of the sensor are translated into radiance values of the earth's surface by using Radiometric Calibration (RC) operation.They are referred to as radiometrically calibrated radiance images.These radiance images are translated into Absolute Reflectance (AR) images by the formula given in Equation 1.The reflectance images provide spectral information of the earth's surface which can be compared to the spectra provided by various standards like USGS and Jet Propulsion Laboratory (JPL) libraries.Environmental Visualization (Envi) 5.3 toolbox provides options like Radiometric Calibration, FLAASH, and QUick algorithms custom setting for achieving the above tasks (Samani et al., 2020).The resultant reflectance images are termed as atmospherically adjusted and geometrically corrected ones which can be further processed for various operations like identification, classification, and many others.The process stated above is termed atmospheric and geometric correction.
(1) These operations are mostly carried out on level 1 (L1) remote sensing data which needs these sorts of corrections prior to undergoing any useful processing.Most of the time the hyperspectral data is provided in L1 format which needs this sort of corrections manually.The multispectral granules (Sentinel 2 and Landsat 9) downloaded for the intended research work were acquired from the USGS Earth Explorer website (https://earthexplorer.usgs.gov/) in level-2 (2A) format.The granules were corrected geometrically and adjusted atmospherically by the launch source and did not require any such operation.However, the authors consider it important to provide a fundamental understanding of related terms and useful tools.

Image Processing
The image processing phase is responsible for enhancing image quality and removal of unwanted data.Image enhancement involves the removal of noise and the sharpening of minute details.The minute details refer to useful edges and lineaments inside image data.The reflectance curve also called the spectral signature of the mineral, shows the reflectance response of the mineral against various wavelengths of the electromagnetic spectrum.As depicted in Figure 3a Calcite (Limestone Ingredient) offers distinguishing reflectance response in the range of 1,000-2,000 mm with dipping points of 2,000 and 1,500 mm and high smooth reflectance between 1,000 and 1,500 mm. Figure 3b represents the spectral curve of Fe3 (Iron) exhibiting a different reflectance response against the same range of wavelengths.Fe3 shows a greater response to higher (greater than 2,000 mm) wavelengths.This helps recognize individual minerals and differentiate them from each other.Spectral curves of materials are recorded in close-range spectroscopic studies.
Techniques like Principle Components Analysis (PCA) and Minimum Noise Fraction MNF are used to reduce the dimensions of RS data.MNF has priority over PCA for its multi-folded operations on RS data.It wipes off the noise and reduces the spectral dimensions of RS data simultaneously.It reduces spectral dimensions by removing all those bands that are polluted.The MNF option was chosen from the Envi 5.3 toolbox which separately displays the image of each band with detailed graphs of noisy/redundant bands and useful bands.The bands with maximum noise/redundant data can be excluded from the final selection.In this way, it also reduces data size by removing unwanted and noisy bands.It is worth mentioning here that most often it is the hyperspectral data that contain noisy and unwanted bands as they record data in hundreds of consecutive bands in which many of them are absorbed by atmospheric pollution, and water vapors and contain bad bands.Therefore, it is necessary to wipe off such bands quite often.The intended research work has utilized multispectral data which contained a limited number of bands (Tables 2 and 3).Though we exercised these operations on our data for experimental purposes, it had only a negligible effect on the data as it was small and clean from bad bands.

Polygon Drawing and Pixel Refinement for Classifiers Training
The reliable classification of data by ML algorithms totally depends on the careful selection of training data.Training data must be clean, unmixed, and big enough to provide a full-fledged training session for ML algorithms.Besides this, the Training data set and test data set must be independent of each other to avoid bias in the classification process.
The above has been ensured by the authors by adopting a multistep pixel selection procedure.First, mineral deposits of granite and limestone were verified from the authenticated and official government website which displays mining lease and open pits polygons for all minerals deposits of the intended study area.We drew training polygons of various class members of the study area and imported their shape files in Envi 5.3.Clean ROIs of each class member were drawn manually using RS imagery with the help of imported shape files and then the pixel reflectance values were recorded and saved in CSV files to serve as training data.One ROI may contain single or multiple pixels depending on the size of the ROI drawn.For instance, Sentinel 2 data with 10 m spatial resolution would mean, a pixel size of 10 × 10 = 100 m 2 .(The area covered by Sentinel 2 sensors in one pixel while capturing the surface image is called pixel size).Enough amount of training polygons and pixels (Table 4) were chosen for the training session.
In order to double confirm the purity of the chosen pixels for training the algorithms, we exercised another procedure on data called Pixel Purity Index (PPI) selection.PPI works on a whole bunch of data in thousands of iterations.In every iteration, it visits each pixel and records its spectral curve (also called spectral signature).Pixels with similar spectral curves are grouped in one cluster.In this way, many clusters of pixels having similar spectral characteristics are formed and grouped separately.This process continues until all the alike pixels are placed in clusters.Once the clusters were formed, we noted their coordinates and projected the coordinates on the remote sensing imagery of our intended study area.It was ensured that separate clusters were formed for separate class members.The spectral curve of certain cluster pixels can also be matched with one provided by standard libraries for further verification of data.Envi 5.3 provides a custom setting for implementing the PPI algorithm where we set 10,000 iterations on the data to get the pure pixel clusters.We tested with multiple options and achieved the optimum results with 2.5 thresholds and 10,000 iterations.Thus, we chose only those pixels for training the ML algorithms which were ensured for each class member by (a) mining lease polygons, (b) survey team, and (c) extracting by PPI procedure.The pixel reflectance values were recorded and saved in CSV files to serve as training data.

The Classification and Confusion Matrix
The reflectance of the chosen pure pixels were recorded and saved in CSV file which were used as training data for various ML algorithms.In order to calculate the classification accuracy and confusion matrix, the same location high-resolution images were downloaded from Google Earth Engine (GEE) to act as an independent Reference source here.The images classified by ML algorithms were considered "User Data" while the images downloaded from GEE images were considered "Producer Data."ArcGIS 10.2 was used to draw points at random locations on a Reference image and was saved as a shape file.The Keyhole Markup Language (KML) file was generated for the recording user and producer values table.It is primarily used to display geographic information in applications like Google Earth, Google Maps, and other mapping software.This file gives complete information about each point, line, polygon, image, and text label on the maps to ensure whether it has been correctly classified or misclassified.Table 5 was created for user data and producer data as standard operators to find the classification accuracy.The overall accuracy was calculated for many ML classifiers and the most optimal performers are enlisted in Table 5.

Supervised Classification (SC)
A bunch of supervised classification algorithms was deployed and tested to achieve results.Pixel reflectance values were recorded and saved in CSV files for training purposes.Supervised classifiers make the placement of pixels in certain classes based on the training they obtain during the training session.The training polygons were kept enough so that classifiers could get proper data for the training session.The supervised classifiers deployed in the present mapping work are described as under.

Spectral Angle Mapper (SAM)
SAM is the most widely used supervised spectral classification method that uses an n-dimensional angle for calculating the spectral similarities between image spectra and reference reflectance spectra.SAM is a technique for classifying pixels inside a multispectral image based on their spectral resemblance to a reference spectrum.It computes the spectral angle between each pixel's spectral signature and a reference spectral signature.The spectral angle is a trigonometric measure of the angular difference between the two spectra.It measures the similarity of the spectra.A smaller angle would mean a closer resemblance to the reference spectrum and hence a closer match to the end member.The spectral angle is calculated by the formula shown in Equation 2.
(2) SAM algorithm has been extensively used for mapping minerals' deposits especially hydroxyl, kaolinite, and calcite-bearing minerals in (Z.Adiri et al., 2016).The algorithm was trained on the spectral data of the said mineralization samples and deployed for mapping them with satisfactory results.Similarly, chromite-bearing mineralization was delineated and mapped through the SAM algorithm using VNIR and SWIR bands of ASTER (Rani et al., 2016).Bands 4 to 9 were used to achieve the task as these bands, especially band 7 exhibit mineral mapping capabilities with fine details.

Support Vector Machine (SVM)
SVM is a nonparametric, supervised classification algorithm that is based on statistical learning approaches.This classifier is suitable for distinguishing class objects in a complicated class distribution data environment.Data is served in an n-dimensional hyperplane and gets separated into two classes of support vectors and misclassified instances based on the evidence data of a predictor variable (Adiri et al., 2016).The greater distance between the neighboring points on the hyperplane would lead to good separation and fine classification.SVM has been used for delineating sandstone, rock types, and many other mineral ingredients in the recent past.SVM was instrumental in the mapping of Regolith-geology mapping and Au deposits mapping.It outwits SAM when it comes to mapping lithological units through hyperspectral sensors.

Unsupervised Classification (UC)
The two most used unsupervised classification algorithms for minerals mapping are k-means and ISODATA classifiers.This category of classification is useful for the verification of various classes that are already classified in the supervised classification stage.In UC clusters are made from data elements/pixel elements with similar features and characteristics.Multiple iterations of scanning are performed over pixels and finally, they are placed in the category where a closer or exact resemblance is found.There are multiple choices to find out the close resemblance.Finding the nearest distance of the pixel element to the center of the cluster through the Euclidean formula is utilized in the UC algorithms which have been used in present research work.The post-cleanup process was performed over the final classified maps before they were presented in the script.

K-Means Clustering Algorithm
This is a classification algorithm that belongs to the pool of unsupervised categories and is best for making clusters out of data with similar features and properties.This algorithm is not trained to get a model but instead, every new instance of data is directly analyzed for its features and placed in a certain category according to these features.This is quite uncharacteristic of supervised classification and hence this is placed in the unsupervised classification group.This type of classification would be expected to require greater resources as the number of training samples increases.

Iterative Self Organizing Data Analysis Technique Algorithm (ISODATA) Classification
The Iterative Self Organizing Data Analysis Technique Algorithm (ISODATA) is an unsupervised classification algorithm which is a general form of k-means algorithm.A threshold value is set to some predefined value and classes are merged based on iterations.Each data member is assigned a class based on the measured distance between the centers of clusters.K-Means clustering would need a predefined value for the number of classes whereas the ISODATA algorithm allows different numbers of classes.This algorithm is mostly used when required training data is not available in hand.Unsupervised classification algorithms have mainly been used for minerals mapping and geological surveys (Loizzo et al., 2022;Shirmard et al., 2020).The distance between the center of the cluster and each pixel is measured by the Euclidean distance formula.The pixels with the shortest distance are assigned to the cluster near it.The distance is calculated by the formula shown in Equation 3.
Through an iterative procedure, ISODATA is meant to automatically calculate the number of clusters in the data.This versatility is especially useful when working with mineral mapping, as the number of mineral kinds or classes may not be known ahead of time.k-Means, on the other hand, needs the number of clusters to be determined in advance, which might be difficult when the actual number of mineral classes is unknown.ISODATA can deal with clusters of all forms and sizes, whereas k-Means implies clusters are spherical and of equal size.
Because mineral distributions in hyperspectral data can be complicated and irregular, this flexibility is critical for mapping minerals.ISODATA refines the cluster centroids repeatedly, modifying them to better represent the data distribution.When compared to the simpler k-means update approach, this refining process can result in enhanced mineral separation and a better fit to the data.

Random Forest (RF)
Random Forest is an advanced ensemble learning approach that mixes many decision trees to provide a more robust and accurate classifier.Each decision tree is trained on a distinct subset of the training data and predicts independently.The Random Forest's ultimate forecast is derived by aggregating the predictions of various trees, which is generally done by majority voting (classification) or averaging (regression).Random Forest's capacity to tolerate noisy input is one of its benefits (Zhang et al., 2022).The noise caused by individual data points is decreased by pooling predictions from many trees.When compared to a single decision tree, a Random Forest is less prone to overfitting to noisy data.Random Forest's "bagging" component entails training each tree on a random portion of the data (with replacement).This approach decreases variance and helps to avoid overfitting, which is very useful when dealing with noisy data.It can learn from data with complex relationships, which might be useful for categorizing data with low discrimination but that is still good enough to acquire useful features and contain relevant patterns.

Confusion Matrix
We split the data first in various ratios 60:40, 80:20, and 70:30 for Training and Test purposes respectively where the most optimum results were obtained with a 70:30 ratio.This can be illustrated with an example like, Out of 1,300 pixels of mineral class, as depicted in Table 4, we kept 910 pixels of minerals for training and 390 pixels for testing the models which is a 70:30 ratio of 1,300.We initiated the training phase with 100 purified pixels from each class and calculated their overall accuracy using Table 5.We kept increasing the number of pixels and noted the improvements in the final results.Validation was performed on a separate and random set of points generated from independent high-resolution images.The optimum results (Table 6) with the resultant classified map (Figure 4) were achieved when the number of pixels stated in -------------------------  A 93.5% overall accuracy was achieved for SAM followed by RF with 89 approx.ISO provided 94% form the pool of unsupervised classification with the rest (best performers) shown in Table 6.

Discussion
The application of the Landsat-9 and Sentinel-2 VNIR-SWIR bands for delineating Granite and limestone deposits of the Mardan region demonstrates that both these multispectral sensors can map nonmetallic minerals with Sentinel 2 standing as the outperformer.A series of image processing techniques, including false color composite, MNF, and PPI were utilized to process the data to bring significant changes in classification accuracy.To sum up the whole process on a comprehensive note, it is stated that Sentinel-2 data accompanied with image processing methods and reference spectra (e.g., USGS, JPL) or field measured) could be an effective technique for mapping biotitic and calcite in areas with less vegetation cover.Supervised ML algorithms can be vital to map and identify mineral deposits provided that the data is taken care of in precise measures.{{Due to widespread mineral deposits in the study area, mineral prospecting methodology is suggested for application adoption to similar geological settings in the belt}}.
The presented research findings can be extended to other areas where the prospective percentage is high.Baluchistan province, Kohat, Bannu, and Abbottabad regions are rich in mineral deposits that can be investigated and utilized for revenue generation.

Conclusion
Keeping cloud cover of less than 10%, RS data of the study area was downloaded for two different multispectral satellites.The satellites chosen for mapping the study area were Landsat-9 and Sentinel-2 as the former is the latest launch of the Landsat series with a better radiometric resolution of 14 bits and the latter has a better spatial

Figure 1 .
Figure 1.Shows granite and limestone mining lease (ML) polygons drawn in the study area.These polygons are drawn by government organizations during surveys to find mineral exposure in Khyber Pakhtunkhwa province.

Figure 2 .
Figure 2. Shows complete methodology diagram.It shows all the treatments and procedures applied to remote-sensing granules.

Figure 3 .
Figure 3. (a) Shows the spectral signature of Calcite while (b) shows the spectral curve of Fe3(Iron).The spectral curves library for various materials is provided by USGS Earth Explorer.

Figure 4
Figure 4 is the final classified map of the study area with legends specified.While training ML models, enough amount of training data (CSV files) was collected from each class of Polygons/patch (Table4).For example, we drew 73 (47 for training and 26 independent polygons for test purposes) polygons of mineral patches in the study area, generated their CSV files, and then split them in Training-Test ratios.A similar approach was adopted for all other classes as well.Once the models were assured to get trained properly, the classified maps of the study area were generated and validated later on with independent data points.The supervised classification proved much better this time and yielded results almost identical to unsupervised ones.The careful selection of supervised data and in optimum size for training sessions boosted the classification accuracy to satisfactory figures.

Figure 4 .
Figure 4. (c) Sentinel-2 classified map of the study area generated by ISO classifier with legends specified.(a, b) enlarged minerals portion.

Table 1
Shows Different Satellites Used for Minerals Mapping in the Recent Past

Table 4
Shows No. of Samples Taken for Training-Test Session

Table 5
Minerals Rocks Trees/Veg Urban Total (User) Shows Sentinel-2 Confusion Matrix of Test Set Data

Table 6
Shows Classification Results for Various Machine Learning Algorithms improvements were noticed with the increase in data.This clearly illustrates that ML algorithms work best when data size is limited and well prepared.