Study Area:

The study area falls in the territory of Mardan situated within Pakistan's Khyber Pakhtunkhwa province and is known for its rich mineral resources, including Limestone. Limestone is an important natural resource in the Mardan area and plays a significant role in the local economy. The area selected comes under the umbrella of 34°23'02.1"N, 72°07'27.1"E, as shown in Figure 1. A complete detail of the methodology is shown in Figure 2.

Figure 1: Generalized Location of Study Area on Map

Data Acquisition:

Remote sensing multispectral data acquisition is the process of collecting data from remote sensing platforms, such as satellites, aircraft, drones, etc., using sensors that capture information in multiple spectral bands. We will use multispectral data from the Landsat-8 satellite, sentinel-2 satellite, and ground data samples [19]. Table 1 shows the acquisition details of both the satellites.

Table 1: Acquisition details of satellite Images

Figure 2: Methodology Flow Diagram

Ground Survey for Data Collection:

Training data is very necessary while developing a robust machine learning model for remote sensing multispectral data. To collect Limestone training samples Geo Survey App was utilized. Geo survey app is designed and developed by the National Center of Big Data and Cloud Computing to help surveyors/scientists/researchers collect data using their smartphones [20]. Figure 3 provides a visual overview of the application, with Fig 3(a) showing the main menu having several survey methods. By tapping the desired method in the main menu, a polygon/polyline can be drawn around the survey area, as illustrated in Fig.3(b)(c)(d). The application was programmed using JAVA language and saved the survey data in Google’s database. Google’s database then provides data in JSON format which is then further converted into KML using Python language. ENVI was employed to convert the KML into shapefiles for testing and training the proposed technique’s/Model's performance.

Figure 3: (a) Main menu of application (b) Polygon drawn by tapping (c) View of survey data on map (d) List view of collected data

Data Processing:

Data processing involves the extraction of useful information by manipulation and analysis of data acquired from remote sensing instruments. The multispectral data processing is divided into the following steps:

Spatial Resampling:

Resampling is commonly performed in order to adjust the spatial resolution of satellite images. Depending on the specific requirements of the analysis or application either increase or decrease the spatial resolution of the image. In this study, we have increased image resolution, by using Bi-linear interpolation, as it creates new pixel values based on the weighted average of the surrounding pixels in a 2x2 neighborhood.

Layer Stacking:

Layer stacking creates a single image with multiple bands from multiple satellite images or raster layers. Each band in the stacked image represents a different spectral band or data layer from the original images. In this study, we have stacked the required bands from Landsat-8 and Sentinel-2 images for our analysis. Sentinel-2 data information is confined in a distinct spectral band which is 13 in number, similarly, Landsat-8 data information is loaded in 11 of its distinct spectral bands. Spectral bands were chosen while taking into account the depth, relevance, and extent of information they present. Keeping in view our study targets only 10 bands from Sentinel-2 were selected with a resolution of 10-20m and 7 bands of Landsat-8 were chosen with a resolution of 30m [21].

Temporal Stacking:

The process of temporal stacking involves aligning the images or raster layers, adjusting their spatial resolution, and assigning each layer to a separate band in the final stacked image. This allows analysts to visualize and analyze temporal changes in the landscape or the environment. Sentinel-2 and Landsat-8 images are temporarily stacked for this study. The resultant temporal stacked image (Composite Image) consists of a total of 17 bands of which 8 bands are from Lansat-8, and the remaining 9 are from Sentinel-2 [22].

Enhancement and Analysis:

The image enhancement techniques improve the visual quality of the data. The most common techniques for image enhancement are Principle Component Analysis (PCA) and Band Ratios, while for image analysis machine learning models are used.

Data Splitting:

A field survey of the proposed area was conducted to gather the samples using the Geo Survey Application. The samples were then segregated into a training dataset and a testing dataset having percentages of 70% and 30%.

Selection of Model:

Support Vector machines and Neural Networks are the two well-known machine learning algorithms/models that we have employed [23].

Support Vector Machines (SVMs):

Figure 4: Classification Using SVM

Figure 5: Simple Output Function of Neural Network with Weighted Inputs

Support Vector Machines (SVMs) fall into the category of supervised machine learning, known for solving the complex problems of regression and classification. SVM separates data into different categories using the best hyper-plane in a high-dimensional space [24]. The hyper-plane is chosen such that it maximizes the distance between the classes, which is known as the margin. In binary classification, two classes are separated by margins, their distance is further maximized to generate the hyper-plane shown in Figure 3. The same is the case with multiclass classification, multiple classes are separated by margins, and their distance is further maximized to generate a hyperplane shown in Figure 4. This hyperplane can be linear or non-linear, depending on the kernel used [25][26].

Once the SVM Model is trained, it has the capability to label the new unlabeled data into different classes. The SVM model maps the new data points into the same feature space as the training data and then predicts the class label. Handling high-dimensional data, non-linearly separable data, and avoiding overfitting are the few key features SVM offers over other machine learning techniques [27].

Neural Network:

A neural network Model is a complex machine-learning model designed after observing keenly the functional structure of the human brain [28]. Neural networks learn from the patterns found in data and then apply this learning to deal with real word problems e.g. image classification and pattern recognition, natural language processing, and predictive analytics [29].

Neural network architecture is divided into layers with interconnected neurons, the building block of a neural network. The neurons in the input layer receive training/testing data, and each subsequent layer hierarchically processes the data, extracting higher-level features as the data passes through the network. The output layer generates the output which can be a prediction, classification, or pattern based on the data fed into the input layer.

Each neuron performs its predefined computation as it receives inputs and then passes on to the next neurons in the next layer, where it is combined with the outputs of other neurons to produce a new set of outputs. Each neuron is assigned a weight which further helps in the interconnection of neurons. These weights can be adjusted, to minimize the error between the predicted output and the actual output, accordingly. During the training phase labeled data is fed into the input layer and weights are adjusted to achieve the maximum accuracy [30].

Neural networks have several different types, each has its own architecture and area of application. Feed-forward neural networks are considered the simplest, having a single layer of neurons that takes the input data and supplies output accordingly as shown in Figure 6. Neural networks are preferred over traditional machine learning models because of their complex nonlinear learning ability, which helps handle large datasets.

Model Evaluation:

The model is evaluated using the test data. Furthermore, the metrics that are used for model evaluation are overall accuracy, user accuracy, and producer accuracy [31]. The value of these metrics helps us identify if any problem lies in model training or parameter tuning.

Classification performance has been evaluated using the following parameters.

Confusion Matrix:

To evaluate the performance of a machine learning model confusion matrix is used. A confusion matrix is a table that provides a detailed view of classification models [32]. It specifies the classes where the model has promised results and the classes with dismal results. The confusion matrix can be extended to calculate other evaluation metrics, such as overall accuracy, user accuracy, and producer accuracy.

Some commonly used metrics calculated based on the confusion matrix include Overall Accuracy: The proportion of properly identified observations, calculated as;

Figure 6: A Simple Feed-Forward Neural Network Architecture

User Accuracy:

It is a measure of the probability that a specific class on the map corresponds to the same class on the ground, it assesses how well a particular land cover class is correctly identified or classified by the algorithm. High user accuracy indicates a low rate of false positives or misclassifications. Calculated as;

Producer Accuracy:

It is a measure of the probability that a specific class on the ground is correctly classified on the map. It evaluates how well the model correctly detects and classifies a specific land cover type. A high producer's accuracy indicates a low rate of false negatives or omission errors. Calculated as;

Kappa Coefficient:

The kappa coefficient is a statistical tool used to evaluate the degree of concurrence between two raters or classifiers [33]. The kappa coefficient is calculated by comparing the observed degree of concurrence between two raters or classifiers with the expected degree of concurrence that would occur by chance. Its value varies from -1 to 1, as the value goes towards 1 the degree of concurrence increases, and as the value moves towards 0 the degree of agreement falls which can be called by chance. Negative values indicate that the level of agreement is worse than the chance agreement. The kappa coefficient value is calculated:

Where Po is the observed degree of concurrence between the two raters or classifiers. Pe is the expected degree of concurrence by chance. The expected level of agreement by chance is calculated as follows:

Where a1 to an and b1 to bn are the marginal totals of the contingency table (the total number of times each rater/classifier assigns each category),n is the sum of all the observations.

Training and Testing of Model:

Models (SVM, NN) were trained and tested using the ground samples collected via the Geo Survey App. The data was categorized into training and testing samples. Each Category has four classes namely Limestone, Urban, Barren, and Vegetation. Each class was indicated with different colors, red indicates the Limestone, blue indicates the urban, brown indicates the barren, and green indicates vegetation. After successful training of the model, we tested the model using test samples collected from the ground survey. Training and testing lead to the generation of prospectivity maps which show the detailed contribution of each class in the pilot region.