Using an Efficient Optimal Classifier for Soil Classification in Spatial Data Mining Over Big Data

4.1 MapReduce Framework

MapReduce is a structure for the competent distribution of the study of huge data on an outsized amount of servers. It is a similar and disseminated large-scale data processing model that has been widely investigated and extensively approved for huge data function recently. Incorporated among communications possessions provisioned by cloud scheme, MapReduce turns out to be much more dominant, flexible, and commercial because of the outstanding uniqueness of cloud computing. The MapReduce framework comprises two phases: map step and reduce step. Figure 2 illustrates the general arrangement of the MapReduce framework. There are five sections in MapReduce: Input (input involved data), Map (filtering and sorting the data), Reduction (redistribute the mapped data using PSA), Reduce (process each group of the redistricted data), and Output (collect all the reduce output). MapReduce courses are normally used to practice big files.

Figure 2:

Flow Diagram of the MapReduce Framework.

A MapReduce framework consists of two major phases: (1) Map phase: This phase inputs input soil data and divides it into M map task; every map task executes differently. (2) Reduce phase: In this phase, the separated soil data from the previous phase is condensed by PCA. The specific clarification of PCA is discussed in the next section.

4.2 PCA

PCA has been described as the majority of the important results from functional linear algebra. PCA is used often in every one of the figures of study, from neuroscience to computer graphics, as it is an easy, nonparametric process of removing appropriate information from confounds of data sets. By a smallest supplementary endeavor, PCA supplies a direction for how to reduce a compound dataset to a lesser measurement to expose the occasionally concealed, shortened dynamics that frequently motivates it. The foremost principles of PCA are the study of data to recognize models and discovery models to reduce the measurement of the dataset among the smallest failure of information.

Our preferred conclusion of the PCA is to offer an attribute space onto a slighter subspace that signifies our data “well”. A potential function would be a model categorization task, where we want to reduce the computational expenses and the fault of limitation assessment by dropping the amount of measurement of our attribute space by removing a subspace that depicts our data “best”.

4.3 Summary of the PCA Approach

Listed below are the six general steps for performing a PCA:

1. Take the whole dataset consisting of d-dimensional samples, ignoring the class labels.

2. Compute the d-dimensional mean vector (i.e. the means for every dimension of the whole dataset).

3. Compute the covariance matrix of the whole data set.

4. Compute eigenvectors (e₁, e₂,…,e_d) and corresponding eigenvalues (λ₁, λ₂,…,λ_d).

5. Sort the eigenvectors by decreasing eigenvalues and choose k eigenvectors with the largest eigenvalues to form a d*k-dimensional matrix M (where every column represents an eigenvector).

6. Use this d*k eigenvector matrix to transform the samples onto the new subspace. This can be summarized by the mathematical equation: y=M^T*x.

From the above procedure, the input soil data is reduced and then the resultant output is fed to the next phase of the MapReduce framework.

In the reduce phase, obtain every one of the condensed data from PCA and construct the productivity of the MapReduce scheme.

Lastly, the condensed productivity from the MapReduce framework is provided for to the categorization procedure. Agriculture is the field that provides food to the entire world [18]. There would be no existence of humans without agriculture. Hence, it is essential to increase the yield of agriculture. The yield of agriculture could be increased only via proper classification of soil [23]. Hence, in our suggestion scheme, the best NN is used for the soil categorization function. Each phase of the system of the best NN is discussed next.

4.4 ONN Classifier

The proposed procedure employs the best NN for soil categorization. The conventional NNs are customized by means of the GWO algorithm. The GWO algorithm is engaged to optimize the influence in the NN. The foremost intention of the ONN is to classify the input spatial soil data. For the preparation function, the backpropagation algorithm is used in our recommended procedure. NN encompasses a sequence of nodes (neurons) that obtain numerous associations among further nodes. Each one association includes an influence related to it, which can be diverse in force, in similarity through neurobiology synapses. The artificial NN (ANN) structure has an input layer and an output layer as well as hidden layers between these two layers. The amount of these layers is reliant on the difficulty we are difficult to explain, that is fundamentally on the consumer. There are two very significant sections in the NN system: the preparation section and the experiment section. The general progression of the best NN is shown next.

4.5 ONN Function Steps

1. Fix weights for every neuron, except the neurons in the input layer.

2. Develop the NN with the input soil data as the input units, H_u as the hidden units, and O_u as the output unit.

3. The computation of the proposed bias function for the input layer is

In the proposed best NN, the influences are optimized using the GWO algorithm. Each system of the GWO is demonstrated in the next section.

4.6 GWO Algorithm

The grey wolves adequately enclose Canidae species ancestors and are esteemed as the apex predators offering their location at the wherewithal’s food sequence [21]. They habitually illustrate a partiality to construct suitable as a cluster. The heads represent a male and a female, marked as α, which is the majority division in incriminating of enchanting appropriate selection presenting different features; they are fundamentally supplementary wolves that effectively suggest a few support to the α in the option constructing or equivalent cluster task. The choices prepared by the α are permitted onto the group. The β conveys to the second grade in the outstanding arrange of the grey wolves. They are basically complementary wolves that adequately treaty a number of assist to the α in the option generating or corresponding group performance. The omega, which is the smallest division of the grey wolf group, by and great task as a replacement contribution into the further primary wolves very practically on each occasion and is acceptable to obtain just the diminutive leftovers enchanting following an enormous blowout by the organizer wolves. In the GWO process, the tracking (optimization) is directed by then α, β, δ, and ω. For picking the best influence, the proposed procedure employs the GWO algorithm. The pseudo code for the customized GWO algorithm is demonstrated below.

The step-by-step process of the GWO algorithm is mentioned below.

Step 1: Initialization process

Initialize the input random weights and a, A, and C as coefficient vectors.

Step 2: Fitness evaluation

Evaluate the fitness performance based on Eq. (2) and then pick the best result.

Step 3: Separate the solution based on the fitness

Now, determine the different result on the foundation of the fitness value. Let the first best fitness results be w_α, the second best fitness results w_β and the third best fitness results w_δ.

Step 4: Update the position

We presume that the α, β, and δ obtain the enhanced facts about the probable position of the prey to replicate precisely the tracking activities of the grey wolves. Because of the result, we accumulate the primary three best influences accomplished until now and necessitate the further search influence (including the omegas) to modify their situation along with the situation of the best search influence. For replication, the innovative weight w_p(t+1) is predictable by the formulas below:

where K→α= |C→1.wpα−wp|,K→β= |C→2.wpβ−wp|,K→δ= |C→3.wpδ−wp|

where wp1=wpα−A→1⋅(K→α), wp2=wpβ−A→2⋅(K→β),  w3p=wpδ−A→3⋅(K→δ) 

where t is the iteration number, p(t) is the prey location, A and C are the coefficient vector, a→ is linearly reduced from 2 to 0, and r₁ and r₂ are the random vector [0, 1]. It can be distinguished that the final influence would be in an arbitrary position prearranged a circle, which is precise by the position of α, β, and δ in the investigate break. It also meant by α, β, and δ evaluate the location of the prey and supplementary wolves update their location randomly in the region of the prey. Examination and utilization are specific by means of the adaptive values of a and A. The adaptive values of limitation a and A allow GWO to effortlessly changeover in the middle of examination and utilization. By retreating A, half of the iterations are dedicated to the examination (|A|<1) and the other half is devoted to the convention. Attaching the demeanor, the following equations are used keeping in mind the conclusion objective to provide arithmetical representation.

Step 5: Fitness calculation

Calculate the fitness of the new search weight using Eq. (2) and then store the best solution.

Step 6: Stopping criteria

Replicate Steps 3–5, awaiting an improved fitness or highest amount of iterations are collected. Derived from on top of the declared procedure, accomplish the best influence. After that, the best influence is used for the additional procedure.

• The activation function for the output layer is estimated as

  (6) Active (S)=11+e−S

• Recognize the learning error as follows:

  (7) Output(O)u=12 ∑n = 0Hu− 1(Desiredn−Actualn′)2

where Desired_n is desired output and Actual_n is actual output.

In the best NN, the mistake should be in the least value. Subsequently, the qualified NN is well qualified for presenting the experiment stage. Derived from the qualified data, the NN arrangement efficiently categorized the soil category. The analysis result of the proposed procedure is discussed in Section 5.

5.1 Dataset Description

The proposed process was investigated by the Harmonized World Soil Database version 1.2.

The Harmonized World Soil Database is a 30-arc second raster database with more than 15,000 different soil mapping components that merge local and countrywide updates of soil information worldwide (SOTER, ESD, Soil Map of China, WISE) from the information restricted within the 1:5,000,000 scale FAO-UNESCO Soil Map of the World. The resulting raster database consists of 21,600 rows and 43,200 columns, which are linked to harmonized soil property data. The dataset is available at http://www.fao.org/soils-portal/soil-survey/soil-maps-and-databases/harmonized-world-soil-database-v12/en/.

5.2 Evaluation Metrics

The assessment metrics for the performance of the proposed scheme are sensitivity, specificity, accuracy, precision, recall, and F-measure. The standard count values such as true positive (TP), true negative (TN), false positive (FP), and false negative (FN) are presented.

5.3 Performance Analysis

The proposed procedure is discussed in the next section. Table 1 illustrates the performance analysis of the proposed procedure. The sensitivity, specificity, accuracy, precision, and recall, and F-measure values are presented. In our procedure, 90% of the input data is allowed for preparation and the outstanding 10% is allowed for the experiment. The performance analysis of the proposed procedure with varying data size is discussed below.

In Table 1, using the proposed process for the data dimension of 1000, the sensitivity value is 0.89676%, the specificity value is 0.78947368%, the accuracy value is 0.9046%, the precision value is 0.98979%, the recall value is 0.89676%, and the F-measure value is 0.940986%. For the data dimension of 2000, the sensitivity value is 0.865787%, the specificity value is 0.81818%, the accuracy value is 0.913846%, the precision value is 0.948979%, the recall value is 0.865787%, and the F-measure value is 0.905476%. For the data dimension of 3000, the sensitivity value is 0.89506%, the specificity value is 0.75%, the accuracy value is 0.9797%, the precision value is 0.9797%, the recall value is 0.89506%, and the F-measure value is 0.935484%. For the data dimension of 4000, the sensitivity value is 0.83553%, the specificity value is 0.78409%, the accuracy value is 0.881538%, the precision value is 0.944737%, the recall value is 0.8355%, and the F-measure value is 0.88678%. For the data dimension of 5000, the sensitivity value is 0.8416499%, the specificity value is 0.77777%, the accuracy value is 0.86399%, the precision value is 0.972163%, the recall value is 0.841649%, and the F-measure value is 0.902211%.

Table 2 presents the performance analysis by the cross-validation method with a steady data dimension of 1000. For cross-validation 1, the sensitivity value is 0.82265%, the specificity value is 0.75%, the accuracy value is 0.854769%, the precision value is 0.9638%, the recall value is 0.826%, and the F-measure value is 0.886443% for the preparation percentage of 70 and the experiment percentage of 30. For cross-validation 2, the sensitivity value is 0.85604%, the specificity value is 0.775862%, the accuracy value is 0.863999%, the precision value is 0.98095%, the recall value is 0.85265%, and the F-measure value is 0.91817% for the preparation percentage of 80 and the experiment percentage of 20. For cross-validation 3, the sensitivity value is 0.89676%, the specificity value is 0.78947368%, the accuracy value is 0.9046%, the precision value is 0.98979%, the recall value is 0.89676%, and the F-measure value is 0.940986% for the preparation percentage of 90 and the experiment percentage of 10. The average accuracy value for the cross-validation method is 87.44%.

5.4 Effectiveness of the Suggested Technique

In this section, the efficiency of the proposed procedure is compared to a further obtainable procedure. The detail clarification is demonstrated in Figure 3.

Figure 3:

Overall Process of the ONN.

5.5 Comparison Analysis of Performance Metrics

The accuracy value is the foremost feature in the soil categorization scheme. It is significant for the process to offer the elevated accuracy value to provide the best process. Figure 4 illustrates the proportional study for the accuracy value by the obtainable categorization process. We allow the obtainable categorization scheme as the conventional NN and k-nearest neighbor (KNN) classifier.

Figure 4:

Accuracy Comparison for Different Classifiers.

Figure 4 illustrates the accuracy of the proposed ANN-GWO compared to the accuracy of the obtainable classifiers such as NN and KNN. The conventional NN classifier has 75.3846% accuracy and the KNN classifier has 75.38% accuracy, but the proposed process has 90.46% accuracy, which is the highest value compared to the obtainable procedure.

Figure 5 illustrates the sensitivity of the proposed ANN-GWO compared to the sensitivity of the obtainable classifiers such as NN and KNN. The conventional NN classifier has a sensitivity value of 0.75% and the KNN classifier has 0.7538% sensitivity value, but the proposed classifier ANN-GWO has 0.89676% sensitivity value, which is the highest value compared to the obtainable classifier procedure.

Figure 5:

Sensitivity Comparison for Different Classifiers.

Figure 6 illustrates the specificity of the proposed ANN-GWO compared to the specificity of the obtainable classifiers such as NN and KNN. The NN classifier has a specificity value of 0.77166% and the KNN classifier has 0.811269% sensitivity value, but the proposed classifier ANN-GWO has 0.789474% sensitivity value. Compared to the obtainable procedure, the recommended procedure has the least specificity value. Although it has the least specificity value, the categorization accuracy is the highest value compared to the further categorization procedure.

Figure 6:

Specificity Comparison for Different Classifiers.

Figure 7 illustrates the precision and recall of the proposed ANN-GWO compared to the precision and recall of the obtainable classifiers such as NN and KNN. The precision value of the NN classifier is 0.8229% and the KNN classifier has 0.81887% precision value, but the proposed classifier ANN-GWO has 0.98979% precision value. Compared to the obtainable process, the proposed categorization procedure has the highest precision value. The recall value of the NN classifier is 0.75% and the recall value of the KNN classifier is 0.7538%, whereas the proposed classifier has 0.89676% recall value, which is better than the obtainable classifiers.

Figure 7:

Precision and Recall Comparison for Different Classifiers.

Figure 8 illustrates the F-measure value of the proposed ANN-GWO compared to the F-measure value of the obtainable classifiers such as NN and KNN. The NN classifier has 0.784768% F-measure value and the KNN classifier has 0.8792% F-measure value, but the proposed classifier has 0.940986% F-measure value, which is better than the obtainable classifiers.

Figure 8:

F-Measure Comparison for Different Classifiers.

In Figure 9, the proposed method compares the result for PCA with and without eigenvector k value. The performance is evaluated by accuracy, sensitivity, and specificity values. Here, the proposed PCA with eigenvector has an accuracy value of 90.46%, the sensitivity value is 0.89676%, and the specificity value is 0.789%, but the proposed system PCA without eigenvector has an accuracy value of 87.35%, the sensitivity value is 0.8661%, and the specificity value is 0.7556%, which is the minimum value compared to PCA with eigenvector. In the graph, PCA with k value improves the proposed performance compared to PCA without k value. The comparison result for with and without optimization is shown in Table 3.

Figure 9:

Comparison Results for PCA with and without k Value.

Table 3 shows the comparison result for soil classification performance with and without optimization. It is clearly shown that the proposed technique with the optimization method has a better result compared to that without the optimization method. Here, the accuracy value of the proposed soil classification with optimization is 90.46% but that without the optimization method is 88.72%, which is minimum value compared to the proposed technique. From the result, the proposed method concludes that classification with the optimization method outperforms the classification without optimization.

The proposed performance is compared to the existing technique [7] in Figure 10. For comparison, the proposed technique considers the existing support vector machine (SVM) and PCA with SVM.

Figure 10:

Comparison Results for Various Existing Methods.

Figure 10 illustrates the accuracy value of the proposed ANN-GWO compared to the accuracy value of the obtainable classifiers such as SVM and PCA + SVM.

For cross-validation 1, the accuracy value of the SVM classifier is 81.44% and PCA + SVM classifiers have 72.63% accuracy value, but the proposed classifier ANN-GWO has 85.47% accuracy value.

For cross-validation 2, the accuracy value of the SVM classifier is 59.61% and PCA + SVM classifiers have 80.61% accuracy value, but the proposed classifier ANN-GWO has 86.399% accuracy value.

For cross-validation 3, the proposed classifier ANN-GWO has 90.46% accuracy value, but the existing classifier has 75.94% for the SVM classifier and 59.31% for the PCA + SVM classifier. Compared to the obtainable process, the proposed categorization procedure has the highest accuracy value. As a result, the recommended procedure has the highest result compared to the obtainable classifier for spatial data soil categorization.