Dwarf Mongoose Optimization with DL Based Soil Classification Model for Precision Agriculture

Precision agriculture allows precise consumption of inputs like fertilizers, seed, water, and pesticides timely for the crop to maximize quality, yields, and productivity [1, 2]. By deploying sensors for mapping fields and data collection, farmers could understand the field in a better way conserve the resource being used, and decrease adversarial effects on the environment. Mostly, the farmers practice outdated farming patterns for deciding on crops to be cultivated in the field [3]. But the farmer doesn’t perceive crop yields as interdependent on climatic conditions and soil characteristics. Forecasting the variety of crops in specific regions is a challenging and very difficult task since it is influenced by the climatic parameters and type of soil [4]. Moreover, the crop is based on the type of technique utilized by farmers in field-to-field, thus forecasting the performance of Crop types in the parametric perspective is challenging. With the growing population, it exists a substantial demand for crops worldwide [5]; therefore, farmer needs that aware of crop variety which is fixable to the soil type and geographical location [6]. Thus, there is a need for providing timely-based, accurate data according to the soil type to the farmers and climatic parameters, helping them make better decisions for their soil, resulting in great productivity and profitability [7].

Climatic parameters like, precipitation, sunlight, temperature and humidity employ a reflective effect on crop yields. Their complex interaction directly influences crop progress, improvement as well as complete productivity. Temperature affects amount of plant development and flowering while precipitation ranks are serious for sufficient soil moisture. Sunlight and humidity are vital for photosynthesis as well as creation of plant structures [7]. But, huge challenge in predicting crop yields lies in complex nature of weather parameters, their temporal and spatial variations, nonlinear relations among these aspects and crop results. Furthermore, weather change presents extra irregularity that makes complex to accurately expect how shifts in climatic situations will influence prospect crop yields that highlighting requirement for advanced modeling as well as data-driven techniques in agricultural estimating. The superior of accurate crop variety is dominant in enhancing agricultural harvest. Dissimilar crop variations varying acceptance levels for exact environmental circumstances, types of soil and geographical areas.

Soil plays a vital part in hydrologic sequence by performing as a reservoir for water and simplifying its effort via landscape. The soil's capability is to grip and relief water effects availability for crops, surface runoff and groundwater recharge. Soil assets and their relations with vegetation and precipitation command its part in changeable water quality and flow. Modeling surface excess is vital to water and soil conservation efforts as it aids to forecast water movement across scenery. Surface runoff techniques consider causes like vegetation cover, type of soil, topography, and precipitation to pretend water movement paths and recognize prospective sedimentation and erosion threats. By evaluating and modifying surface runoff, land managers and conservationists can able to decrease soil erosion, guard water excellence, and improve complete sustainability in environmental and agricultural situations. Flooding and Erosion caused by uncontrollable runoff, mostly down-stream, cause damage to agricultural land and manmade structure [8]. Consequently, modelling surface runoff is a critical part of soil and water conservation effort including but has not been constrained to, predictive floods, and monitoring soil, soil erosion, and water quality [9]. Soil could be classified into silty sand, clayey sand, sandy clay, humus clay, clay, clayey peat, and peat. Numerous soil classifier techniques are presented in this work. The boundary and signal energy model for feature extraction [10].

This study develops a Dwarf Mongoose Optimization with DL Based Soil Classification (DMODL-SC) model. The presented DMODL-SC technique majorly recognizes different kinds of soil using CV and DL models. In the presented DMODL-SC technique, bilateral filtering (BF) technique is used for noise removal process. In addition, the presented DMODL-SC technique employs capsule network (CapsNet) model for feature extraction process. Moreover, DMO with denoising auto encoder (DAE) is exploited for the identification and classification of soil. To demonstrate the enhanced performance of the projected DMODL-SC system, an extensive range of experiments were performed.

3.1 Image pre-processing

In the presented DMODL-SC technique in Figure 1, the BF technique is used for noise removal process. The BF approach offers the advantage of automated censoring, less noise, rotation symmetric, and ease to design [17]. The input image may have noises, involving salt pepper noise, Gaussian, etc. Noise removal application preserves the information on the input data similarly. The BF method is employed for denoising the input images. This is accomplished by integrating two Gaussian filters, viz., the spatial domain, another one that operates the intensity domain, and the one that is functioning. The resultant at $p$ pixel place was determined as:

$\bar{F}\left( p \right)=\frac{1}{N}\mathop{\sum }_{z\epsilon S\left( p \right)}e\frac{-\|q-p{{\|}^{2}}}{2\varepsilon _{e}^{2}}\frac{-|F\left( q \right)-F\left( p \right){{|}^{2}}}{2\mathcal{E}_{S}^{2}}F\left( q \right)$                      (1)

Now, the normalization constant is denoted by $N$, $S\left( p \right)$ characterize a pixel spatial neighborhood $F\left( p \right)$, and variable ${{\varepsilon }_{e}}{{\varepsilon }_{r}}$ are governing weightedin the domain of spatial and intensity begin with fall off.

${{e}^{\frac{-\|q-p{{\|}^{2}}}{2\varepsilon _{e}^{2}}}}{{e}^{\frac{-|F\left( q \right)-F\left( p \right){{|}^{2}}}{2\varepsilon _{e}^{2}}}}$                  (2)

The BF was employed in texture removal, tone mapping, volumetric de‐noising, and another application like de‐noising the image is utilized. They could produce simpler conditions for downsampling the procedure and accomplish acceleration by demonstrating in the augmented space; with 2 nonlinearities, the BF has been implemented as linear convolutional.

image010.jpg

Figure 1. Block diagram of DMODL-SC system

3.2 Feature extraction using CapsNet

At this stage, the presented DMODL-SC technique employed the CapsNet model for feature extraction process. A primary advantage of CapsNet is that holds the features of further concrete characteristics that aretakenfor understanding how and what is the network learning. The CapsNet can able to encode spatial information and distinguish amongst different orientations, poses, and textures [18]. The capsule is a collection of neurons, thereby each capsule has an activity vector interconnected with it that captures many instant parameters to recognize specific type of object or its part. The orientation and length of vector presented the possibility or probability of existence of that object and the generalization pose. This vector was passed onto the upper level capsule in lower layer capsule. The coupling coefficient occurs amongst the capsule layers. Once the prediction by the lower level capsule equals the result of existing capsule, the values of coupling coefficient amongst them increase, computed by using softmax function. Particularly, once the existing capsule recognizes a tight cluster of previous prediction, and intensely represent the existence of that object, its results in a high probability are also identified as routing by agreement. Figure 2 illustrates the architecture of CapsNet module:

${{\hat{u}}_{{{j}_{|i}}}}\wedge ={{W}_{ij}}{{u}_{i}}$                 (3)

In Eq. (3), ${{\hat{u}}_{{{j}_{|i}}}}$ represent the results of prediction vector of upper‐level ${{j}^{th}}$ capsules, ${{W}_{ij}}$ and ${{u}_{i}}$ indicates the weight matrixes and prediction vector of $i$-$th$ capsules in lower layer. It captures spatial connection and interaction amongst objects and sub‐objects. In Eq. (4), based on degree of agreement among nearby layer capsules, the coupling coefficient was evaluated by means of softmax function:

${{c}_{ij}}=\frac{exp\left( {{b}_{ij}} \right)}{\sum exp\left( {{b}_{ik}} \right)}$               (4)

where,${{b}_{ij}}$ indicates the$~log$ probability among 2 capsules, initialization to 0, and $k$ denotes the capsule count. The input vector ${{s}_{j}}$ to ${{j}^{th}}$ layer capsules that a weight sum of ${{\hat{u}}_{{{j}_{|i}}}}\wedge $ vector learned by routing method is calculated by the following equation:

${{s}_{j}}=\underset{i}{\mathop \sum }\,{{c}_{ij}}{{\hat{u}}_{{{j}_{|i}}}}$               (5)

Finally, a squashing purpose which incorporates squashing and unit scaling (Eq. (6)) was implemented for restraining the value of outcomes in the range among [0, 1], thus evaluating the probability as:

$\|{{s}_{j}}{{\|}^{2}}{{v}_{j}}=\underline{1+\|{{s}_{j}}{{\|}^{2}}}\frac{{{s}_{j}}}{{{s}_{j}}}$             (6)

The loss function (interconnected to capsule from the concluding layer, while $m+ar\iota d$ m‐ are set as 0.9 and 0.1 resp.

${{l}_{k}}={{T}_{k}}\text{max}{{(0,~{{m}^{+}}-\left| \left| {{v}_{k}} \right| \right|)}^{2}}+\lambda \left( 1-{{T}_{k}} \right)~max~{{(0,~\left| \left| {{v}_{k}} \right| \right|-{{m}^{-}})}^{2}}$            (7)

In Eq. (7), the value ${{T}_{k}}$ is 1 to accurate labels and $0$ otherwise, $\lambda $ represents the constant that value is 0.5. The first term can be evaluated for correcting labels, and the next term evaluates incorrect labels.

image031.jpg

Figure 2. Structure of CapsNet

3.3 Soil classification using DAE

For soil classification, the DAE model is applied. DAE is a novel deep network which encompasses AE using different hidden layers and takes sensitive power [19]. For classifier issue, the softmax classification is extensively chosen by the resulting layer. Then, reformation of input instance with less error:

$\left( X,~Y \right)=\{\left( {{x}^{\left( n \right)}},~{{y}^{\left( n \right)}} \right)|n=1,2,\ldots ,N)\}$                (8)

Now, ${{y}^{\left( n \right)}}$ represent aninstance trademark ${{x}^{\left( n \right)}}.~$The instance count can be represented by N. For all the instances of trained data ${{x}^{\left( n \right)}}$, code encoding through ${{h}^{\left( n \right)}}=f\left( {{x}^{\left( n \right)}} \right)$ afterward, decode ${{h}^{\left( n \right)}}$ for recreating with ${{x}^{\left( n \right)}}=g\left( {{h}^{\left( n \right)}} \right),~f$ and $g$ are encoder and decoder parameters.

${{h}^{\left( n \right)}}=s\left( W{{x}^{\left( n \right)}}+b \right)$             (9)

${{x}^{\left( n \right)}}=s\left( W{{h}^{\left( n \right)}}+b \right)$              (10)

The sigmoid function was characterized by $s\left( \cdot \right)$ a trained data by means of energy consumption in the following:

$\left( \theta  \right)=\frac{1}{N}\mathop{\sum }_{n=1}^{N}\frac{1}{2}|\left| {{x}^{\left( n \right)}}-x \right||_{2}^{2}$           (11)

Variable absence from $s$&$\theta $ from linear transformation. Standard AE is fundamental for the DAE system which encoded ${{x}^{\left( n \right)}}$ to hidden notation ${{h}^{\left( n1 \right)}}$ that is given to subsequent input port of DAE. The resulting layer is comprised of top hidden layer to monitor the trained system. Each layer produces a better outcome due to the training of designed parameters. Fine-tuning is widely employed in NN as a global optimized algorithm therefore it improves accuracy of the DAE.

$J\left( W,~b;~{{x}^{\left( n \right)}},{{y}^{\left( n \right)}} \right)=\frac{1}{2}|\left| {{y}^{\left( n \right)}}-{{y}^{\left( n \right)}} \right||_{2}^{2}$              (12)

The energy function $J\left( W,~b \right)$ forces the outcomes closer to true label all over the whole preparation and determines the fine-tuning procedure.

$J\left( W,b \right)=\frac{1}{N}\mathop{\sum }_{n=1}^{N}J\left( W,b;{{x}^{\left( n \right)}},{{y}^{\left( n \right)}} \right)$           (13)

In Eq. (13), $\left( W,~b \right)=\{\left( {{\text{W}}^{\left( l \right)}},~{{b}^{\left( l \right)}} \right)|1=1,2,\ldots ,L\}$ indicates encoder constraint of the entire layer.

3.4 Parameter tuning using DMO algorithm

In the final stage, the DMO algorithm has been exploited for the hyper parameter tuning of the DAE. DMO algorithm stimulates the performance of dwarf mongooses while defining the food [20]. In general, the DMO starts with setting the primary values to a group of solutions utilizing the succeeding equation:

${{x}_{i,j}}={{l}_{j}}+rand\times \left( {{u}_{j}}-{{l}_{j}} \right)$           (14) 

In Eq. (14), rand represents arbitrary value within [0, 1]. ${{u}_{j}}$ and$~{{l}_{j}}$ indicates the restriction of the search area. The swarming of DMO has 3 groups namely scouts, babysitters, and alpha groups. Each group has individual performance to capture the food, and particular of this group is given below.

The fitness of each solution is evaluated when the population was determined. Eq. (15) calculates the potential value for each fitness population, and alpha female $\left( \alpha  \right)$ is selective and depending on these probabilities $n~$represents the count of mongooses from the alpha group. The babysitter count was characterized as bs. Peep refers to the vocalization of dominant females that keeps the family on track.

$\alpha =\frac{fi{{t}_{i}}}{\Sigma _{i=1}^{n}fi{{t}_{i}}}$            (15)

Each mongoose sleep from the initial sleeping mound that is set $\varnothing $. The DMO exploits to generate a candidate food location.

${{X}_{i+1}}={{X}_{i}}+phi\times peep$           (16)

The sleeping mound was shown below, then each repetition, while $phi$ indicates the uniform distribution number within [-1 and 1$].$

$s{{m}_{i}}=\frac{fi{{t}_{i+1}}-fi{{t}_{i}}}{max~\left\{ \left| fi{{t}_{i+1}},~fi{{t}_{i}} \right| \right\}}$            (17)

Eq. (18) includes the average value of sleeping mounds.

$\varphi =\frac{\Sigma _{i=1}^{n}s{{m}_{i}}}{n}$             (18)

When the babysitting modifies criteria are satisfied, this technique progress to the scouting stage, whereas the next food source/sleeping mound is regarded.

Since mongooses are recognized to not back to historical sleep mounds, the scout appearance is for the second sleeping mounds, ensuring to search. In these methods, forage and scout are performed simultaneously. Then, this drive demonstrated an unsuccessful or successful look for novel sleeping mounds. Particularly, the migration of mongooses is dependent on the whole efficacy.

${{X}_{i+1}}=\left\{ \begin{matrix}

   {{X}_{i}}-CF\text{*}phi\text{*}rand\text{*}\left[ {{X}_{i}}-\vec{M} \right]\text{ }\!\!~\!\!\text{ }if{{\varphi }_{i+1}}>{{\varphi }_{i}}  \\

   {{X}_{i}}+CF\text{*}phi\text{*}rand\text{*}\left[ {{X}_{i}}-\vec{M} \right]  \\

\end{matrix} \right.$                  (19)

In Eq. (19), rand indicates a random value within [0, 1]$,\text{ }\!\!~\!\!\text{ }CF={{(1-\frac{iter}{\text{Ma}{{\text{x}}_{iter}}})}^{\left( 2\frac{iter}{\text{Ma}{{\text{x}}_{iter}}} \right)}}$ whereas the variable regulates the mongoose group.

The babysitters were usually lesser group members that continued with the young allowing the alpha female (mother) to lead remaining groups on everyday forage expeditions. The babysitter count was proportionate to the population size; it can be stimulus the technique by decreasing the whole population size reliant on the percentage set.