Enhanced Coconut Yield Prediction Using Internet of Things and Deep Learning: A Bi-Directional Long Short-Term Memory Lévy Flight and Seagull Optimization Algorithm Approach

Abstract

In coastal areas, coconuts are a common crop. Everyone from farmers to lawmakers and businesses would benefit from an accurate forecast of coconut production. Internet of Things (IoT) sensors are strategically positioned to continuously monitor the environment and gather production statistics to obtain accurate agricultural output predictions. To effectively estimate coconut prediction, this study presents an enhanced deep learning classifier called Bi-directional Long Short-Term Memory (BILSTM) with the integrated Lévy Flight and Seagull Optimization Algorithm (LFSOA). LASSO feature selection is applied to eliminate the superfluous characteristics in the yield estimation. To further enhance the coconut yield estimate, the optimal set of hyperparameters for BILSTM is tuned by the LFSOA, which helps to avoid the overfitting issue. For the results, the BILSTM is compared against different classifiers such as Recurrent Neural Network (RNN), Random Forest Classifier (RFC), and LSTM. Similarly, LFSOA-based hyperparameter tuning is contrasted with different optimization algorithms. The outputs show that LFSOA-based hyperparameter tuning in BILSTM achieved accuracy, precision, recall, and f1-score of 98.963%, 99.026%, 99.155%, and 95.758%, respectively, which are higher when compared to existing methods. Similarly, the BILSTM-LFSOA accomplished better results in statistical measures, including the Root Mean Square Error (RMSE) of 0.105, Mean Squared Error (MSE) of 0.011, Mean Absolute Error (MAE) of 0.094, and coefficient of determination (R²) of 0.954, respectively. From the overall analysis, the proposed BILSTM-LFSOA improves coconut yield prediction by achieving better results in all the performance measures when compared with existing models. The results of this study are important to many stakeholders, including but not limited to policymakers, farmers, banks, and insurance companies. As coconuts are an important crop in developing countries, accurate coconut yield forecasting will lead to greater financial and food security in these regions.

1. Introduction

Agriculture has a substantial impact on the development of the Gross Domestic Product (GDP) of developing countries. Thus, accurate crop yield estimation is important for food security and maintaining population health [1,2]. Crop yields are estimated by using different factors such as nutrients, management practices, and weather. These factors are used for predicting yields over a large scale in a timely and concise manner. The purpose of this crop yield calculation is to guard against climate risk and guarantee food security, which is becoming increasingly important given climate change and the worsening of extreme weather conditions [3,4,5]. As the global population increases, the available arable land decreases, necessitating more effective utilization of land and enhancement of yields over expanding cultivated land [6]. Policymakers rely on various advisory and management services that include yield forecasts and uncertainty assessments to make timely import and export policy decisions. At the national level, yield prediction is mostly important for strategic and staple crops [7,8].

Over time, land measurements for crop yields have been replaced by more sophisticated methods of estimation that rely on crop growth and management-based approaches. These are not only used to estimate yields but also provide suggestions for crop yield management [9]. As a result, data and information generated by agricultural equipment are the foundation of smart farms that are putting manufacturers in contact with digital technologies. Smart farms collect precise data using sensors and drones, such as weather information and soil mapping, among other things.

To fulfill the population’s food needs and ensure sustainable use of natural resources, it is becoming more vital to extract information from this data and create decision-support systems for farms [10]. Therefore, the agricultural industry uses Internet of Things (IoT) technology for predictive analysis on a large scale. The global information society is built on a foundation of fundamental and emergent Information and Communication Technologies (ICT), which together make up the IoT. Through the integration of virtual and physical items, the IoT offers advanced services [11].

New developments in IoT technology have made it possible to gather exact climatic data alongside precise and reliable yield data. High-quality data is readily available, but current yield prediction techniques still have a lot of obstacles to overcome, including difficulties in handling high-dimensional data, an inability to capture complicated non-linear correlations, and a lack of fundamental knowledge. These drawbacks highlight the need for a novel deep-learning model that captures complex connections, makes accurate forecasts, and efficiently uses reliable yield and climate data. Therefore, this study employs a deep learning architecture with hyperparameter optimization to predict coconut yields.

It is difficult to calculate how several climatic elements, such as temperature, humidity, wind, and precipitation, interact with one another and affect crop yields. These requirements can be met by a deep learning model (LSTM), which enables data-driven agricultural decision-making by producing precise and reliable crop predictions.

One of the most important tropical crops, coconuts are grown on 12.08 million hectares in 92 countries and yield 69 billion nuts annually [12]. Through processing, farming, marketing, and trade-related activities, coconuts provide 20 million people worldwide with food security and a chance at a healthy life, including 10 million people living in India alone. Consequently, rural economies are significantly impacted by the global coconut value chain [13]. Rainfall and air temperature are considered the main parameters since they affect the female flower’s receptivity and pollen’s feasibility which determines the yield under open field conditions [14,15].

The existing models are insufficient to reliably estimate coconut yields because of their inability to handle complex environmental factors and non-linear connections. Furthermore, the influence of environmental conditions on coconut yields has not been taken into account properly by the earlier models. This study presents a reliable and accurate coconut yield prediction model to support informed decision-making in coconut farming and agricultural research by filling these gaps and accomplishing the research objectives. Based on the results of this study, Hyperparameter Optimization (HPO)-based classification is recommended. Further, automated HPO is designed to minimize human effort and improve classification.

The concise contributions are as follows:

• To perform crop yield prediction, continuous data from coconut yields is collected using IoT devices. The gathered data includes information about the past 10 years from 2011 to 2021.
• Further, redundant attributes in the collected data are removed by using LASSO before performing the prediction.
• The BILSTM classifier uses historical and future data to estimate the coconut yield while the LFSOA optimizes hyperparameters. The BILSTM offers an effective prediction by using the information from both the backward and forward directions.
• The incorporation of LF into conventional SOA provides a means to effectively navigate complex search spaces and identify optimal results, leading to improved handling of the control search and helping in achieving a better set of hyperparameters.
• The developed LFSOA-based hyperparameter tuning helps to perform reliable prediction by avoiding overfitting risk. The proposed BILSTM-LFSOA predicts coconut yields in terms of statistical and classification measures.

The preexisting literature has suffered from ineffective data processing during the prediction process, a failure to analyze complex and non-linear dependencies, and risks of overfitting. This research seeks to fill this research gap in the following manner: IoT sensors are used to obtain field data, thus overcoming the problem of reliable data. Meanwhile, the usage of past and future data in the BILSTM model captures complex and non-linear dependencies and therefore enhances the predictive power of the model. Finally, hyperparameter tuning using the LFSOA reduces the overfitting risk.

The structure of the paper is arranged as follows. The related works about the forecast of agricultural output are included in Section 2. Section 3 provides detailed information about the BILSTM-LFSOA, while the results and comparisons are given in Section 4. Finally, Section 5 presents the conclusion.

2. Related Work

Related works about coconut yield along with its prediction and some other types of crop yield estimation are discussed in this section.

2.1. Coconut Yield and Its Prediction

Novarianto [16] developed research for discovering, evaluating, and estimating the local tall coconut production and productivity at Taliabu Island. Drone technology was integrated with the classical sampling population to improve the efficiency of data collection. The collected data was used to maximize the production of coconuts by utilizing input technology. This increased coconut production enabled coconut palms to generate high yields per unit area, which improved the sustainability of the integrated coconut industry.

The Ordinary Least Square (OLS) regression and Quantile Regression (QR) approaches were introduced by Samarakoon et al. [17] for estimating production in Cobb Douglas functional form. If input was considered without lags, the amount of bearing palms created a positive effect during OLS estimation. An amount of bearing palms and one-year lag rainfall created a positive effect in nut production during QR results. The strong/weak relationship over the coconut yield distribution was found using the QR. However, the developed OLS overestimates and underestimates an appropriate relationship among covariates and response variables.

Hadi [18] studied problems in coconut plantations that impact selling price, income, and economic value to welfare. A quantitative study using regression analysis was developed as a method for estimating causal relationships between variables. Here, samples for this work were randomly acquired from 150 coconut farmers in the Kalimantan region. The decrease in the number of trees was attributed to the transformation of mining, putting other plantations in the coconut industry at risk. Another risk identified was operational costs, which were not balanced with the selling price.

Hebbar et al. [19] developed a MaxEnt model for evaluating bioclimatic variables. In order to identify the ideal habitat for coconut production and locations with climates that would be viable given future climate circumstances, the MaxEnt model was used to evaluate bioclimatic parameters. Because of the created MaxEnt model’s high prediction accuracy, it was frequently used for species dispersion. Species dispersion refers to the spatial distribution or arrangement of individuals within a species population or community. It describes how individuals are spread out or dispersed across a landscape, region, or habitat. However, the results were generalized from independent data due to spatial bias and overfitting.

Karunakaran and Narmadha [20] examined the growth and variability of the coconut sector in the world context. Secondary data from 1990–1991 to 2020–2021 was obtained from the Faostat and ICC websites on area, coconut yield, and production. Coppock’s instability index was used to investigate the pattern of growth according to the compound growth rate and calculate the instability. The area role and production yield were examined using decomposition analysis. The study found poor management practices, dangerous diseases, and pests were causing the negative growth.

Madeshwaran et al. [21] presented four different models, including LASSO, Ridge, Artificial Neural Networks (ANNs), and Elastic Net (ELNET) regression, which were used to determine the best model for predicting coconut yields. This yield estimation utilized external factors and weather parameters of the Coimbatore district. The ELNET was better when compared to the other approaches as it achieved an effective regularization of the statistical model.

The yearly output of data and monthly weather data for several districts on India’s west coast were produced by Das et al. [22] between 2000 and 2015. The weather indices were produced by adding together the various monthly average characteristics, including, relative humidity, solar radiation maximum and lowest temperatures, rainfall, and wind speed. Using the produced data as input, a variety of models were employed to simulate the coconut yield, including ANN, ELNET, LASSO, Stepwise Multiple Linear Regression (SMLR), Principal Component Analysis (PCA) with SMLR, and PCA-ANN.

2.2. Different Types of Crop Yield Prediction

Joshua et al. [23] described many machine learning techniques including backpropagation neural networks, radial basis functional neural networks, support vector regression, and general regression neural networks (GRNN) to estimate the paddy yield. The eastern part of Tamil Nadu, South India, which is home to the Cauvery Delta Zone (also known as the CDZ), was the target of this work’s rice yield calculation. Utilizing GRNN’s forward data propagation, the results of the prediction were improved. Pre-processing and redundant data removal should be included to further improve the prediction.

In order to estimate agricultural productivity, Elavarasan et al. [24] developed a hybrid feature extraction technique that paired a Random Forest Recursive Feature Elimination (RFRFE) wrapper with a correlation-based filter (CFS). A variety of machine learning techniques, including random forest, gradient boosting, and decision trees, were used to analyze the generated CFS-RFRFE. The non-essential and superfluous features were removed by designing the filter approach using the correlation. The developed CFS-RFRFE only considered the correlation measure during the feature extraction.

To calculate the productivity of agriculture, Elavarasan and Vincent [25] introduced a hybrid regression-based system based on reinforcement learning and random forests. This hybrid regression was used in order to apply reinforcement learning to each choice of splitting characteristics between tree buildings. Variable significance measure was examined in order to choose significant variables for the node splitting process during the model’s development and to facilitate efficient use of training data. By using internal cross-validation, the problem of overfitting was avoided and minimal parameter adjustment was achieved. However, the data was not divided using the proposed hybrid technique in a way that would have had an impact on prediction results.

To forecast agricultural output, Iniyan and Jebakumar [26] created an improved ensemble regression based on mutual information. The main statistical method for identifying feature association between the datasets was the mutual information-based feature selection. Reducing the amount of data input was necessary to assist the classification or regression problem. The data that varied over time was required to be considered for an effective prediction of the crop.

From the overall literature review, the following issues are highlighted: ineffective data incorporated into the prediction, a failure to analyze the complex and non-linear dependencies, and overfitting risk. It is difficult to calculate how several climatic elements, such as temperature, humidity, wind, and precipitation, interact with one another and affect crop yields. These requirements can be met by a deep learning model, which enables data-driven agricultural decision-making by producing precise and reliable crop predictions. In order to provide effective coconut prediction, this study implemented a deep learning method, BILSTM, with integrated LFSOA. To overcome the aforementioned issues, the following solutions are provided by the proposed method. For processing effective data to perform reliable prediction, the IoT sensors are used to obtain the data from the field. The information utilization from both the backward and forward directions is used to capture complex and non-linear dependencies for enhancing prediction. Finally, the tuning of hyperparameters using the LFSOA helps to avoid overfitting risk. Along with the hyperparameters, a penalty term to the loss function is added to restrain the large weights which supports to prevention of overfitting.

3. Background and Methods

In this BILSTM-LFSOA, an effective coconut yield prediction is ensured by using the data monitored by IoT devices in the field. The continuous data monitored using IoT devices are processed by using min–max normalization and redundant features are eliminated using LASSO feature selection. Further, the prediction results are improved using the BILSTM where the hyperparameters are optimized by using the LFSOA. Figure 1 shows the block diagram of BILSTM-LFSOA.

3.1. Data Acquisition and Processing

3.1.1. Dataset Acquisition

The data to perform the coconut yield estimation was acquired from the Trivandrum region. Here, every 15 min, the yield environment was monitored by IoT sensors, and the data was transformed into a daily variable for study. Here, the daily variables were collected from the year 2011 to 2021 (10 years of data). The following sensors were used in the monitoring: temperature sensors, humidity sensors, soil moisture sensors, rainfall sensors, wind speed, and direction sensors, solar radiation sensors, soil pH sensors, and nutrient sensors. The variables that they measured provide the input features of the model.

• The temperature sensor helped measure the existence of heat energy present in the soil; thus, the minimum and maximum temperatures were taken as 20.7 °C and 21.1 °C, respectively.
• The humidity sensor frequently monitored the humidity, which confirmed that the crop persisted in the perfect range (70% average).
• The soil moisture sensors measured the water content present in the soil (average 370 to 600).
• The rainfall sensor was used to determine whether there was a risk of flooding based on the rainfall range in various regions, which was taken as 7 mm on average.
• The wind speed and direction sensor measured the speed and direction of the wind in the crop field, which was taken as 2.4 m/s on average.
• The solar radiation sensors measured broadband and flux density irradiance with an average of 450, 600, and 470 W/m².
• The soil pH sensors measured the pH of the soil, which was essential since it affected the accessibility of plant nutrients with an average of 7 pH.
• The nutrient sensors precisely measured the nutrient levels in crop fields with an average of 1.67.

The data generation is depicted in the following Figure 2. The overall process of the proposed model is mentioned as follows,

Stage 1: Initially, the agricultural field was monitored using IoT devices (sensors) such as temperature, humidity, wind speed, rainfall, precipitation, and solar radiation.

Stage 2: Depending on the obtained data from these IoT devices, the datasets were collected. In-depth information about the collected data is given in Section 3.1.1.

Stage 3: The collected environmental data was pre-processed using min–max normalization, then the redundant features were removed in the feature selection through LASSO.

Stage 4: Moreover, the prediction outcomes were improved by BILSTM where the hyperparameters were optimized by means of the LFSOA. These performances were measured in terms of recall, accuracy, f1-score, and precision.

3.1.2. Data Pre-Processing

In the pre-processing, normalization was used for converting the values into the normal form without changing the differences in the value’s range. This stage ensures the data quality and reliability of abnormal data with the help of the normalization process. In order to modify the lowest value to 0 and the highest value to 1, as well as to convert the remaining value to decimal, the min–max normalization [10] was taken into consideration. Equation (1) shows the process of min–max normalization.

$$Inpu{t}_{new}={\displaystyle \frac{Inpu{t}_{old}-Inpu{t}_{min}}{Inpu{t}_{max}-Inpu{t}_{min}}}$$

(1)

where input minimum and maximum values are indicated as $Inpu{t}_{min}$ and $Inpu{t}_{max}$, respectively; and input and preprocessed values are designated as $Inpu{t}_{old}$ and $Inpu{t}_{new}$, respectively. The pre-processed outcome is given as input to the LASSO feature selection to avoid redundant features.

3.1.3. Feature Selection Using LASSO

LASSO [27,28] is a fundamental process that was utilized for eliminating insignificant features. Regularization and feature selection were the two main LASSO processes that were used while using Ordinary Least Square (OLS) regression to minimize the other sum of squares. A restriction of overall absolute parameter values was incorporated in the minimization process. Using the minimization function presented in Equation (2), the model coefficient β is determined.

$$LASSO\left({\beta }_i,{\beta }_0\right)=\underset{\beta }{\mathrm{argmin}}\sum _{i=1}^n{\left(R{V}_i-\left({\beta }_{i}Inpu{t_{new}}_i+{\beta }_0\right)\right)}^2+\alpha \sum _{j=1}^k\left|{\beta }_j\right|$$

(2)

where the feature data is denoted as $Inpu{t_{new}}_i$; the coefficient of the feature I is denoted as ${\beta }_i$; the regularization parameter is denoted as $\alpha$; $R{V}_i$ denotes the response vector; and $n$ is the number of features. The smaller number of features were acquired with larger values of $\alpha$ which made the coefficients controlled to be zero.

3.2. Yield Prediction

3.2.1. Classification Using BILSTM

The BILSTM [29,30,31] used the chosen LASSO characteristics as input to forecast the coconut production for an effective decision-making purpose [32]. In general, the LSTM [33] is an improved version of a recurrent neural network with a similar design. The data is moved from one stage to the next using the RNN and LSTM. Long-term dependencies are where LSTM achieves a high degree of success. However, an LSTM can only categorize the output according to the prior data. Consequently, if the forward data is not taken into account, there is a chance that a single LSTM will result in misclassification. As a result, to improve the categorization, the BILSTM takes into account both historical and future data. As shown in Figure 3, the BILSTM was made up of two LSTM models that functioned simultaneously, whereas Figure 4 depicts the operation of the BILSTM’s single LSTM unit. In the overall architecture, BILSTM trained two LSTM models simultaneously, one trained on the data moving from the first time period to the most recent time period (also called the forward direction), and one from the most recent time period to the first time period (also called the backward direction). Following the completion of the models’ training, their hidden states were combined. Thus, the BILSTM handled both the incoming and outgoing data.

The old data may be retained by the LSTM’s hidden layers for a brief period of time. Memory cell $C_t$, which is updated in response to the input gate $\left(i_t\right)$ and forget gate $\left(f_t\right)$, is a crucial component of the LSTM.

The $i_t$ is used to decide the data that is required to be kept in the memory cell, while $f_t$ is required to decide the memory that needs to be discarded from the memory cell. Equations (3)–(8) are used to update the $C_t$ of forward LSTM.

$$u_t^{\left(f\right)}=\tanh \left(w_{xu}^{\left(f\right)}{x}_t+w_{hu}^{\left(f\right)}{h}_{t-1}+b_u^{\left(f\right)}\right)$$

(3)

$$i_t^{\left(f\right)}=\sigma \left(w_{xi}^{\left(f\right)}{x}_t+w_{hi}^{\left(f\right)}{h}_{t-1}+b_i^{\left(f\right)}\right)$$

(4)

$$f_t^{\left(f\right)}=\sigma \left(w_{xf}^{\left(f\right)}{x}_t+w_{hf}^{\left(f\right)}{h}_{t-1}+b_f^{\left(f\right)}\right)$$

(5)

$$C_t^{\left(f\right)}=f_t^{\left(f\right)}⨀C_{t-1}^{\left(f\right)}+i_t^{\left(f\right)}⨀u_t^{\left(f\right)}$$

(6)

$$O_t^{\left(f\right)}=\sigma \left(w_{xo}^{\left(f\right)}{x}_t+w_{ho}^{\left(f\right)}{h}_{t-1}+b_o^{\left(f\right)}\right)$$

(7)

$$f{h}_t=O_t^{\left(f\right)}⨀{\tanh (C}_t^{\left(f\right)})$$

(8)

Equations (9)–(14) show the update of $C_t$ for backward LSTM.

$$u_t^{\left(b\right)}=\tanh \left(w_{xu}^{\left(b\right)}{x}_t+w_{hu}^{\left(b\right)}{h}_{t+1}+b_u^{\left(b\right)}\right)$$

(9)

$$i_t^{\left(b\right)}=\sigma (w_{xi}^{\left(b\right)}{x}_t+w_{hi}^{\left(b\right)}{h}_{t+1}+b_i^{\left(b\right)})$$

(10)

$$f_t^{\left(b\right)}=\sigma (w_{xf}^{\left(b\right)}{x}_t+w_{hf}^{\left(b\right)}{h}_{t+1}+b_f^{\left(b\right)})$$

(11)

$$C_t^{\left(b\right)}=f_t^{\left(b\right)}⨀C_{t-1}^{\left(b\right)}+i_t^{\left(b\right)}⨀u_t^{\left(b\right)}$$

(12)

$$O_t^{\left(b\right)}=\sigma (w_{xo}^{\left(b\right)}{x}_t+w_{ho}^{\left(b\right)}{h}_{t+1}+b_o^{\left(b\right)})$$

(13)

$$b{h}_t=O_t^{\left(b\right)}⨀{\tanh (C}_t^{\left(b\right)})$$

(14)

where $x_t$ is the input, ${fh}_t$ t and ${bh}_t$ are the output of forward and backward LSTM, and $w_{xi}, b_i, w_{xu}, w_{hu}, w_{xo}, b_o, w_{xf}$, and $b_f$ are the parameters that need to be learned in BILSTM. Combining the forward and backward LSTM results yields the final output of BILSTM $\left(H_T\right)$, as shown in Equation (15).

$$H_T=w_x^{\left(h\right)}f{h}_t+w_h^{\left(h\right)}b{h}_t+b^{\left(h\right)}$$

(15)

Hence, the BILSTM uses the past and future background for a better classification. This model measured the divergence between expected and actual crop yields using the loss function. The trained model predicted crop yields by using environmental variables, such as temperature, humidity, and precipitation. By employing this strategy, the model had the ability to comprehend complex relationships among crop yields and variables related to the environment, facilitating precise forecasting.

3.2.2. LFSOA-Based Hyperparameter Tuning for BILSTM

Using the LFSOA, the hyperparameters of BILSTM were optimized to provide a reliable crop yield forecast that helped to avoid the overfitting issue. The SOA was converted into the LFSOA by incorporating the Lévy flight, which was used to avoid premature convergence. LF is a mechanism for random walk that utilizes a long-tailed distribution of distances for dispersion. This results in a high probability of short jumps and a low probability of large jumps during the random walk. Lévy flight is used to supplement the SOA model as an examination of the SOA model shows that the model is prone to reducing the diversity of search agents in the early stages of the attack phase of the model. This results in a weaker exploration ability of the model. The inclusion of Lévy flight allows for the inclusion of greater randomness within the model. The lowest and maximum settings of Max-Epoch, L2Regularization, Dropout, and Learning Rate were used to initialize the solutions in this LFSOA. The ranges for L2Regularization, Dropout, and Learning Rate were $[0.003,0.1]$, $[0.1,0.4]$ and $[0.003, 0.1]$, respectively, while the options for Max-Epochs were $[5, 10, 15, 20]$.

To determine the ideal set of hyperparameters, the LFSOA was fed these randomly produced solutions. The random walk mechanism was provided by the LF which led to achieving the appropriate control over the local search. Therefore, the LFSOA was used with an effective local search to optimize the hyperparameters. The conventional SOA [34] is a modern swarm optimization approach that mimics the migration and attacking behavior of seagulls. This migration is said to be a seasonal behavior of seagulls that move to different locations in search of a wide variety of food that would provide them with enough energy. This process is displayed as follows:

• Migration is initialized by moving a seagull swarm. Initial locations are dissimilar from each other to avoid collisions.
• The gulls use their collective knowledge to their advantage, traveling in the direction of maximum survival to obtain the lowest possible cost.

In general, seagulls attack birds traveling over the sea, using a spiral pattern in their attack process. The overall process of SOA is explained as follows.

Exploration (Migration)

The mathematical equations of migration and prey attacking are deliberated here. The exploration phase mimics the seagull swarm traveling towards the locations. The following three conditions, such as avoiding collision, moving toward the best neighbor, and near-optimal search agent are required to be satisfied.

• Collision avoidance: An extra variable A is defined for eliminating the collision between the neighbor seagulls. This helps to update the new location of certain seagulls (search agent) as per Equation (16).

$${\overrightarrow{P}}_N=A\times {\overrightarrow{P}}_c\left(T\right),T=\mathrm{0,1},2,\dots ,Max(T)$$

(16)

where the location which avoids the other search agent is denoted as ${\overrightarrow{P}}_N$; the location of the candidate in the current iteration is denoted as ${\overrightarrow{P}}_c\left(T\right)$; and the search agent’s motion action at the search space is defined by using the A expressed in Equation (17).

$$A=FC-\left(T\times \left({\displaystyle \frac{FC}{\mathit{Max}\left(T\right)}}\right)\right)$$

(17)

where the iteration is denoted as $T$, and the frequency control of $A$ in the period $[0,FC]$ is denoted as $FC$.

• Movement towards best neighbor’s direction: Following their avoidance of neighboring collisions, the candidates attempt to travel in the direction of the optimal solution, as shown by Equation (18).

$${\overrightarrow{d}}_e=B\times ({\overrightarrow{P}}_b\left(T\right)-{\overrightarrow{P}}_c\left(T\right))$$

(18)

where the candidate location moves in the direction of the optimum solution ${\overrightarrow{P}}_b\left(T\right)$ is denoted as ${\overrightarrow{d}}_e$. Further, the balance between exploitation and exploration is achieved by using the coefficient B expressed in Equation (19).

$$B=2\times A^2\times R$$

(19)

where R represents the random value inside [0, 1].

• Remain close to the best search agent: Further, the search agents modify the location according to the optimum solution using Equation (20).

$${\overrightarrow{D}}_e=\left|{\overrightarrow{P}}_N-{\overrightarrow{d}}_e\right|$$

(20)

Exploitation (Attacking)

The seagulls vary their angle and speed in the migration process. Seagulls fly in the air using their wings and weight. The seagulls move in a spiral way at the $q, r$, and $s$ planes in the attack process, as shown in Equations (21)–(23).

$$\widehat{q}=rad\times \cos \left(RV\right)$$

(21)

$$\widehat{r}=rad\times \sin \left(RV\right)$$

(22)

$$\widehat{s}=rad\times RV$$

(23)

where in Equation (24), the radius of the spiral turn is represented as rad, and the random value between $[0, 2\pi ]$ is indicated as $RV$.

$$rad=\delta \times e^{\gamma RV}$$

(24)

where the logarithm base is denoted as e, and $\delta$ and $\gamma$ denotes the spiral shape. Equation (25) shows the new location of seagulls.

$${\overrightarrow{P}}_c\left(T\right)=\left({\overrightarrow{D}}_e\times \widehat{q}\times \widehat{r}\times \widehat{s}\right)+{\overrightarrow{P}}_b\left(T\right)$$

(25)

where the optimum result is denoted as ${\overrightarrow{P}}_c\left(T\right)$.

Enhancement Using Lévy Flight

The random walk approach was provided by LF for controlling the local search. This random walk is expressed in Equations (26)–(28).

$$Le\left(w\right)\approx w^{-1-\tau }$$

(26)

$$w={\displaystyle \frac{A}{{\left|B\right|}^{1/\tau }}}$$

(27)

$$\sigma ={\left\{{\displaystyle \frac{\mathsf{\Gamma }(1+\tau )}{\tau \mathsf{\Gamma }\left(\right(1+\tau )/2)}} {\displaystyle \frac{\sin (\pi \tau /2)}{2^{(1+\tau )/2}}}\right\}}^{2/\tau }$$

(28)

where $0<\tau \le 2;A~N(0,{\sigma }^2)$;$B~N(0,{\sigma }^2)$; The gamma function is denoted as $\mathsf{\Gamma }(.)$; the step size is denoted as w; the Lévy index is $\tau$; and $A/B~N(0,{\sigma }^2)$ defines the samples created using Gaussian distribution where the mean is 0 and the variance is ${\sigma }^2$.

A new enhanced portion according to the above-defined mechanism to update the SOA’s solution is shown in Equation (29).

$${\overrightarrow{D}}_{el}=\left\{\begin{array}{l}{\overrightarrow{D}}_{el}, F\left({\overrightarrow{D}}_{el}\right)>F\left({\overrightarrow{D}}_e\right)\\ {\overrightarrow{D}}_e, Otherwise\end{array}\right.$$

(29)

Better exploration, enhanced exploitation, and balanced search are just a few of the advantages that result from the integration of LFA and SOA into an effective optimization algorithm. Several optimization issues have been solved with the LFSOA, proving its effectiveness in producing superior results. The LFSOA was executed until it reached the maximum iteration. The flowchart for the LFSOA is shown in Figure 5.

The hyperparameters optimized using the LFSOA were used in the BILSTM for an effective coconut yield prediction. Moreover, the unwanted features were eliminated using the LASSO, which also helps to minimize the error rate and enhances the classification.

The statistical measures comprised the MAE, MSE, RMSE, and coefficient of determination (R²), which are expressed in Equations (30)–(33).

$$R^2={\left\{{\displaystyle \frac{1}{N}}\times {\displaystyle \frac{{\sum }_{i=1}^N\left(U_i-\overline{U}\right)\times \left(V_i-\overline{V}\right)}{{\left({\sigma }_U-{\sigma }_V\right)}^2}}\right\}}^2$$

(30)

$$MSE={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\left(V_i-\overline{V}\right)}^2$$

(31)

$$RMSE=\sqrt{MSE}$$

(32)

$$MAE={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N\left|V_i-\overline{V}\right|$$

(33)

where $U$ and $V$ are the observed value of features and dataset class; the number of observations is denoted as $N$; $U_i$ and $V_i$ are the observed value of features and dataset class at $i^{th}$ observation, and $\overline{U}$ and $\overline{V}$ are the mean value of $U$ and $V$, respectively. Equations (34)–(37) represent the recall, accuracy, F1-score, and precision of the classification measures.

$$Accuracy={\displaystyle \frac{TP+TN}{TN+TP+FN+FP}}\times 100$$

(34)

$$Precision={\displaystyle \frac{TP}{TP+FP}}\times 100$$

(35)

$$Recall={\displaystyle \frac{TP}{TP+FN}}\times 100$$

(36)

$$F1-measure={\displaystyle \frac{2Precision\times Recall}{Precision+Recall}}\times 100$$

(37)

where the symbols for false positive, false negative, true positive, and true negative are, respectively, TP, TN, FP, and FN.

4. Results and Discussion

This section presents the results of the BILSTM-LFSOA. The design and simulation of the BILSTM-LFSOA method were carried out using Python 3 and Anaconda Navigator V 2.3.1. The system used to perform the yield prediction was configured with 8 GB RAM and an i5 processor with 2.2 GHz. The dataset generated for evaluating this research was divided into 80% for training and 20% for testing. Input features were temperature, humidity, soil moisture, rainfall, wind speed and direction, solar radiation, CO₂, soil pH, and soil nutrients. Similarly, the proposed BILSTM-LFSOA was compared against baseline models such as RNN, RFC, and LSTM in terms of various performance metrics.

4.1. Performance Analysis of BILSTM-LFSOA

The performance analysis of the proposed BILSTM-LFSOA was analyzed using various classifiers and optimization algorithms such as statistical measures (MAE, MSE, RMSE, and (R²)) and classification measures (recall, accuracy, F1-score, and precision) are explained in this section. To analyze the LFSOA, many optimization-based hyperparameter tuning techniques were taken into consideration, including Grey Wolf Optimization (GWO) [35], the Whale Optimization Algorithm (WOA) [36], and SOA [34]. In this analysis, the GWO [35], WOA [36], and SOA [34] were developed for hyperparameter optimization for coconut yield prediction in order to analyze the effectiveness of the LFSOA. Figure 6 displays the fitness graph for the LFSOA with GWO [35], WOA [36], and SOA [34]. Figure 7 shows the comparison of the LFSOA with GWO [35], WOA [36], and SOA [34] for different population sizes. From this analysis, it is found that an LFSOA with a population size of 40 provides a better performance.

BILSTM-LFSOA’s performance with various classifiers and optimization strategies for hyperparameter optimization is examined. Random Forest Classifier (RFC), Recurrent Neural Network (RNN), and LSTM are the three classifiers that are compared against the BILSTM. The analysis of statistical and prediction measures for classifiers with default hyperparameters is shown in

Table 1. Analysis of statistical measures for classifiers with default hyperparameters.

Classifiers	MAE	MSE	RMSE	R2
RNN	0.412	0.443	0.666	0.618
RFC	0.304	0.399	0.632	0.697
LSTM	0.278	0.281	0.530	0.770
BILSTM	0.133	0.108	0.329	0.857

and

Table 2. Analysis of prediction measures for classifiers with default hyperparameters.

Classifiers	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
RNN	90.390	89.879	89.704	90.205
RFC	92.756	92.166	93.999	92.676
LSTM	94.502	93.908	94.516	94.327
BILSTM	96.467	97.627	97.172	96.135

. According to this investigation, the BILSTM outperforms the RNN, RFC, and LSTM in performance. For instance, the accuracy of BILSTM with default hyperparameters is 96.467%, whereas that of RNN, RFC, and LSTM is 90.390%, 92.756%, and 94.502%, respectively. In comparison to previous classifiers, the crop yield estimate is improved by using both historical and future data throughout the prediction process.

Table 3. Analysis of statistical measures for classifiers with optimized hyperparameters.

Classifiers	MAE	MSE	RMSE	R2
RNN	0.326	0.408	0.468	0.697
RFC	0.230	0.363	0.399	0.768
LSTM	0.202	0.260	0.348	0.812
BILSTM	0.094	0.011	0.105	0.954

and

Table 4. Analysis of prediction measures for classifiers with optimized hyperparameters.

Classifiers	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
RNN	93.431	93.716	93.358	93.842
RFC	96.390	96.176	96.941	96.872
LSTM	97.137	98.231	97.205	97.061
BILSTM	98.963	99.026	99.155	98.837

show the statistical and prediction measures of classifiers with optimized hyperparameters. This investigation demonstrates that compared to RNN, RFC, and LSTM, the BILSTM with optimized hyperparameters performs better. For example, the accuracy of BILSTM with optimized hyperparameters is 98.963%, whereas the RNN has an accuracy of 93.431%, RFC has 96.390%, and LSTM has 97.137%. Furthermore, compared to the one with default hyperparameters, the BILSTM with optimized hyperparameters performs better. To minimize the loss function, the ideal collection of hyperparameters discovered by the LFSOA is being used.

The analysis of statistical and prediction measures for different optimizations is shown in

Table 5. Analysis of statistical measures for different optimization approaches.

Classifiers	MAE	MSE	RMSE	R2
GWO	0.327	0.219	0.639	0.693
WOA	0.252	0.159	0.602	0.716
SOA	0.246	0.121	0.510	0.770
LFSOA	0.094	0.011	0.105	0.954

and

Table 6. Analysis of prediction measures for different optimization approaches.

Classifiers	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
GWO	90.570	90.595	90.600	91.754
WOA	93.623	93.881	93.022	93.260
SOA	96.046	95.402	96.817	95.384
LFSOA	98.963	99.026	99.155	98.837

. This investigation demonstrates that the LFSOA outperforms the GWO [35], WOA [36], and SOA [34]. For example, the accuracy of the LFSOA with default hyperparameters is 98.963%, whereas the GWO [35] is 90.570%, WOA [36] is 93.623%, and SOA [34] is 96.046%. The incorporation of LF in SOA is used to handle the local search which helps to achieve a more optimum set of hyperparameters than the other methods. Moreover, the tuning of hyperparameters avoids the overfitting issue and additionally helps to enhance the estimation performances.

The analysis of prediction results using BILSTM with/without the LFSOA is shown in

Table 7. Analysis of prediction results using BILSTM with/without LFSOA.

Classifiers	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
BILSTM	96.064	95.928	95.911	95.758
BILSTM with LFSOA	98.963	99.026	99.155	98.837

. From Table 7, it can be clearly seen that BILSTM with LFSOA has achieved superior results in terms of accuracy (98.963%), precision (99.026%), recall (99.155%), and f1-score (95.758%). Meanwhile, BILSTM without using LFSOA has accuracy, precision, recall, and f1-score of 96.064%, 95.928%, 95.911%, and 95.758%, respectively. Utilizing the LFSOA grants the coconut yield prediction model greater accuracy, faster convergence, and increased resistance to noise, which results in more accurate and dependable predictions.

4.2. Discussion

This research developed an enhanced deep learning classifier name called BILSTM with the LFSOA to effectively predict coconut production. From the overall analysis, the proposed BILSTM with the LFSOA is evaluated against the different optimization approaches such as GWO [35], WOA [36], and SOA [34] to evaluate its efficiency. Furthermore, the classification by BILSTM-LFSOA is analyzed against some other classifiers such as RNN, RFC, and LSTM. In the results, the LFSOA achieved an enhanced prediction accuracy of 98.963% which was better than the existing SOA [34], GWO [35], and WOA [36] which obtained 96.046%, 90.570%, and 93.623%, respectively. Similarly, the BILSTM with LFSOA achieved an enhanced prediction accuracy of 98.963% which is superior to existing RNN, RFC, and LSTM which obtained 93.431%, 96.390%, and 97.137%, respectively. Overall, the proposed BILSTM with LFSOA achieved higher results in all the performance metrics when compared with the conventional methods. The incorporation of both the backward and forward directions in BILSTM helps obtain complex and non-linear dependencies that lead to enhanced prediction performances. Even though BILSTM is a much slower model since it needs additional time for the training process, the incorporation of the LFSOA reduces the training time.

The BILSTM model has some other limitations. Firstly, the model requires an adequately large training dataset to fetch input data from. In the case that the training dataset is not sufficiently large, a traditional LSTM model tends to outperform the BILSTM model in terms of loss [37]. Secondly, LSTM is faster in reaching its equilibrium loss values in comparison to BILSTM, which is likely attributable to the latter model requiring to be trained twice (once in the forward direction, and once in the backward direction) as compared to the LSTM which is only trained once. Finally, while both LSTM and BILSTM store prior sequences of data to enhance the accuracy of predictions, the size of the sequences that can be stored is greater in LSTM than in BILSTM due to the model training two LSTM models simultaneously, reducing the size of the sequences to half the size of sequences LSTM can store [37].

It is noted that, while the proposed model exceeds the performance of all other benchmark models, BILSTM on its own has lower accuracy (96.064%), precision (95.928%), recall (95.911%), and f1-score (95.758%) in comparison to both the RFC and LSTM models. This is likely to be attributable to the aforementioned limitations of the BILSTM model wherein it has both lower memory for stored sequences as compared to the traditional LSTM model, and requires a larger training dataset in order to tune its parameters.

Given that LFSOA-based hyperparameter tuning results in BILSTM outperforming the tested models, it is likely that the reason for the difference in scores is attributable to insufficient training data to tune parameters. This indicates that the inclusion of the LFSOA may shorten the training time of the BILSTM model.

Another factor to consider is that the BILSTM model is a “black box”—it is not completely understood how it arrives at the specific values that it does, which can make it difficult to properly explain to small stakeholders the crucial factors that result in the particular decision that is reached [38]. BILSTM especially suffers from this problem, as it incorporates hidden factors that may not be evident in the datasets but become essential to accurate forecasting. This can impact the perceived trustworthiness of the predictions.

Rural adoption of smart technologies is also a critical factor that can hinder adoption of the model. Rural areas present a challenge to the successful implementation of the proposed model for a variety of reasons. Simply put, there is a lack of sufficient research to accurately predict rural responses [39]. What research has been conducted on rural adoption of technologies is mostly isolated to Western rural areas, with a relative dearth of research on the adoption of technologies in rural areas of the developing world. Additionally, “rural areas” compose a broad spectrum of attitudes and cultures, even within the same country, making it difficult to generalize any one response across every rural community [39].

Given the aforementioned difficulties in explaining how BILSTM-LFSOA achieves its results, a careful approach will be necessitated to ensure cooperation from rural communities. The advantages of the model would need to be thoroughly explained to key local stakeholders, and a robust and holistic strategy for implementation of the model that caters to the specific features of the community would need to be devised. Local expertise will be a key element of successful adoption, as local experts can provide information on local perspectives and needs.

Finally, we conclude this discussion by noting the potential opportunities that smart farming affords to local communities. The integration of data-driven forecasting into traditional agriculture will be essential in the coming years as worldwide arable land shrinks while populations continue to grow [40]. Machine learning and data-driven techniques can help reduce the uncertainties that farmers face in decision-making and reduce the need for costly and time-consuming methods [40,41]. Successful adoption of IoT-based sensors can allow for real-time monitoring of crop conditions, while the usage of drones and unarmed vehicles enables better modeling of crop conditions. Robotics in agriculture have also increased productivity by automating the watering and spraying processes [40].

All these elements contribute to increasing the financial security of farmers and their incomes, resulting in greater integration into the agricultural value chain [40].

5. Conclusions

In this research, the hyperparameters for BILSTM were optimized using the proposed LFSOA that helped to avoid the overfitting risk and effectively forecast the coconut yield. Also, the optimized features and optimized hyperparameter with BILSTM were used to enhance the crop yield prediction performances. The contributions of this study are as follows:

• The LFSOA model was compared against GWO, WOA, and SOA. The results indicate that the LFSOA outperforms these benchmark models in terms of accuracy, precision, recall, and f1-score. The LFSOA scored 98.963%, 99.026%, 99.155%, and 98.837% in these areas, respectively, which is higher than the next-highest score attained by SOA. SOA attains 96.046%, 95.402%, 96.817%, and 95.384% in the aforementioned metrics, respectively.
• The prediction results of BILSTM with and without LFSOA were analyzed, showing that BILSTM without LFSOA scores lower than BILSTM with LFSOA. BILSTM without LFSOA was shown to score 96.064%, 95.928%, 95.911%, and 95.758% in accuracy, precision, recall, and f1-score, respectively. Meanwhile, BILSTM-LFSOA scored 98.963%, 99.026%, 99.155%, and 98.837%, respectively, in each of these areas.
• These results were compared against RNN, RFC, and LSTM models. It was found that while BILSTM-LFSOA outperformed all the benchmark models, with the next-highest score being that of the LSTM model (97.137%, 98.231%, 97.205%, and 97.061%, respectively), both LSTM and RFC outperformed the simple BILSTM model. This is likely attributable to the size of the dataset. As noted in [37], while the BILSTM model has the potential to outperform unilateral LSTM models in time series forecasting, the model requires a sufficiently large dataset so as to properly tune its parameters. The inclusion of the LFSOA to tune hyperparameters was likely one reason for the higher score of the BILSTM-LFSOA model as compared to the simple BILSTM model.
• The superior predictive power of the BILSTM model when compared to benchmark models is likely to be attributable to the incorporation of forward and backward directions, which allows it to more accurately obtain long-term complex and non-linear dependencies in the dataset.

In the future, the BILSTM can be integrated with the attention mechanism for enhancing the coconut yield estimation by handling the long-term dependencies. One possible limitation of the model presented can be the reliance on sensors to gather accurate data readings, which can be costly to implement at a larger scale. Additionally, the model requires large amounts of data to create accurate estimations of yield outcomes. The limited geographical area in which the study was conducted also puts constraints on the generalization potential of the study. Finally, the variables considered within the scope of this study are not necessarily the only factors influencing crop yields.

Further, future research will focus on factors such as soil composition, pesticide use, and agricultural site characteristics to create a more complete production prediction model. However, future research could focus on alternate methods to gather data without the loss of accuracy in the dataset. Additionally, the BILSTM could be integrated with the attention mechanism for enhancing the coconut yield estimation by handling the long-term dependencies. Another avenue for future research could be to examine alternative methods of addressing the overfitting risk. While this study uses hyperparameter tuning to reduce overfitting risk, other potential techniques could be deployed to reduce overfitting, such as early stopping mechanisms. There is also potential to examine whether the inclusion of the LFSOA in the other benchmark models raises their performance to the point that it can challenge BILSTM. Finally, future research could incorporate other factors influencing crop yields in their study.