Artificial intelligence framework for modeling and predicting crop yield to enhance food security in Saudi Arabia

Datasets

Predicting crop productivity is a significant issue in agriculture. Producing high-quality food for human consumption heavily relies on a variety of factors, chief among them being the weather (such as temperature and rainfall), insecticides, and historical data on crop productivity. In the end, we all need the same fundamental essentials for survival. Corn, wheat, rice, and other basic crops make up the bulk of our diet. AI was used in this study to anticipate the consumed yields throughout Saudi Arabia. Ten of the most often grown crops were included. Regression was an issue we had to deal with. The datasets contained crop yields, temperature, insecticides, and rainfall. For the crop yields of maize, potatoes, rice, sorghum, and wheat, we collected data between 1994 and 2016. The dataset is available at the following link: https://www.kaggle.com/code/kushagranull/crop-yield-prediction/notebook.

Figure 2 shows the dataset after normalization, as well as the statistical metrics means and standard division that we calculated. It can be observed that the dataset after min-max normalization values are as follows: mean = 0.20973 and STD = 0.270; the Y-axis represents the scaling of data, and the sample identifications are presented on the X-axis.

Normalization method

Normalizing data with min-max normalization is an often-used practice. The minimum and maximum values of each feature are translated into zero and a decimal between 0 and 1, respectively, for each feature. (1)${\displaystyle z_n=\frac{x-x_{\min }}{x_{\max -x_{\min }}}\left({\mathrm{New}}_{{\max }_x}-{\mathrm{New}}_{{\min }_x}\right)+{\mathrm{New}}_{{\min }_x}.}$ The x_max and x_min are the maximum and minimum values, respectively. New (min_ x) is the smallest number, while New (max_x) is the largest number.

Proposed model

ANNs are a kind of AI that mimics the human brain (Gavahi, Abbaszadeh & Moradkhani, 2021). Densely coupled neurons, the network architecture, and the learning technique influence a neural network’s function. These are simulations of biological brain networks. The ANN technique may help in pattern identification and data classification (Alkahtani & Aldhyani, 2022; Mehedi et al., 2021; Alkahtani & Aldhyani, 2021b; Aldhyani & Alkahtani, 2022). The MLP model is the most often utilized ANN, notably in environmental research. This method may be used to match features and solve pattern recognition problems. MLP may also be used to categorize various linear patterns. These are feedforward neural networks (FNNs) that include several layers of units between the input and output layers. Examples of how a neuron’s output might be expressed are as follows:

(2)${\displaystyle \xi =\sum _{i=1}^n{w}_i{x}_i-b=w^{T}x-b}$ (3)${\displaystyle y=\sigma \left(\xi \right)}$ (4)${\displaystyle \sigma \left(\xi \right)=\frac{1}{1+{\mathrm{e}}^{-\left(\xi \right)}},}$

where xi is the number of the ith input, w_i isthe link weight from the ith input, w = (w₁…w_n)T is the total weight, where x_i is the number of the ith input (x₁…x_n). A threshold or bias is denoted by the letter b, while the number n indicates the total number of inputs. The job of the activation function s(x) is typically to transfer the real numbers into the interval, and this duty can be performed by a continuous or discontinuous function. It is also possible to utilize the sigmoidal activation function. It is possible to express it using the form (Alkahtani & Aldhyani, 2021a).

Training (learning), testing, and validation helped to improve the network design. The sum-of-squares error function was used to evaluate the performance of each neural network in the subsequent stages of model creation. The ANN training method involved an iterative adjustment of the strength of connections between neurons in adjacent layers and the parameters of activation functions. An attempt was made to reduce the training error (El) using the training data set (Chen, 2020). The test error (Et) was also determined for each iteration of the training process to evaluate its accuracy. Network overtraining can occur when E1s cease decreasing or when they decrease but Ets increase, which usually implies overtraining. This can be detected by looking for an increase in E1s and a decrease in Ets. The Levenberg–Marquardt function was used to train the neural networks.

Model performance

Different MLP neural network models and their practical appropriateness were evaluated using statistical criteria for prognostic model validation in this study. The models’ accuracy in terms of fitting was assessed using the coefficient of determination (R2). In this study, the normalized root mean square error (NRMSE), mean square error (MSE), and root mean square error (RMSE) were used to calculate the average absolute difference between forecasts and observations.

(5)${\displaystyle \mathrm{MSE}=\frac{1}{n}\sum _{i=1}^n{\left(y_{i,\exp }-y_{i,\mathrm{pred}}\right)}^2}$ (6)${\displaystyle \mathrm{RMSE}=\sqrt{\sum _{i=1}^n\frac{{\left(y_{i,\exp }-y_{i,\mathrm{pred}}\right)}^2}{n}}}$ (7)${\displaystyle R\text{\%}=\frac{n\left(\sum _{i=1}^n{y}_{i,\exp }\times y_{i,\mathrm{pred}}\right)-\left(\sum _{i=1}^n{y}_{i,\exp }\right)\left(\sum _{i=1}^n{y}_{i,\mathrm{pred}}\right)}{\sqrt{\left[n{\left(\sum _{i=1}^n{y}_{i,\exp }\right)}^2-{\left(\sum _{i=1}^n{y}_{i,\exp }\right)}^2\right]\left[n{\left(\sum _{i=1}^n{y}_{i,\mathrm{pred}}\right)}^2-{\left(\sum _{i=1}^n{y}_{i,\mathrm{pred}}\right)}^2\right]}}\times 100}$ (8)${\displaystyle R^2=1-\frac{\sum _{i=1}^n{\left(y_{i,\exp }-y_{i,\mathrm{pred}}\right)}^2}{\sum _{i=1}^n{\left(y_{i,\exp }-y_{\mathrm{avg},\exp }\right)}^2}}$ (9)${\displaystyle \mathrm{NRMSE}=\frac{\sqrt{\frac{1}{n}\sum _{k=1}^n{\left(y_{i,\exp }-y_{i,\mathrm{pred}}\right)}^2}}{y_{i,\mathrm{pred}}}.}$

In this case, the y_i,exp represents the experimental value of the data point i and y_i,pred is the predicted value of the data point i. The y_avg,exp is the represented average of the experimental values, and n is the total training values

Development of the MLP model

The MLP model contains an input layer, a hidden layer, and an output layer; the first two have 15 neurons, while the output layer contains one neuron. The input variables are represented by the number of neurons in the input layer, while the projected output value is represented by a single neuron in the output layer. It is utilized for cross-validation of the prediction model, which is further constituted of two hidden units and executes computational tasks. In training, one unit is employed; in validation, trainlm is a network training function that uses Levenberg–Marquardt optimization to update weight and bias values in the network. With a loss function that is the sum of squared errors, the Levenberg–Marquardt algorithm is best suited for this use.

Figure 3 depicts the MPL model for predicting crop yields in Saudi Arabia, while Table 1 lists the model’s parameter values.

Training process of the MLP model

Training is a key step in the development of a highly effective model based on some experimental data. Of the total datasets, 70% were used at this stage for this purpose. As can be seen in Fig. 4 and Table 2, the constructed MLP model performs admirably in terms of the evaluation metrics. The correlation between the prediction values and the crop yield parameters values is presented. The MLP model has high values of R (>100%) for predicting insecticide values and R2 (>0.9284) for predicting the crop yield, in addition to low values of MSE, RMSE, and NRMSE. These values show that the system has been optimized to meet the specified objectives.

The histogram inaccuracy of the predicted values at the training state is depicted in Fig. 5. To determine the amount of deviation that exists between the predicted values and the target values, the error histogram metrics were examined. Because these error values explain how the anticipated values differ from the target values, these values can be negative. Also, it specifies how the predicted values deviate from the target values. It was reported that the greatest errors were 0.000946, 0.000544, and 0.000258 for the crop yield, temperature, and insecticides, respectively.

Testing process of the MLP model

To verify the accuracy of the MLP model, the testing phase utilized 30% of the datasets, which consisted of previously undisclosed data. The results of the MLP model performance in the testing stage are shown in Table 3, respectively. As can be seen in Fig. 6, there is an outstanding agreement between the values that were predicted and the values that were sought to be found through experimentation. In addition, it was found that the values of R% were very high (100%), and the value of R² was very high (96.33%), for predicting insecticides while the values of MSE and RMSE were very low (0.0045 and 0.021 respectively) for predicting temperature values. These demonstrate that the MLP model that was created to forecast the crop yield, temperature and insecticides and it is solutions is both accurate and reliable.

Figure 7 shows the histogram inaccuracy of the MLP model during testing for forecasting crop yield. Histogram errors are metrics that are employed to determine the discrepancies between the observed and predicted data. The mean errors in the histograms were 0.000094, 0.00544, and 0.00025 for the crop yield, temperature, and insecticides, respectively.

Analysis of the important parameters for increasing the crop yield in Saud Arabia

Our economy and sustainable growth depend on accurate forecasts of agricultural production, which is why crop production forecasting has become a major concern. It aids farmers and the government in developing better post-harvest management in terms of transportation, storage, and distribution at the local, regional, and national levels. Agricultural production optimization and intensification face numerous challenges, one of which is predicting crop yields. Natural conditions can have a considerable impact on crop selection, crop rotations, applied agrotechnical approaches, and long-term land use planning. All these items are essential components of systems that assist farmers in making knowledgeable, expert decisions.

One of the most important aspects of the AI approach is the requirement for a large enough number of training examples based on high-quality observations of a complicated system. The precision of training in intelligent systems is closely related to the amount of information provided and the dependability of that information.

To estimate the effect of meteorological conditions of the previous year on the yield of the current year in Saudi Arabia, we employed a vector formed of the average monthly values of rainfall, temperatures, and insecticides. These three essential elements have a greater impact on Saudi Arabia’s crop harvest than others, but many factors affect the agricultural output in Saudi Arabia. Thus, long-term stationary experiments are utilized to determine them. Figure 8 displays the structure of an MLP model for determining the effects of rainfall, temperature, and insecticides on different crop yields, such as those for potatoes, rice, sorghum, and wheat.

We also used the MLP model to determine the relationship between crop yields and these parameters to formulate appropriate recommendations for agricultural technology for the upcoming year, taking into account the collected experience and weather conditions monitored over time. Hence, we have looked into the possibility of predicting crop yields using AI. We have applied the MLP model to determine the relationship between the temperature, insecticides, and rainfall amounts with different crop yields, such as potatoes, rice, sorghum, and wheat. Figure 9 shows the regression plot of the MLP model for finding the correlation between temperatures and crop yields It can be observed that temperature had more influence on the crop yields in Saudi Arabia. The MLP score was R > 98% for all the crops.

Figure 10 shows the regression between insecticides and crop yields for potatoes, rice, sorghum, and wheat. It shows that insecticides had the most influence on increasing crop yields in Saudi Arabia. The MLP model had the highest regression scores (between R > 90% and 99%) for crop yields in the testing and training processes.

Temperatures in Saudi Arabia (SA) can vary widely from place to place and depending on the time of year. Spring and winter have the largest incidences of rainfall, according to an analysis. The rainfall is an important resource for increasing the percentage of the predicted crop yield. Therefore, we have applied the MLP model to determine the relationship between rainfall and crop yields for rice, sorghum, and wheat. Figure 11 shows the regression graph, which shows that rainfall helps increase the crop yield. The scores were R > 91% for the testing phase and R > 98% for the training phase.

According to the findings of this research project, the effect of weather conditions from the previous year, such as rainfall, temperature, and the use of insecticides, on the yield of the current year is comparable to the overall effect of the agricultural practices that are implemented in Saudi Arabia. Therefore, to accurately predict crop yields for the subsequent agricultural period, it is essential to take into account not only the agricultural practices that are going to be utilized but also the anticipated temperature range, amount of precipitation, and composition of insecticides. Table 4 shows the MLP result against different existing food security systems.