Predicting rice diseases across diverse agro-meteorological conditions using an artificial intelligence approach

Dataset collection

The dataset was collected for Dapoli taluka of Ratnagiri district (Maharashtra, India) which is one of the high productivity districts for rice. The dataset is compiled from the data obtained from the online visual crossing weather platform, the Indian Meteorological Department, and All India Coordinated Research Project on Agrometeorology (AICRPAM). The dataset is composed of 1,634 rows of weekly weather data from the year 1989 to 2019 as per Indian Meteorological Department (IMD) with eight numerical input weather parameters and one numerical target variable. The parameters such as Temperature, Relative Humidity, Precipitation, Wind speed, etc. are comprised in the dataset. The dataset involves predicting the occurrence of five different classes of rice diseases based on the weather parameters in the district. To predict a particular value of a particular parameter the algorithm needs to learn from available historical data. The algorithm understands the correlation among different parameters in the dataset.

Dataset pre-processing

A data visualization Python library called Seaborn is used to find significant characteristics in the dataset. It uses the SNS.heatmap( ) function to build a heat map as an output. It shows a correlation matrix for different combinations of agro-meteorological variables in the dataset. The range of correlation values is 0 to 1, with the values above 0.3 and below −0.3 indicates a significant relationship. The data was pre-processed after the features were selected. There were alpha-numeric values in certain fields, such as the date field in the dataset so before processing, these data must be transformed to a single numeric value. The label encoder( ) function from the Pandas library was used for this. The dataset was having null values so to remove these null values, the function isnull( ) was invoked. The data are pre-processed through standardization, which changed the actual scale of the input values to an interval between 0 and 1.

Model adopted based on ANN using Keras for predicting weather parameter values

Initially, a sequential model by adding a sequence of layers is created. A fully-connected network structure with three layers is used. They are defined using the dense class. The number of neurons in the layer is the first argument and the activation function using the activation argument is the second argument. The next step is to select an activation function from sigmoid, ReLU, softmax and tanH. As the number of inputs is 8 parameters, the input layer and the output layer will have eight neurons for predicting weather parameter values. After creating the model, it is necessary to compile a model. In this step, it is necessary to specify the loss function to evaluate a set of weights and the optimizer to search through different weights for the network. Here, mean squared error is used as loss function and optimizer used is adam. The metrics used is MAE for prediction of values and accuracy, precision, F1 score for classification of rice diseases. The model is trained or fit on the loaded data by calling the fit( ) function on the model. Training occurs over epochs and each epoch is split into batches. The number of epochs considered here is 500 and batch size is 32. Lastly evaluation of the model is done. The evaluate( ) function is used to perform the same. It will return a list with two values. The first will be the loss of the model on the dataset and the second will be the accuracy of the model on the dataset.

Dataset splitting

The historical weather data obtained were divided into two sets, namely the training set and training set. Initially, the artificial intelligence (AI) model is built on a dataset called a training set and the built model is tested on a new set called the test set. The newly developed ML model is applied to test the dataset to measure the performance. This dataset was further split in two different ways. As per the bifurcation of the dataset, if 70–30% split is considered then the data values for the year field that are less than equal to 2010 and greater than equal to 1989 will be considered for training. The data values for the year field that are greater than 2010 and equal to 2019 will be considered for testing. If 80–20% split is considered, then the data values for the year field that are less than equal to 2013 and greater than equal to 1989 will be considered for training and the data values for the year field that are greater than 2013 and equal to 2019 will be considered for testing. Table 1, shows these two variations in dataset division.

Algorithm of proposed work

The proposed model is a combination of regression and classification. ANN model is used for implementation. Following are the algorithmic steps of the model.

Step 1: Collect week-wise data of agro-meteorological parameters for disease forecasting.

Step 2: Load data by using the pandas library

Step 3: Perform data pre-processing on the instances of the dataset.

Step 4: Build a model based on ANN using Keras.

Step 5: Divide the dataset into two parts: training dataset and test dataset. The dataset is split into two cases (70–30%, 80–30%) and results are verified accordingly.

Step 6: Training the network by utilizing the dataset.

Step 7: Prediction of future values of climatic parameters.

Step 8: Evaluation of the prediction model

Step 9: Prediction of crop disease occurrence. The final output will be in the form of five classes namely Healthy, Rice Blast, Blight, Brown Spot, and False Smut.

Step 10: Evaluation of the classification model.

Class 1—Healthy, Class 2—Rice Blast, Class 3—Bacterial Blight, Class 4—Brown Spot, Class 5—False Smut.

Artificial neural network

The biological nervous system influenced the creation of the ANN. It is made up of closely connected nodes known as neurons. It is a network made up of artificial neuron clusters that are linked together with different weights allocated to each connection. There are certain learning laws that determine the states of the neurons as well as the values of each relation. There are two types of learning namely supervised and unsupervised learning. Supervised learning is utilized in the proposed approach. The relationship between input and output variables is determined using ANN, a non-linear mathematical procedure. A neuron’s computational model can be defined as:

Initially, the n inputs are denoted as

(1) $$\mathrm{X}=x_1,x_2,......x_{\mathrm{n}-1},x_{\mathrm{n}}$$

These inputs are fed into a neural node, where they are multiplied by their weights, which are denoted by

(2) $$\mathrm{W}=w_1,w_2,.....w_{\mathrm{n}-1},w_{\mathrm{n}}$$

The results of these multiplications are then added together. The summed value is denoted by Sum variable.

(3) $$Sum=X\times W$$

(4) $$Sum=\sum _{i=1}^n{x}_i\times w_{\mathrm{i}}$$

After that, the sum value is transferred into an activation function denoted by φ. So output can be defined as:

(5) $$Y(output)=\text{φ}(\mathrm{S}\mathrm{u}\mathrm{m})$$

Along with the actual inputs to the node, an external bias also known as the threshold term denoted by θ, is applied to the node, which has the effect of decreasing or increasing the net input to the node, based on its sign. The bias can be modeled as a constant zeroth input $x_0$ = 1, and an associated weight $w_0$ = θ whose value can be negative, positive, or zero. Figure 2, represents the overview of the ANN model used in this paper.

The Back propagation algorithm is used in the proposed process, which operates on the concept of error correction learning law. The basic one-layer perceptron can only solve linearly separable problems, but it has trouble with linearly dependent scenarios. The partial derivative of the error is taken considering the corresponding weight of the network. This helps to understand how the error on the network is going. if we add the weight to the negative of this error, the error will minimize and shift towards the local minima. This algorithm is known as back propagation since it begins from the output layer, moves to the hidden layer, and then back to the input layer in a backward direction. There have been several weather-based models produced for different types of crops and different diseases. Potatoes, tomatoes, and grapes are the major crops for which the forecast model has been proposed in the literature. They are focused on multiple linear regression analysis. However, the ANN method has never been used to forecast multiple diseases in rice.

Activation function

It is an internal state of a neuron. It is used to convert the input signal on a node of ANN to an output signal. It is a weighted sum of input that becomes an input signal to the activation function to give one output signal. They introduce non-linear properties to the network. Therefore, it helps to solve complex problems that include images, videos, and audio. The accuracy of the model is majorly dependent on the activation function. Many types of activation functions are available. However, in the proposed model, four different AF are demonstrated namely sigmoid, hyperbolic Tangent (TanH), ReLU, and softmax. In the regression part, the activation function is applied to hidden layers as the model predicts the value of climatic variables. For classifying the rice diseases these classifiers are evaluated by applying at output layers.

Sigmoid

The sigmoid function can be represented as

(6) $$\mathbf{s}\mathbf{i}\mathbf{g}\mathbf{m}\mathbf{o}\mathbf{i}\mathbf{d}\left(\mathbf{x}\right)=\frac{\mathbf{1}}{\mathbf{1}+{\mathbf{e}}^{-\mathbf{x}}}$$

Figure 3, shows the graphical representation of the sigmoid activation function.

This function converts the input into a range from 0 to 1. The sigmoidal graph is having S-shaped curve. The product is any value, by applying a sigmoid function, the value comes in the range of 0 to 1. If the result is less than 0.5 then consider the value as 0 and if the result is more than 0.5 then consider the value as 1. If the value obtained is 0.5 then mark it as 0. The center of the curve lies at 0.5. The value 1 means the neuron is activated and is transferring the signal and is helping in classification.

(7) $$\mathbf{f}\left(\mathbf{x}\right)=\left\{\begin{array}{c}0\mathbf{i}\mathbf{f}\mathbf{x}<0\\ 1\mathbf{i}\mathbf{f}\mathbf{x}>0\\ 0.5\mathbf{i}\mathbf{f}\mathbf{x}=0\end{array}\right\}$$

This function eliminates outliers or extreme values in data without replacing them. It is used in models where there is a need to predict the probability of an output. It is best suited for multiclass classification.

Hyperbolic tangent (tanH)

Hyperbolic tangent is commonly termed as tanH function. It is similar to Sigmoid AF; however, tanH AF has better performance as the tanH AF ranges between −1 to 1. It is a nonlinear AF. It is centered at 0 so it is symmetric around the origin. Figure 4 represents the graph of tanH AF. It also has S-shaped curve similar to sigmoid AF.

tanH can be represented as follows:

(8) $$\tanh \left(\mathbf{x}\right)=\frac{{\mathbf{e}}^{\mathbf{x}}-{\mathbf{e}}^{-\mathbf{x}}}{{\mathbf{e}}^{\mathbf{x}}+{\mathbf{e}}^{-\mathbf{x}}}$$

It also faces a vanishing gradient issue similar to sigmoid AF. tanH activation function can be represented by the following mathematical model

(9) $$\mathbf{f}\left(\mathbf{x}\right)=\left\{\begin{array}{c}-1\mathbf{i}\mathbf{f}\mathbf{x}<0\\ 1\mathbf{i}\mathbf{f}\mathbf{x}>0\\ 0\mathbf{i}\mathbf{f}\mathbf{x}=0\end{array}\right\}$$

If the input is a negative value then it will map to -1, if the input values are positive then the output will be mapped to 1 and if the input is 0 then it will be mapped to zero itself.

ReLU

ReLU stands for Rectified Linear Unit. Figure 5 represents the graphical representation of the ReLU activation function. The graph has a linear nature.

ReLU activation function can be represented by the following mathematical model

(10) $$f\left(x\right)=\left\{\begin{array}{cc}x& if x\ge 0\\ 0& if x<0\end{array}\right\}$$

The above ReLU mathematical model says that if the input value is a negative value then the output value will be 0 and if the input value is a positive value then the output will be the inputted positive value itself. As a result, it essentially eliminates the negative aspect of the feature. It can also be represented as

(11) $$\mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}(\mathrm{y})=\mathrm{m}\mathrm{a}\mathrm{x}(\mathrm{y},0)$$where y is the product of the weight of the neuron and the input value.

If the value of y is negative, then the output is transformed to 0 and if the value of y is positive, then the output is a positive number itself. If the value of y is negative, then the maximum of a negative number and zero is zero and if the value of y is positive then the maximum of a positive number and 0 is a positive number itself. ReLU removes the vanishing gradient limitation experienced by tanH and sigmoid activation functions.

Softmax activation function

The output produced with this activation function ranges between 0 to 1. It generates a multiclass probability distribution as a vector over target classes and the summation of the probabilities equals 1. Softmax activation function can be represented as

(12) $$\mathbf{S}\mathbf{o}\mathbf{f}\mathbf{t}\mathbf{m}\mathbf{a}\mathbf{x}\left({\mathbf{x}}_{\mathbf{i}}\right)=\frac{\exp \left({\mathbf{x}}_{\mathbf{i}}\right)}{\sum _{\mathbf{j}}\exp \left({\mathbf{x}}_{\mathbf{j}}\right)}$$

The higher the value in the vector means higher is the probability of the output mapping to a particular class. It is typically used in multiclass classification and performs best if used in the output layer of the network. Table 2 summarizes the activation functions used in the research paper.

Table 2:

Summary of activation functions.

Activation function	Equation	Graph
Sigmoid	$f\left(x\right)=\left\{\begin{array}{cc}0& ifx<0\\ 1& ifx>0\\ 0.5& ifx=0\end{array}\right\}$
ReLU	$f\left(x\right)=\left\{\begin{array}{cc}x& ifx\ge 0\\ 0& ifx<0\end{array}\right\}$
Softmax	$\mathbf{S}\mathbf{o}\mathbf{f}\mathbf{t}\mathbf{m}\mathbf{a}\mathbf{x}\left({\mathbf{x}}_{\mathbf{i}}\right)=\frac{\exp \left({\mathbf{x}}_{\mathbf{i}}\right)}{\sum _{\mathbf{j}}\exp \left({\mathbf{x}}_{\mathbf{j}}\right)}$	This is not a function of a single fold from the previous layer
tanH	$f\left(x\right)=\left\{\begin{array}{cc}-1& ifx<0\\ 1& ifx>0\\ 0& ifx=0\end{array}\right\}$

DOI: 10.7717/peerj-cs.687/table-2

Evaluation metrics for the regression model

In the proposed framework, the input to the model is the agro-meteorological parameters collated as per the Indian Meteorological Department week wise for past 30 years. Based on these values, the new values for the parameters are predicted by the model for the future date provided. Once the new values are predicted by the model for the said period, these values are further considered for predicting the rice diseases. This is where regression is applied. There are various performance metrics used to measure the accuracy of the regression model such as Mean Absolute Error(MAE), Mean Square Error(MSE), Root Mean Square Error(RMSE), etc. (Kaundal, Kapoor & Raghava, 2006). An error can be defined as the difference between the actual value and the predicted value by the built model (Piekutowska et al., 2021). The evaluation parameter used to measure the regression accuracy for the proposed model is the mean absolute error (MAE).

(13) $$\mathbf{M}\mathbf{A}\mathbf{E}=\sum _{\mathbf{i}=\mathbf{1}}^{\mathbf{n}}\frac{|\mathbf{p}\mathbf{r}\mathbf{e}\mathbf{d}\mathbf{i}\mathbf{c}\mathbf{t}\mathbf{e}\mathbf{d}\mathbf{v}\mathbf{a}\mathbf{l}\mathbf{u}\mathbf{e}(\mathbf{i})-\mathbf{a}\mathbf{c}\mathbf{t}\mathbf{u}\mathbf{a}\mathbf{l}\mathbf{v}\mathbf{a}\mathbf{l}\mathbf{u}\mathbf{e}(\mathbf{i})|}{\mathbf{n}}$$where n is the number of instances considered in the climatic dataset.

The mean absolute error is the mean of the sum of all the absolute values of the error. It is simply calculated by averaging the errors. As there are outliers in the climatic dataset that we have used in the research model, MAE is the best suitable metric to measure the regression accuracy of the model. An outlier is a data point in a dataset that lies outside the overall distribution of the dataset. As there is variability in the collected climatic dataset, outliers are present in the dataset. After fitting the model, the performance of the model is assessed by comparing the model predicted values of the agro-meteorological parameters to actual data in the climatic dataset. The lower the value of MAE higher is the crop disease prediction accuracy. If MAE has a value equal to zero, then model predictions are perfect.

Hyperparameter of ANN

Epochs are the number of times the data is given to the neural network. Here the performance of the model was observed on various combinations of epochs to get minimum loss in the predicted values. It was observed that if the epoch was set to 100 the data was under fitted. If the number of epochs was set to 500 then the MAE value is 0.6018. If epochs are set to 700, a minimum MAE value of 0.5632 is obtained, but the time required to train the model was more. So 500 epochs are selected as an optimal solution. Figure 6 shows values of MAE for the various number of epochs. Batch size is the segmentation of the whole data into smaller parts so that the data can be processed individually. Here the batch size is 32. Optimizers are used to minimize the loss so that the model performs better. Adam is the optimizer used to build the model.

Conventionally, an artificial neural network comprises three different types of layers namely input layer, hidden layer, and output layer. The number of input layers used to build the model is eight as the number of features in the dataset is eight, the number of hidden layers used is two. The number of neurons used in hidden layers was selected in such a way that too large neurons would overfit the data and too low number of neurons would underfit the data. There are three thumb rules to calculate the number of neurons in hidden rules first being the number of hidden neurons should be between the size of the input and output layer. The second rule is hidden neurons should be 2/3 the size of input plus output layer size and the third is the number of hidden neurons that should be less than twice the size of the input layer. The third rule is used in the proposed model to calculate the number of hidden layers. Therefore, the number of hidden layers used is 15. The output layer must be equal to the number of target classes. In the proposed model, the number of target classes is five so there are five neurons in output layers. So, precisely (8–15–5) is the topology of the model (Abdipour et al., 2019). After the optimized number of epochs have been obtained MAE is calculated for four types of AF and for two cases of dataset split. Table 3 shows the MAE values for different AF with different data split.

It can be observed from Table 3, that ReLU AF outperforms other AF to predict agro-meteorological parameter values if used with a 70–30% dataset split.

Evaluation metrics for classification model

In our model, true negatives (TN) are the instances where the plants did not have the disease and the model also predicted that plants are not affected by the disease. True positives (TP) are cases in which the plants have diseases and our model estimated that they will have it as well. However, there are several circumstances in which the plants are not affected by the diseases but our model predicts that they are affected. This is known as false positives (FP) and also known as a Type I Error. Similarly, in some circumstances, the plants do have a disease, but our model predicts that it is not affected by the disease. This is known as false negatives (FN) and is also called a Type II Error.

Precision: It is the percentage of plants that are accurately classified as having disease out of all those who have it based on weather parameters. In terms of mathematics:

(14) $$\mathit{P}\mathit{r}\mathit{e}\mathit{c}\mathit{i}\mathit{s}\mathit{i}\mathit{o}\mathit{n}=\frac{\mathit{T}\mathit{P}}{\mathit{T}\mathit{P}+\mathit{F}\mathit{P}}$$

Recall: The recall is a test of how well our model detects true positives. As a result, recall shows us how many instances were accurately reported as having plant disease out of all those that have it. In terms of mathematics:

(15) $$\mathit{R}\mathit{e}\mathit{c}\mathit{a}\mathit{l}\mathit{l}=\frac{\mathit{T}\mathit{P}}{\mathit{T}\mathit{P}+\mathit{F}\mathit{N}}$$

F1 score: In our model, the instances that were wrongly diagnosed with plant disease are equally critical because they may be signs of another plant disease, therefore achieving not only a high recall but also a high precision is important. In such instances, we use a technique known as F1-score.

(16) $$\mathit{F}\mathbf{1}\mathit{S}\mathit{c}\mathit{o}\mathit{r}\mathit{e}=\frac{\mathit{P}\mathit{r}\mathit{e}\mathit{c}\mathit{i}\mathit{s}\mathit{i}\mathit{o}\mathit{n}\ast \mathit{R}\mathit{e}\mathit{c}\mathit{a}\mathit{l}\mathit{l}}{\mathit{P}\mathit{r}\mathit{e}\mathit{c}\mathit{i}\mathit{s}\mathit{i}\mathit{o}\mathit{n}+\mathit{R}\mathit{e}\mathit{c}\mathit{a}\mathit{l}\mathit{l}}$$

The more nearer the F1 score is equal to 1 the more accurate is the model.

Accuracy: The ratio of the total number of accurate predictions to the total number of predictions is known as accuracy.

(17) $$\mathit{A}\mathit{c}\mathit{c}\mathit{u}\mathit{r}\mathit{a}\mathit{c}\mathit{y}=\frac{\mathit{T}\mathit{P}+\mathit{T}\mathit{N}}{\mathit{T}\mathit{P}+\mathit{F}\mathit{P}+\mathit{T}\mathit{N}+\mathit{F}\mathit{N}}$$

Comparison of Activation functions with 70–30% dataset split

The results are shown from Figs. 7 to 10 represent the values of precision, recall, and F1 score for each rice disease class when used with 70–30% training and testing split of dataset parameters. The results are shown for sigmoid, tanH, ReLU, and softmax functions respectively.

In Fig. 7, Class 5 rice diseases False Smut are more accurately identified as compared to other classes of diseases followed by Class 1 of healthy conditions. The rest classes are moderately classifying the diseases.

In Fig. 8, Class 5 and Class 3 are identified accurately with 0.83 and 0.79 F1 scores respectively. These values are low as compared to sigmoid AF. Classes 1,2 and 4 are classified with F1 score accuracy between 0.73 to 0.76 values.

Figure 9 represents that with ReLU AF, only classes 3 and 4 have F1 scores above or equal to 0.90 value. Class 1, 2 and 5 have scores within the range of 0.86 to 0.89 so it needs to be improved.

By observing Fig. 10, the F1 score metrics for all the classes are above 0.90 which means softmax AF when combined with 70–30% dataset split gives maximum accuracy.

Comparison of activation functions with 80–20% dataset split

The results are shown from Figs. 11 to 14 represent the values of Precision, Recall, and F1 score for each rice disease class when used with 80–20% training and testing split of dataset parameters. The results are shown for sigmoid, tanH, ReLU, and softmax functions respectively.

Figure 11 represents that Class 1, 2, and 4 are classified with maximum accuracy but Class 3 and 5 are moderately classified. If compared to a 70–30% split, performance is better.

In Fig. 12, Class 5 is misclassified, reciprocally with 70–30% split Class 5 was more accurately classified with tanH AF.

In Fig. 13, Class 1, and 5 have good accuracy with ReLU AF as compared to Class 2, 3 and, class 4.

Figure 14 depicts that Class 1,3 and 5 are achieving good accuracy however Class 2 and 4 are still misclassified.

After comparing both the cases, the solution that is obtained must be optimal which means the model should be able to predict the occurrence of all diseases with maximum accuracy. Softmax AF gives maximum accurate results for predicting all five classes of rice diseases when data was split with 70–30% capability. It can be concluded that the proposed model performs best when the model was applied to 70–30% training and testing sets. The model performance decreased when it was applied to 80–20% data split.

Receiver operating characteristics (ROC) for four different types of AF implemented in this research paper are shown in Fig. 15. It calculates the micro-average curve for all the activation functions. It clearly shows that softmax AF produces maximum accuracy as the ROC curve is highly inclined towards true positive rate. Also, tanH AF misclassifies the instance in most cases among the four classifiers demonstrated in the paper as its curve is inclined more towards false positive rate.

Table 4 shows the comparison of the overall accuracy of four activation functions with two different cases of dataset division.

It can be concluded that by making use of softmax AF the model achieves a maximum prediction overall accuracy of 92.15% which is the highest of all. The model performs best at data split 70–30% of training and testing sets as compared to model when applied to 80–20% data split. ReLU activation function performs in an average manner for both the cases of data split. It performs better when the training set is larger. The sigmoid activation function gives better performance with 80–20% data split. Tanh activation function has demonstrated poor performance when compared with all other activation functions implemented in the research paper. Softmax activation function outperforms when applied with 70–30% data split.