Introduction

To date, the principles of digitalization and intellectualization of technological processes are one of the global trends in improving the efficiency of production processes at enterprises of various profiles, scales and forms of ownership. Agriculture is one of the strategic sectors of the national economies of many countries, and therefore requires constant search and implementation of scientifically substantiated approaches to the sustainable development of agricultural practices. Smart farming is a key concept relevant to running production processes in agricultural enterprises, particularly open-field crop production. This concept, in turn, is made possible by introducing technologies such as: Internet of Things, machine learning and artificial intelligence, remote sensing, drones and robotics. This approach allows achieving a significant socio-economic, environmental and technological effect, which consists in: efficient use of material, land and labor and time resources; increasing the resistance of field crops to changing climatic conditions; increasing yields and minimizing negative environmental impact [1,2].

Based on a priori analysis of current statistics on agricultural practices at the global level, it has been established that cereals (wheat, rice, corn, barley, and others) are the most popular crops in terms of cultivated areas and specific yields [3,4]. Over the past decade, wheat has been the leader among cereals at the national level in terms of cultivated areas: from 5.28 million hectares to 7.1 million hectares. It is also necessary to emphasize that with an increase in cultivated areas, there is no proportional increase in the yield of cereals [5]. This phenomenon is rooted in the fact that during the full cycle of cultivation, grain crops are subject to destabilizing effects of physical (changing agroclimatic conditions) and biological (pests and diseases) factors, which negatively affect the integral stress resistance and productivity and, as a result, crop yields.

Therefore, an actual scientific and practical task is to develop and implement methods, models, and software and hardware for predicting the occurrence of crop diseases at the pre-symptomatic stage in real time, which will allow timely planning and implementation of agrotechnical measures to increase the stress resistance of crops and preserve the harvest in changing agroclimatic conditions.

In the present-day world practice, there is a significant number of high-quality research and developments of information and computer technologies for agrotechnical purposes to detect the parameters and characteristics of the processes of occurrence and progression of crop diseases based on various data collection technologies [6], in particular: obtaining and analyzing graphic images from satellites [7,8] and UAVs [9,10], as well as collecting and processing data from ground-based sensor networks [11,12].

One of the main tasks, which is the focus of many relevant studies in developing intelligent information technologies for predictive monitoring of crops, is the precise and reliable analysis of observation results. To date, machine learning methods have gained considerable popularity in solving problems of intellectual analysis of large amounts of data when creating intelligent information technologies for various applied fields [13,14]. In agriculture, such technologies are used for intelligent processing of agroclimatic data distributed in time and space [15,16], as these approaches allow aggregating, analyzing, and interpreting large amounts of measurement data with subsequent automatic support for making management decisions to optimize agrotechnical procedures.

The perspectives of integrating sensor networks for agroclimatic monitoring and machine learning methods have been proven by the authors of scientific studies on: the introduction of precision farming systems [17], analysis of promising practices for managing agrotechnical processes and resources [18], accounting for the impact of changing climatic conditions on crop cultivation regimes [19], and others.

Thus, the results of the analysis of the current state of scientific and applied research prove the potential and effectiveness of creating and implementing information technologies for predictive monitoring of crop diseases based on online measurements of soil and climatic parameters with their subsequent processing by software based on machine learning algorithms. Consequently, the current research task is to develop computer components for complex intelligent processing of agromonitoring measurement data that implements a full cycle of data transformation (primary statistical processing, selection of informative features, and predictive analytics) within an integrated hardware and software architecture, taking into account the specifics of detected diseases and agroclimatic growing conditions for specific types and periods of grain crops.

Therefore, the aim of the article is to further develop information technologies for agrotechnical purposes to predict the risk of occurrence of grain crop diseases (Powdery Mildew Blumeria graminis in wheat) at the pre-symptomatic stage through the development and research of computerized method of intelligent analysis of measured data of agroclimatic monitoring utilizing machine learning models. The object of research is information processes of software analysis of agroclimatic data distributed in time and space. The subject of research is methods and computer models of complex predictive processing of agroclimatic monitoring data.

Accordingly, researching the development and validation of computerized methods and software components of predictive transformation and analysis of agroclimatic measurement data during the creation of information technologies for agrotechnical purposes to increase the stress resistance of grain crops is an actual scientific and practical task.

Materials and methods

Two primary software environments were used in this research: Python 3.10.12 and MATLAB R2024a. Preliminary data analysis was performed in Python using libraries such as Pandas and NumPy. MATLAB was used to train the classifier model with the possibility of further generating code for microcontrollers. For this purpose, the Simulink and Statistics and Machine Learning toolboxes were used.

In Figure 1, a generalized structure of the research in this article is illustrated. The data for the research was obtained using professional Metos weather stations from Pessl Instruments via the FieldClimate IoT platform, access to which was provided by Metos Ukraine LLC. The experimental data reflects the results of monitoring soil and climatic parameters collected in two agroclimatic zones of Ukraine from September 2022 to September 2023:  northern steppe of Ukraine: a zone characterized by an arid and very warm climate. The hydrothermal coefficient varies from 0.7 to 1.0, and the annual sum of temperatures ranges from 2900 °C to 3300 °C;  forest-steppe of Ukraine: insufficiently humid and warm zone with a hydrothermal coefficient of 1.0 to 1.3 and an annual sum of temperatures from 2500 °C to 2900 °C.

The data sample from the northern steppe zone (Dnipro region) consisted of 8656 records with a sampling interval of 1 hour and 14 attributes. Similarly, the data sample from the forest-steppe zone (Cherkasy region) consisted of 8687 records with the same sampling interval and number of attributes. Both samples have the probability of occurrence of Powdery Mildew Blumeria graminis as the target function for further analysis and modeling. A description of all attributes included in the two data sets is given in Table 1. During the preliminary data analysis, all rows containing missing values were removed to avoid distortion of the results and to ensure the correctness of the modeling. There were two such rows and given that 'PMBG' is calculated once a day, their removal does not affect the value of the target function. In addition, for the initial analysis of the data in Python, the describe() function from the Pandas library was used to obtain statistical information about the numerical characteristics of the data, such as mean, standard deviation, minimum and maximum values, and quartiles. These indicators allowed assessing data distribution and identifying possible anomalies and general trends in the dataset. The statistical indicators obtained as a result of using describe() for the combined sample from the two regions are shown in Table 2. The average 'AT' value of 10.6 °C indicates a moderate climate in the region. The range of values from 10.9°C to 37°C indicates the presence of both cold and very warm periods. The 'AT' values are centered around 9.8°C (median), with the bottom quartile (25%) falling within 2.7°C and the top quartile (75%) falling within 18.2°C, indicating a significant temperature variation. The average 'RH' value is 72.6%, indicating generally humid conditions. The humidity varies widely, with a minimum of 19% and a maximum of 100%, with most values concentrated between 61% and 87%. A mean 'PR' value of 0.1 mm indicates low precipitation. In fact, most of the records show no precipitation, as the median and 25th quartile are 0. The maximum value of 15.2 mm indicates significant but rare precipitation. The average 'LW' value is 8.6 minutes per hour, indicating predominantly dry conditions. 75% of the 'LW' values are 0, which means that the leaves remain dry most of the time. 75% of 'ET' values below 0.1 mm indicate generally low evapotranspiration. The average 'PMBG' value is 12.2%, indicating a low average risk of developing the disease. The maximum 'PMBG' is 70%, which indicates a significant risk in certain periods. At the same time, the median of 0% indicates that a significant part of the sample had no or low risk of the disease, but the 75th quartile at 20% shows that there are periods with an increased risk of developing the disease.

The data in Table 2 shows moderate climatic conditions with relatively low precipitation and moderate temperatures. The risk of developing Powdery Mildew is generally low, although periods of higher risk require attention.

The risk of Powdery Mildew varies by 10% on average during the day under favorable conditions [20]. Thus, it was decided to aggregate the hourly measurement results to a daily resolution. Additionally, eight attributes were introduced: 'FVT12S' -the number of hours when the temperature ranges from 12°C to 21°C; 'TL16S' -the number of hours when the temperature is below 15°C; 'TG21S' -the number of hours when the temperature exceeds 21°C; 'FVT16S' -the number of hours when the temperature ranges from 16°C to 21°C; 'TG25S' -the number of hours when the temperature exceeds 25°C; 'SR_count' -the number of hours when solar radiation was greater than 0; 'RHG85S' -the number of hours when relative humidity was greater than or equal to 85%; 'LW_count' -the number of hours when the leaves were wet.

Taking into account the above transformations, the columns of the final table are renamed according to the predefined names stored in the variable INPUT_SUMMARY_COLUMNS = ['AT_mean', 'FVT12S', 'TL16S', 'TG21S', 'FVT16S', 'TG25S', 'DP_mean', 'SR_sum', 'SR_count', 'VPD_mean', 'RH_mean', 'RHG85S', 'PR_sum', 'LW_count', 'WS_mean', 'WD_mean', 'WG_mean', 'ST_mean', 'ET_sum', 'PMBG_mean'], and the calculations correspond to the Python code (see Appendix A). The difference between the value of the disease risk for the previous day and the current day is calculated using diff(). Then, based on this difference, a new attribute 'PMBG_class' is created in which the class of risk change is stored: 1 -risk has increased, 2 -risk has decreased, 0risk has not changed.

The data was then split into training and test samples in a 70:30 ratio. 70% of the data was used to train the models, and 30% to evaluate their performance.

After performing these steps, the imbalance of classes in the target variable of the training dataset was detected (class 1 -6.6%, class 2 -6.4%, class 0 -87%). This can lead to the model learning to favor a more common class, ignoring important features of less represented classes, which ultimately degrades the overall performance of the classifier model. Two approaches were used to address this problem: 'undersampling' and 'oversampling'. The 'undersampling' approach is about reducing the number of instances of the majority class to achieve a balance with the minority. This allows the model to better account for the minority, as the number of samples of all classes becomes proportional. To implement this approach, the pandas.DataFrame.sample() function was used, which allows randomly selecting instances from the majority class. Figure 2(a) shows a comparison of the original dataset before dividing it into training and test samples and the data after resampling. The second approach is 'oversampling', which implies increasing the number of minority samples to achieve balance with the majority. One of the most common methods for this is the Synthetic Minority Over-sampling Technique (SMOTE) [21,22]. This method generates synthetic minority samples by interpolating between real samples. It works by selecting a few nearest neighbors for each minority pattern and creating new patterns based on these neighbors. These two approaches help to achieve a balance in the data and improve the overall quality of the model, preventing it from being biased towards the class with more instances and increasing classification accuracy for less represented classes.

The next step in the process of preparing data before training classification models is scaling, including standardization and normalization. Scaling is an important step in data processing because different features can have different scales, which can negatively affect the performance of machine learning models, especially those based on distance or gradient methods.

The mean and standard deviation are computed from the training data, and these values are then used to standardize the training data. The same mean and standard deviation from the training set are also applied to standardize the test data. This ensures that both datasets are transformed consistently, allowing for accurate evaluation of the performance of the model. The standardization is performed in MATLAB Classification Learner App [23] automatically before training the models.

The last step is to extract significant features. To do this, the 'Feature Selection Wrapper Class' algorithm from a GitHub toolbox was used [24]. In the research, the Harris Hawk Optimizer (HHO) method was used to extract significant features. This feature selection method mimics the behavior and joint hunting strategy of Harris's hawks when chasing prey.

The algorithm proposed by Heidari and co-authors [25] is based on the swarm approach, does not use gradients, and includes evolutionary optimization elements. HHO consists of two main phases, exploration and exploitation, which alternate in time to find the best parameters.

This algorithm has a high convergence rate and excellent local search capabilities, which makes it effective for feature selection problems. In this research, the HHO algorithm proposed five significant features ('TL16S', 'FVT16S', 'DP_mean', 'ST_mean', 'ET_sum'), which were extracted for further analysis and model training.

After feature selection, the training set was used to train various Simulink-compatible machinelearning classifiers. Types of classifiers chosen and their relevant hyperparameters [26,27] are summarized in Table 3. Based on the described methodology for assessing the risk of Powdery Mildew Blumeria graminis in wheat, the research steps of this article were presented in the form of a structural algorithmic scheme, as shown in Figure 3.

When training the models, the five-fold cross-validation methodology was used to assess their performance [28]. This approach allows for efficient use of available data and reduces the risk of model overtraining.

The data is divided into five subsets of equal size. At each iteration, four subsets are used to train the model, and one is used to test it. As a result, average accuracy rates are obtained, which gives a more stable and reliable assessment of model quality on different chunks of the dataset.

Results and discussion

After training and testing the classification models, the following results were obtained, as shown in Table 4. For each model, training was performed using undersampling and oversampling data balancing approaches, in particular, the accuracy rates on the validation (Val.) and test (Test) datasets were compared. The tree models (Fine Tree, Medium Tree, Coarse Tree) showed the best accuracy when using oversampling, with the best accuracy for Fine Tree (84% on validation and 73% on testing). Coarse Tree significantly reduced the accuracy on the undersampled test set (56%). SVM models had stable results, especially when using oversampling. For example, Quadratic SVM and Cubic SVM demonstrated high accuracy rates (up to 90% on the validation set). Fine KNN and Weighted KNN achieved the best results in oversampling, showing 90% and 89% accuracy, respectively, on validation. Medium and Wide Neural Networks showed the highest accuracy (91% and 90%, respectively) with oversampling, indicating their ability to efficiently process more balanced data. The tree-based methods (Boosted Trees, Bagged Trees, RUSBoosted Trees) also performed well, especially Bagged Trees, with 88% accuracy on the validation and 74% on the oversampled test set. Thus, the use of oversampling generally improved the performance of the models, especially for SVMs, KNNs, and neural networks.

To analyze the performance of the models in more detail, the Confusion Matrix for the training and test data, as well as the ROC curves for the three classes, were constructed, as shown in Figure 4 and Figure 5. This analysis was performed for two models: Bagged Trees (on undersampled data) and Medium Neural Network (on oversampled data).

The Confusion Matrix for both models showed similar results. For the Bagged Trees model on the undersampled data, it can be seen that the model generally copes with the classification of classes, although there are some errors in the classification of smaller classes. The Medium Neural Network model, on the other hand, showed better results on oversampled data, reducing the number of misclassifications, especially for less represented classes.

The ROC curves for both models show high sensitivity and specificity for each of the three classes. Both models perform well for most classes, with the Medium Neural Network on oversampled data showing slightly higher performance in terms of area under the curve (AUC) for class '2'. The Bagged Trees model trained on undersampled data was then exported [29] to the Simulink environment (Figure 6) to implement a computerized model for automatic data processing to predict the Risk of Powdery Mildew in Wheat. An example of the model output in Figure 6 for the northern steppe scenario in the form of time graphs is shown in Figure 7, comparing actual and predicted data on the risk of Powdery Mildew Blumeria graminis in wheat. The analysis of time graphs shows that the predicted data from the model is able to reproduce the main trends of growth and reduction of disease risk in the relevant time periods. Although there are some discrepancies between the actual and predicted values at certain points, in general, the model reflects risk behavior well and can be used for operational monitoring and decision-making in openfield conditions.

Conclusion

In this research, the relevant scientific and practical task has been solved by developing and researching the computerized method for predicting the risk of powdery mildew in wheat based on software analysis of soil and climatic monitoring data. The research allowed for the further development of information technologies for agrotechnical purposes to predict the risk of occurrence of grain crop diseases (Powdery Mildew Blumeria graminis in wheat) at the pre-symptomatic stage.

The main results of the research include: 1. Using undersampling and oversampling methods to solve the problem of class imbalance in the training sample. 2. The application of feature selection algorithms, in particular Harris Hawk Optimizer, reduced the number of features to five. 3. Classification models such as Bagged Trees and Medium Neural Network performed well on both validation and test datasets, demonstrating good generalizability. 4. The export of the Bagged Trees model to the Simulink environment and the subsequent generation of program code for microcontroller devices allows it to be used for real-world prediction and control in agroclimatic systems based on fog architecture. To improve the results, future research should pay attention to the following aspects: 1. Improving the parameters of algorithms such as Bagged Trees and Neural Networks by further tuning hyperparameters, which can lead to even more accurate results. 2. Involvement of new data sources, such as information on field cultivation, which can improve the recognition of conditions that contribute to the occurrence and development of diseases. 3. Further integration of the model with IoT platforms for automatic real-time monitoring can provide more up-to-date and accurate data for predicting disease risks.