Integrating IoT for Soil Monitoring and Hybrid Machine Learning in Predicting Tomato Crop Diseases in a Typical South India Station

1. Introduction

Precision agriculture technology has revolutionized crop management and disease control. Managing plant diseases in profitable crops like tomatoes is a significant concern for farmers because diseases can reduce tomato yield and quality. It is known that tomato plants are mostly affected by living things like fungi, bacteria, viruses, and worms, as well as non-living things like temperature, sunlight, and lack of nutrients [1]. Various diseases associated with tomato crops are Early blight, Septoria crop spot, Late blight, and Cercospora crop mold [1], which are caused by fungal diseases.

Disease diagnosis involves skilful, time-consuming, and error-prone visual exams. To properly address crop diseases, they must be quickly and accurately identified. Researchers have developed effective machine-learning approaches to predict tomato crop disease [2], due to the fact that ML approaches improve accuracy, efficiency, and scalability. Crop visual differences can help ML approaches discover and classify disorders [3,4]. Specifying disease indicator threshold values significantly enhances the accuracy of machine learning models in predicting plant illnesses, as evidenced by recent advancements in agricultural technology.

This methodology leverages both professional agronomic insights and historical data analytics, enabling a nuanced comparison that underscores the importance of integrated approaches for precise diagnosis. In particular, the severity of plant diseases, such as those affecting tomatoes, can be meticulously quantified through observable parameters like crop coloration, spot patterns, and lesion appearances, as detailed in [5]. Table 1 lists tomato crop infection symptoms and signs.

3. Objectives of the Present Research

According to the literature review, hybrid machine learning is required since it can analyse complicated data more thoroughly than traditional methods. Secondly, and most importantly, region-specific models based on real-time data are scarce. In a broader sense, as different regions have distinct climatic factors, including varying degrees of humidity, different types of soil, and varying rates of disease, it is essential to develop machine learning models that are specifically tailored to each location. Thirdly, comprehensive analyses of metrics are needed to improve resource management.

Therefore, the current research study aims to:

a)

Develop a hybrid machine-learning technique that yields optimal results.

b)

Research on crop disease detection in various climates results in the creation of models unique to a given area.

c)

Offer a comprehensive list of evaluation metrics, such as accuracy, precision, recall, and F1-score, ROC-AUC.

4. Data and Methodology

4.1. Tomato Crop Location

The present study considered tomato fields located at Anakapalle (Geographic Latitude 17.6896° N, Geographic Longitude 83.0024° E, MSL ~ 29 m), a major municipality in the state of Andhra Pradesh, south India. Figure 1 shows the map of India and the locaiton of Anakaplle. Tomatoes (Solanum lycopersicum) flourish in Andhra Pradesh's diverse agro-climatic conditions, which span from coastal regions to fertile interior plains. The state's strategic location and varied seasons enable year-round cultivation, ensuring a steady supply of this essential vegetable. Andhra Pradesh, the largest tomato-producing state in India, achieved a notable production of 2,744.32 thousand metric tonnes last financial year, i.e., FY 2023-2024.

This study utilized one-minute datasets gathered meticulously from fields in Anankapalle (India) over five consecutive days, from 27 November to 01 December 2023. For clarity, Figure 2 presents probability density functions (PDFs) of these datasets collected during our research.

The provided histograms with PDFs for Nitrogen (%), Potassium (%), Phosphorus (%), Humidity (%), Soil pH, and Temperature (°C) all display approximately normal distributions, with most data points clustering around the mean values. Nitrogen is centered around 2%, Potassium around 3%, Phosphorus around 1%, Humidity around 80%, Soil pH around 6.5, and Temperature around 25°C, each showing a slight right skew. The close alignment of the PDFs with the histograms indicates that the datasets are well-behaved and suitable for statistical analysis and modeling.

4.2. Data Collection- Arduino Microcontroller

In the pursuit of agricultural excellence, modern farms employ a suite of sophisticated sensors, each designed to unravel the environmental intricacies of soil and air that critically influence crop health and yield. The sensor array, a fusion of metal and plastic engineering, operates in a symphony of data collection and analysis. One such sensor, adept at monitoring thermal variations, vigilantly regulates the temperature within the fields, safeguarding the warmth essential for the optimal growth of grapevines [20]. Figure 3 shows an Arduino microcontroller to collect soil parameters, pH value, temperature, humidity, and others in the presenet research.

A GPS sensor surveys the farm's topology, ensuring uniform sunlight distribution for even growth. An RGB color sensor, called Sarah’s tool, monitors soil hues to assess health and fertility, essential for vine sustenance [28]. An air quality sensor discreetly measures particulate matter [29], wirelessly sending data to Sarah's mainframe to maintain pure air. Temperature and humidity sensors protect the fruits, with their crucial readings streamed to the cloud via an ESP8266 Wi-Fi microcontroller, enabling real-time data processing [30].

A dedicated NPK sensor ensures optimal nutrient levels and balanced pH for the vines, streaming data to a cloud-based platform via an Arduino board. This board acts as the farm's AI, learning, adapting, and issuing directives based on continuous data flow, maintaining vigilant environmental monitoring. Powered by a 24V battery system, this network of sensors and technology upholds the necessary equilibrium for the vineyard’s flourishing, as shown in the attached diagram [31]. Table 3 depicts various sensors used in the present study.

An attempt is also made to make a comparison between soil parameters (temperature and humidity) measured by the Arduino sensor and those provided by the India Meteorological Department (IMD, https://mausam.imd.gov.in/), and almost a one-to-one similitude has been found, which implies that this research used only quality data.

4.3. Calculation of Kendall’s Correlation (τ)

Kendall's correlation, named after Maurice Kendall, is essential in hybrid machine learning for feature selection and model evaluation due to its robustness to outliers and ability to handle ordinal data. In agricultural data management, it analyses datasets comprising temporal markers, temperature, and humidity, pH, and NPK levels. The process begins with rigorous data preparation, addressing missing values through imputation or removal to prevent bias. Categorical and timestamp data are encoded to numerical values to calculate Kendall's correlation.

The Kendall correlation coefficient (τ) is calculated as:

$$\mathrm{\tau }={\displaystyle \frac{2(n_c-n_{d)}}{n(n-1)}}$$

where n_c is the number of concordant pairs, n_d is the number of discordant pairs, and n is the total

number of pairs.

The algorithm iteratively evaluates all column pairs, identifying concordant and discordant pairs based on their relative order. Concordant pairs are those where both elements are greater or lesser than another pair, while discordant pairs have differing orders. The coefficient τ ranges from -1 (complete disagreement) to 1 (perfect agreement), with 0 indicating no association. Interpreting these coefficients reveals intervariable relationships, aiding in factor selection, trend analysis, and predictive modeling. This process unveils insights for informed decision-making and strategic planning in agricultural management.

4.4. Bayesian Optimization with the KNN Algorithm

Bayesian optimization is a powerful method for optimizing objective functions, especially in hyperparameter tuning, where each evaluation requires significant computational resources and time [32]. It is effective for finding the optimal hyperparameters for a K-Nearest Neighbors (KNN) model by defining an objective function, such as model accuracy. The process starts with setting a prior distribution over the hyperparameters and using an acquisition function like Expected Improvement (EI) to balance exploration and exploitation.

The algorithm begins with random initial hyperparameters, evaluates the objective function, and updates a surrogate model that approximates the objective function. Iteratively, it uses the surrogate model and acquisition function to select new hyperparameters, refining the model until convergence criteria or iteration limits are met. The optimal hyperparameters are then used to train the KNN model on the entire dataset, aiming to outperform models trained with randomly chosen hyperparameters. Finally, the fine-tuned KNN model is tested on an independent dataset to assess its generalization and performance, efficiently reducing the computational cost and time of exhaustive search methods.

4.5. Data Preparation and Plant Health Classification

To leverage IoT, a complex Arduino-based circuit terminal was used to systematically collect soil samples and gather data on environmental parameters such as humidity, temperature, and nutrient content. After data collection, the data is securely transmitted to the cloud for further analysis. The first step is data cleansing, which is accurately performed using Python. This involves removing any extraneous or disruptive information to retain only the most pertinent data. Python's capabilities are then used to optimize and synchronize the data structure. Next, Python's analytical tools are employed for correlation research to uncover links and trends within the processed data, essential for gaining insights into environmental factors affecting crop health.

In the predictive modeling stage, sophisticated algorithms like KNN and Bayesian approaches are applied. Python's scikit-learn and Matplotlib are used for comparative analysis of various prediction techniques. This research demonstrates that incorporating IoT technology into data processing and analysis can enhance agricultural operations and open new avenues in precision farming. It illustrates the potential success of integrating traditional farming with cutting-edge technology. Field IoT peripherals provide real-time farm data with time-stamped entries, offering minute-by-minute time series updates. This granularity allows for precise monitoring of field conditions. Table 5 shows how we prepared collected data to propose a hybrid model, while Table 6 shows how the disease classification has done.

a)

Each record is annotated with a date and time stamp. The format is consistent with DD-MM-YYYY dates and HH:MM times, suggesting one-minute automated logging.

b)

NPK Level: This measures how much nitrogen (N), phosphorus (P), and potassium (K) are in the soil. It's given as a number between 1 and 3, indicating low, medium, or high levels of these nutrients.

e)

Temperature (°C): The ambient temperature is recorded in degrees Celsius (°C), ranging from 21°C to 24°C and this tight range reflects consistent weather during data collection.

f)

Humidity (%): Measures air moisture, ranging from 45% to 50%. Paddy fields are usually moderately humid.

g)

pH value: Recorded soil acidity or alkalinity ranges.

h)

Diseased: This binary classification serves as the target variable for machine learning models, where '1' represents the presence of disease and '0' signifies a healthy state.

5. Results and Discussion

By applying the Kendall correlation coefficient, the study seeks to establish monotonic relationships between environmental factors and the health status of the plants. The non-parametric nature of this correlation is suitable for datasets that may not adhere to a normal distribution and can handle ordinal and continuous variables alike. The processed data set is thus poised for use in classification algorithms, such as decision trees (DT), support vector machines (SVM), or neural networks (NN), to predict disease presence in the paddy field. The goal is to leverage this model to forecast and mitigate crop diseases, thereby enhancing yield and reducing losses due to plant health issues. The dataset provides a basis for developing a predictive model that can be trained, tested, and validated for accuracy and reliability in real-world agricultural settings.

Figure 4 shows the proposed framework, where a hybrid ML enabled is proposed that is backed by an IoT system for soil nutrients monitoring in a real-time environment.

The various phases of the present research framework are explained, as under:

a)

Data Collection Phase: The first step is to collect environmental data using an IoT device. It's very important that this data is accurate and complete because it forms the basis for all future analysis and decisions.

b)

Data Processing and Evaluation: After ensuring the data is accurate and complete, it goes through processing where we assess soil health. This step is key because it turns raw data into useful information about soil health, which is important for good crop management.

c)

Data Analysis and ML Implementation: The processed data is then analysed to see if it meets our standards. If it does, we move on to using machine learning (ML) algorithms. In this stage, we compare different ML algorithms to find the best one for improving soil health.

d)

Model Assessment and Adjustment: After implementing the ML algorithms, we check how well the model performs. If it doesn't meet our standards, we make adjustments or retrain the model. If it does meet the standards, we analyse its performance based on set metrics.

e)

Operational Implementation and Continuous Evaluation: Once the model proves it can improve crop management, it is used in real-world settings. This step is critical to see how well the technology works in practical situations.

The model yields good results over a wide range of criteria. The accuracy rate is 95 out of 100 instances, which is almost 95.35%. The model's precision allows it to generate precise predictions, which is useful in practical applications. Table 7 shows various performance metrics obtained using the hybrid ML.

The hypothetical ROC curve, shown in Figure 5, with an AUC (area under the curve) of 0.95, demonstrates a highly effective classifier. Recent publications highlight the significant impact of soil nutrients on managing tomato crop diseases and improving productivity. Soil amendments with green manures like fodder radish and bio-enriched organic manures combined with macro and micronutrients enhance soil fertility, growth, yield, and disease resistance in tomato crops [33,34]. The ROC curve likely represents the effectiveness of these nutrient management strategies in improving tomato crop health, with an AUC of 0.95 indicating high discriminative ability in distinguishing healthy from diseased crops.

The graph shown in Figure 6 presents a comparative analysis of the performance metrics between training and testing datasets for a machine learning model. Observing the minimal divergence between the training and testing results, we can infer a consistent pattern across all evaluated metrics: accuracy, precision, recall, and F1 score. The closeness of these metrics signifies that the model generalizes well to new, unseen data, indicative of a well-tuned model that avoids overfitting. In particular, the training metrics exhibit a slight elevation, which is expected as models typically perform better on data they have been trained on. However, the test metrics are not far behind, suggesting that the model retains its performance robustness when exposed to new data. This alignment between train and test performance is crucial for ensuring that the model is reliable and effective when deployed in real-world scenarios.

The graph reflects strong overall performance, with high values in accuracy, precision, recall, and the F1 score. These metrics collectively point towards the model's capability to correctly identify positive cases, precisely predict the positive class, and balance the trade-off between precision and recall effectively. Such a performance profile indicates the model's suitability for practical application, although careful consideration for potential false positives and negatives remains necessary to fine-tune its predictive accuracy further. This vigilance will enhance the model's utility and trustworthiness in practical settings, solidifying its strength as a predictive tool.

The graph presented in Figure 7, an assumed confusion matrix, displays the model's prediction abilities. With 5900 true positives and 315 genuine negatives, the model can accurately classify situations as healthy or unhealthy. It must be admitted that the model made 315 false positives and 670 false negatives in disease diagnosis. The statistics indicate that classifier performs exceptionally well in identifying positive instances, as evidemced by the high number of true postiives (5900). On the other other, with 315 fasle postiives, the classifier oftentimes incorrectly identifies negtiave isntances as postive and the classifier misses 670 positive instances, which could be critical depending on the application context.

Performance metrics, shown in Figure 8, highlight similarities among various machine learning algorithms. The best algorithm is "Bayesian Optimization with KNN," with an accuracy, precision, and recall of 0.95, and an F1 score of 0.94, making it highly adaptable. "SVM" and "Logistic Regression" also perform well, with accuracies of 0.82 and 0.81, respectively. "Logistic Regression" excels in recall at 0.94, ideal for applications prioritizing true positives. Conversely, "Decision Tree" underperforms across all metrics, facing accuracy and precision challenges. "Random Forest" achieves balanced performance, fitting all criteria. A comparison between the hybrid model (Bayesian Optimisation with KNN) and other models show how much better the hybrid model performs overall. It is expected that data scientists and practitioners may find these insights useful in choosing the right algorithms. Table 8 provides an overview of earlier studies and sets our work apart from theirs.

6. Conclusion

Modern agriculture faces significant challenges in soil monitoring and disease prediction for crop management. Integrating advanced machine learning algorithms with IoT technology offers a potential solution, particularly for tomato crops. Using ML techniques this research reports that tomato crop disease detection performance can be improved that subsequently improve farming operations. The present study proposes a hybrid ML, which combines multiple ML techniques to improve predictive performance.

Importantly, to fine-tune the model’s hyperparameters, this study has used a sophisticated optimisation method, i.e., Bayesian optimisation with KNN. The performance metrics, as depicted in Figure 8, demonstrate that the proposed optimisation model outperforms existing methods, achieving high accuracy (95%), precision (95%), and recall (95%). Its steady F1 score of 0.94 shows its forecasting accuracy. By leveraging IoT for real-time data acquisition and hybrid ML for sophisticated data analysis, this integrated approach aims to enhance decision-making in tomato cultivation, thereby improving yield quality and quantity.

Secondly, the present study has considered tomato crop prediction from a relatively new area in south India, where no such studies have been reported to the best of the authors’ knowledge. The present research, therefore, can be considered a case study by the agriculture scientists, farmers, and other stakeholders who will benefit from it.

Coming to the future scope, to further improve the performance of these creative works, future research should concentrate on enhancing the scalability and accuracy of machine learning algorithms, consumer food security, and the integration of other data sources. Future research should also focus on expanding the dataset to include more diverse agricultural environments and disease types. Research efforts should focus on developing robust and scalable ML algorithms to handle complex soil nutrient data and integrate ML techniques with sensor networks and remote sensing (satellite-based) data to fully realise its potential in soil science.