Prediction of wheat fusarium head blight severity levels in southern Henan based on K-means-SMOTE and XGBoost algorithms

Collection of disease data and meteorological information

This study has collected basic data on the occurrence of wheat FHB in 17 cities in the southern part of Henan Province from 2016 to 2023, as shown in Fig. 1.

The wheat FHB incidence data used in this study was obtained from the Plant Protection and Quarantine Station of Henan Province, while meteorological data was sourced from the Henan Meteorological Bureau (http://ha.cma.gov.cn/). The incidence data includes the annual average diseased ear rates for wheat from 2016 to 2023 across 17 cities. The meteorological dataset encompasses seven key factors—rainfall (RAIN), average sunshine hours (ASH), average wind speed (AWS), average temperature (AT), highest temperature (HT), lowest temperature (LT), and average relative humidity (ARH)—collected from January 1 to June 1 each year across 17 cities over the same period, capturing the entire growth cycle of wheat. Based on the investigation results and according to the national technical specification for the monitoring and reporting of wheat FHB, recommended national standard (GB/T) 15796-2011, wheat FHB is classified into four levels: not occurring (Level 0, spike infection rate <0.1%, no significant impact on wheat production), mild prevalence (Level 1, 0.1% $\le$ spike infection rate <5%, causing partial yield reduction in wheat production), moderate prevalence (Level 2, 5% $\le$ spike infection rate <10%, causing partial yield reduction in wheat production), and severe prevalence (Level 3, spike infection rate $\ge$10%, causing significant yield reduction in wheat production). In addition to data on spike infection rates, this study also collects and records meteorological information for the corresponding periods, including seven key meteorological variables: daily precipitation, wind speed, average temperature, maximum temperature, minimum temperature, relative humidity, and hours of sunshine from January 1 to June 1 each year from 2016 to 2023.

K-means-SMOTE

Due to the real-world field data on wheat FHB severity levels, the issue of sample imbalance is inevitable. Without proper handling, the prediction results for the wheat FHB severity levels would be biased towards the categories with larger sample sizes, leading to larger prediction errors for categories with smaller sample sizes. In the current collected data, there are 64 instances of wheat FHB at severity Level 0, 58 instances at severity Level 1, six instances at severity Level 2, and eight instances at severity Level 3. Clearly, there is a significant difference in the number of data points for each level, with an imbalance ratio (IR) greater than 10:1 (Jing et al., 2021). For such highly imbalanced datasets, classification results often show high accuracy for the majority classes while approaching zero for the minority classes. This can lead to a decrease in the overall accuracy of the model. To address the issue of data imbalance, this study uses the K-means-SMOTE oversampling technique. According to literature (Douzas, Bacao & Last, 2018) due to the smaller number of instances in certain classes, typical classifiers tend to overlook the detection of minority classes.

SMOTE is one of the oversampling techniques used to solve this problem (Chawla et al., 2002). It generates minority instances within overlapping regions. However, SMOTE randomly synthesizes minority instances along the line connecting minority instances and their selected nearest neighbors, ignoring nearby majority instances. In contrast, the K-means-SMOTE technique applies the K-means clustering algorithm to group minority instances in the dataset and then performs oversampling within each cluster. This approach avoids the problem of SMOTE ignoring nearby majority instances. Within each cluster, the K-means-SMOTE technique randomly synthesizes minority instances along the line connecting minority instances and their selected nearest neighbors, thus maintaining the distribution characteristics of the data. This method better captures the local structure of minority class samples, thereby improving model accuracy performance. The basic idea is to balance the dataset by synthesizing new minority class samples. The algorithm of the K-means-SMOTE technique is described in Douzas, Bacao & Last (2018), with the specific steps shown in Table 1.

Computational complexity is an important indicator to measure the efficiency of an algorithm. The computational complexity of K-means-SMOTE primarily lies in the clustering step, with a time complexity of $O(n\cdot k\cdot t)$, where $n$ is the number of samples, $k$ is the number of clusters and $t$ is the number of iterations of the K-means algorithm. Although the computational cost may increase significantly with larger data sizes, this complexity can be effectively controlled through reasonable selection of the number of clusters $k$, and optimization of the algorithm.

In contrast, the traditional SMOTE and ADASYN methods have a computational complexity of $O(n\cdot k\cdot \log n)$ mainly arising from the need to find the $k$ nearest neighbors for each minority class sample. When dealing with larger sample sizes, SMOTE and ADASYN may face a higher computational burden, especially when processing datasets with high-dimensional features, as the complexity of k-nearest neighbor searches further increases.

Overall, the efficiency of the K-means-SMOTE method is more guaranteed compared to other methods.

XGBoost

In each iteration, XGBoost builds a new decision tree model based on the residuals between the previous model’s predictions and the actual labels. The prediction results of the new model are then weighted and added to the previous model results to obtain the final integrated prediction result. The XGBoost prediction model can be expressed as

(1) $${\hat{y}}_i^{(t)}=\sum _{k=1}^t{f}_k(x_i),f_k\in F.$$

In the formula: k represents the number of decision trees, F corresponds to the set of all decision trees, f_k is the $k$-th decision tree generated in the $k$-th iteration.

The loss function of the results can be represented by the predicted value y_i and the actual value ŷ_i.

(2) $$L=\sum _{i=1}^{n}l(y_i,{\hat{y}}_i).$$

The objective function O consists of the model’s loss function L and a regularization term Ω that penalizes model complexity.

(3) $$O=\sum _{i=1}^{n}l(y_i,{\hat{y}}_i)+\sum _{i=1}^n\mathrm{\Omega }(f_i).$$

Finally, according to the additive model, an overall XGBoost prediction model is obtained, as shown in Fig. 2.

Similarly, this study selects XGBoost as the model algorithm mainly due to its significant advantages in computational efficiency and optimization performance. The computational complexity of XGBoost is $O(n\cdot m\cdot \log m)$, where $n$ represents the number of samples and $m$ represents the number of features. By constructing and optimizing multiple decision trees through incremental learning based on weighted loss, XGBoost significantly improves model training efficiency. In contrast, the computational complexity of GBDT is also $O(n\cdot m\cdot \log m)$, but GBDT constructs trees sequentially and lacks parallel processing, making XGBoost generally faster in training, especially on large datasets.

The computational complexity of RF is $O(t\cdot n\cdot \log m)$, where $t$ is the number of trees, $n$ is the number of samples, and $m$ is the number of features. RF predicts by simultaneously constructing multiple decision trees, and its computational complexity increases linearly with the number of trees. Although RF has strong parallel capabilities, the construction of each tree is independent, lacking the mechanism in XGBoost that optimizes each tree through weighted loss. Therefore, XGBoost demonstrates better performance when handling large datasets, effectively reducing training time and preventing overfitting.

Model construction

Based on the principles of K-means-SMOTE and the XGBoost algorithm, the steps to establish the wheat FHB severity levels monitoring model are as follows:

1. Data Division: Based on historical data and previous research, select meteorological factors closely related to the occurrence and development of wheat FHB severity levels, and divide the data to prepare for correlation analysis.

2. Correlation Analysis: Analyze the correlation between various meteorological factors and wheat FHB severity levels using the Kendall coefficient, identify key periods that have a significant impact on the prediction of wheat FHB severity levels, and organize the corresponding dataset.

3. K-means-SMOTE Oversampling: To address the class imbalance issue in the dataset, the K-means-SMOTE algorithm is used to enhance minority class samples.

4. XGBoost Classification Prediction and Meteorological Factor Selection: Further train the model using the XGBoost model, which combines the predictive capabilities of multiple decision trees and optimizes model performance through the gradient boosting algorithm. This is suitable for handling large-scale complex data. The XGBoost model also provides a ranking of the importance of meteorological factors, allowing for the selection of more strongly correlated meteorological factors.

5. Reconstruct the Dataset and Retrain the Model Using XGBoost: Use the data of the four selected key meteorological factors as the new dataset for model training, continuously adjusting model parameters until the model achieves optimal prediction performance.

The flowchart is shown in Fig. 3.

Data division

This study processes the seven meteorological variables for each county (city, district), considering the high incidence period of wheat FHB to be around the heading and flowering stages (Zhou, 2024). For southern Henan, the period from March 1 to May 20 is considered. To accurately identify the specific stage at which each meteorological variable is most significantly correlated with wheat FHB severity levels, this study integrates all relevant data from 17 cities from 2016 to 2023, dividing the data into a total of 81 time periods.

Specifically, precipitation (RAIN) is calculated on a 15-day cycle, from March 1 to May 20 each year, with the total rainfall calculated every 15 days. For example, labels such as RAIN_3.1−3.15, RAIN_3.2−3.16, etc., up to RAIN_5.6−5.20, represent the total rainfall from March 1 to March 15, March 2 to March 16, and so on, up to May 6 to May 20.

The other seven meteorological factors are divided into 5-day periods. The total sunshine hours (ASH) for each 5-day period are calculated from March 1 to May 20 each year. The average wind speed (AWS), average temperature (AT), highest temperature (HT), lowest temperature (LT), average relative humidity (ARH), and the maximum temperature difference (MTD) between the highest and lowest temperatures within each 5-day period are also calculated. For example, labels such as ASH_3.1−3.5, ASH_3.2−3.6 , etc., up to ASH_5.16−5.20, represent the total sunshine hours from March 1 to March 5, March 2 to March 6, and so on, up to May 16 to May 20.

Then, the meteorological data for the same period from all counties are combined with the corresponding wheat FHB severity level data into one dataset. Thus, for each group of meteorological data (e.g., the group labeled 3.1–3.5), the corresponding data from 17 cities from 2016 to 2023 are aggregated to form the first dataset. This process is repeated, forming a total of 81 such datasets, each containing the meteorological information and wheat FHB severity levels data for all relevant counties within a specific time period. Before feature selection, we first conducted a statistical analysis of missing values and outliers to ensure the integrity and accuracy of the data.

Feature selection

The Kendall’s tau coefficient is a non-parametric statistical method used to measure the ordinal association between two random variables, with values ranging between −1 and 1; It is based on the number of concordant and discordant pairs in the data, providing a way to measure the strength and direction of the monotonic relationship between two variables (Bae & Kim, 2014). In practical applications, the Kendall’s tau coefficient does not require assumptions about the data distribution, has good adaptability to small samples and non-normally distributed data, and is relatively robust to outliers. The Kendall’s tau coefficient algorithm is as follows:

Suppose there are two sequences, X and Y, each with elements. The i-th element ( $1\le \mathrm{i}\le \mathrm{N}$) in the two sequences is represented by ${\mathrm{X}}_{\mathrm{i}}$ and ${\mathrm{Y}}_{\mathrm{i}}$ respectively. When any two elements ( ${\mathrm{X}}_{\mathrm{i}}$, ${\mathrm{Y}}_{\mathrm{i}}$) and ( ${\mathrm{X}}_{\mathrm{j}}$, ${\mathrm{Y}}_{\mathrm{j}}$) in the sets and have the same order, i.e., ${\mathrm{X}}_{\mathrm{i}}>{\mathrm{X}}_{\mathrm{j}}$ and ${\mathrm{Y}}_{\mathrm{i}}>{\mathrm{Y}}_{\mathrm{j}}$ or ${\mathrm{X}}_{\mathrm{i}}<{\mathrm{X}}_{\mathrm{j}}$ and ${\mathrm{Y}}_{\mathrm{i}}<{\mathrm{Y}}_{\mathrm{j}}$, these two elements are considered concordant. When ${\mathrm{X}}_{\mathrm{i}}<{\mathrm{X}}_{\mathrm{j}}$ and ${\mathrm{Y}}_{\mathrm{i}}>{\mathrm{Y}}_{\mathrm{j}}$ or ${\mathrm{X}}_{\mathrm{i}}>{\mathrm{X}}_{\mathrm{j}}$ and ${\mathrm{Y}}_{\mathrm{i}}<{\mathrm{Y}}_{\mathrm{j}}$ occur, these two elements are considered discordant. When ${\mathrm{X}}_{\mathrm{i}}={\mathrm{X}}_{\mathrm{j}}$ or ${\mathrm{Y}}_{\mathrm{i}}={\mathrm{Y}}_{\mathrm{j}}$ occur, these two elements are neither concordant nor discordant. Then the calculation of Kendall’s tau correlation coefficient is as follows:

(4) $$\tau =\frac{C-D}{\sqrt{(N_0-N_1)(N_0-N_2)}},$$where C represents the number of concordant pairs, D represents the number of discordant pairs, and N represents the total number of elements.

(5) $$N_0=\frac{1}{2}N(N-1),$$

(6) $$N_1=\sum _{i=1}^{s_1}\frac{1}{2}{U}_i(U_i-1),$$

(7) $$N_2=\sum _{i=1}^{s_2}\frac{1}{2}{V}_i(V_i-1),$$where ${\mathrm{S}}_1$ and ${\mathrm{S}}_2$ represent the number of types of repeated elements in X and Y respectively, and ${\mathrm{U}}_{\mathrm{i}}$ and ${\mathrm{V}}_{\mathrm{i}}$ represent the number of elements in the $i$-th set of identical elements in X and Y, respectively.

Using Kendall’s tau coefficient, we analyzed the correlation between different meteorological factors during specific periods and the wheat FHB severity levels. The specific approach involved applying Kendall’s tau coefficient to the 81 pre-split datasets sequentially. For each meteorological factor, we selected the period with the highest correlation, and so forth. This resulted in obtaining the data for the eight meteorological factors with the strongest correlations, along with the corresponding year’s wheat FHB severity data, thus forming a new dataset. Eight key meteorological factors that influence the occurrence of wheat FHB severity levels were identified: total precipitation RAIN_4.5−4.19, average wind speed AWS_4.24−4.28, average temperature AT_3.7−3.11, highest temperature HT_3.7−3.11, lowest temperature LT_4.19−4.23, average relative humidity ARH_4.12−4.16, average sunshine hours ASH_3.4−3.8, and maximum temperature difference MTD_5.14−5.18.

To visually display the correlation between each selected meteorological factor and the wheat FHB severity levels, we calculated the correlations using Kendall’s tau coefficient and presented the results in the form of a heatmap, as shown in Fig. 4. The absolute value of the correlation between each meteorological factor and the wheat FHB severity levels exceeded 30%, which confirms the significant role of key meteorological factors in the development of FHB. This indicates that RAIN, LT, and ARH are positively correlated with the levels of FHB, which aligns with Zhou’s (2024) explanation of the disease’s development patterns. This provides a scientific basis for further formulating targeted prevention and control measures.

Parameter selection

Parameter selection is a crucial part of the experiment, based on the oversamplers listed in Fig. 2 as benchmarks for evaluating the proposed method, as well as the hyperparameter sets used for each oversampler. This work aims to create comparability between oversamplers; therefore, it is most important for oversamplers to achieve the same IR. Since there are only six level 2 data in the dataset, K (number of clusters to be found by K-means) in K-means-SMOTE is set to 1, 2, if K is greater than or equal to 3, there may be only two data points in each cluster, which would prevent effective oversampling. In comparison to other oversampling methods such as SMOTE and ADASYN, KNN (number of nearest neighbors considered by SMOTE) is set to 1, 2, 3, 4. In order to make IR as close to 1:1 as possible, the n2 (the number of samples to be generated at level 2) of all oversampling methods is set to be 8, 9, and 10 times that of level 2 samples, and the n3 (the number of samples to be generated at level 3) is set to be six, seven, and eight times that of level 3 samples. The dataset was randomly divided into 80% training set and 20% test set to evaluate the model’s performance under different parameter settings. The specific parameter settings are shown in the Table 2 below.

To further optimize the model’s performance, this study tuned the key parameters of the XGBoost model. Specific settings are shown in Table 3. During the XGBoost model parameter tuning process, the GridSearchCV method was used to find the optimal parameter combination. Specifically, grid search traverses each parameter combination and evaluates the performance of each combination through cross-validation to find the best parameter settings.

To demonstrate the reasonableness of the parameter pool and the chosen parameters, this research conducted a parameter sensitivity analysis.

First, n_estimators was fixed at 100 and the learning_rate at 0.001, systematically evaluating max_depth from 1 to 300 with a step size of 1. The results indicated that the model achieved its highest accuracy when max_depth was set to 3, highlighting optimal performance at this depth. As max_depth increased, accuracy remained stable at four and five but began to decline after reaching six and then remained unchanged, as shown in Fig. 5A.

Afterwards, in the case of fixing max_depth at 3 and learning_rate at 0.001, the accuracy of n_estimators was evaluated from 1 to 300, which showed that the optimal solution was achieved when n_estimators was set to 13, and maintained it thereafter, as shown in Fig. 5B.

Finally, the parameters were fixed at max_depth = 3 and n_estimators = 100, and the learning_rate was varied from 0.001 to 1 in increments of 0.001. The results showed that the model reached its optimal value at the beginning and did not show any obvious pattern afterwards, as shown in Fig. 5C.

Then combine each oversampling method with each classifier for hyperparameter selection. Specifically, to achieve an IR as close to 1:1 as possible, the values for n2 and n3 are divided into three groups: n2 = 8 with n3 = 6, n2 = 9 with n3 = 7, and n2 = 10 with n3 = 8. For K-means-SMOTE, these three groups are combined with K = 1 and K = 2. For SMOTE and ADASYN, these three groups are combined with KNN = 1, KNN = 2, KNN = 3, and KNN = 4. It is important to note that the optimal results varied across different combinations. Subsequently, the parameter combinations for each oversampling method were tested with the classification models, and the combination that demonstrated comparatively better performance was ultimately selected, as shown in Table 4.

The selection of parameters is also in line with the expectations of this study, such as the final determination of KNN, as Sun et al. (2024) mentioned, blindly continuing to increase the k nearest neighbors value will undoubtedly increase the computational load, which is often undesirable while maintaining optimal accuracy. Therefore, based on these findings and experimental comparison results, this study chooses to fix the KNN value at 1.

At this point, the evaluation indicators of the model are shown in Table 5, with an accuracy of 0.8936, precision of 0.9249, recall of 0.8936, and F1 score of 0.8851. Additionally, the importance ranking of each meteorological factor is provided, in order: AWS_4.24−4.28, RAIN_4.5−4.19, MTD_5.14−5.18, ARH_4.12−4.16, HT_3.7−3.11, LT_4.19−4.23, ASH_3.4−3.8 and AT_3.7−3.11. The importance of each factor is 0.307662, 0.287486, 0.191471, 0.118757, 0.036119, 0.031564, 0.013735, and 0.013207, respectively.

Results analysis

To ensure the scientific and rigorous design of the experiment, this study also selected four other oversampling methods: SMOTE, ADASYN and HDAWCR, as well as the mainstream machine learning models RF and GBDT, and KNN, to compare with the XGBoost model.

First, using t-SNE visualization, we compared the quantity and distribution of each class of data before and after oversampling. We found that the original dataset, after different oversampling methods, showed an increase in the amount of 2nd and 3rd grade data. It is not difficult to see that the data clustering effect after K-means-SMOTE is more pronounced and better than other oversampling methods, as shown in Fig. 6. At the same time, the effectiveness of the proposed method is demonstrated through multiple experiments, and the experimental results are shown in Table 5. Based on the data in the table, it is clear that the K-means-SMOTE-XGBoost model used in this study outperforms other models in various performance metrics. Specifically, the model achieved an accuracy of 0.8936, a precision of 0.9249, a recall of 0.8936, and an F1 score of 0.8851.

Based on this, and considering the feature importance ranking in the “Results” along with the biological characteristics of wheat FHB, the severity of wheat FHB each year largely depends on variations in factors such as rainfall, humidity, and temperature (Li et al., 2021, 2022; Matengu et al., 2024). Therefore, four meteorological factors were selected—AWS_4.24−4.28, RAIN_4.5−4.19, MTD_5.14−5.18, and ARH_4.12−4.16 were selected to establish a dataset incorporating these variables. Following this, the performance of different models was re-evaluated using this dataset, according to the parameter selection method outlined in the “Results”. At this point, the classification model parameters were set as follows: learning_rate = 0.03, max_depth = 3, and n_estimators = 200. The specific experimental results are shown in Table 6.

Finally, in order to enhance wheat FHB prevention, combined with the ranking of meteorological factors in the “Results” as well as the biological characteristics of wheat FHB, this means that wheat FHB is more likely to occur during the heading and flowering stages (Xu, 2022; Pang, 2023), in the southern region of Henan, the heading and flowering stages of wheat primarily occur from mid to late April. Therefore, this study replaced MTD_5.14−5.18 with LT_4.19−4.23, by concentrating meteorological factors in April, the period of high prevalence for wheat FHB. Ultimately, four key meteorological factors that contribute to the occurrence of wheat FHB were selected—AWS_4.24−4.28, RAIN_4.5−4.19, ARH_4.12−4.16, and LT_4.19−4.23 —to establish a dataset based on these four factors. Similarly, the performance of different models was re-evaluated according to the parameter selection method outlined in the “Results”. At this point, the classification model parameters were set as follows: learning_rate = 0.001, max_depth = 3, and n_estimators = 100. The specific experimental results are shown in Table 7.

Comparing the results of Tables 5–7, it is evident that Table 6 performs better overall than Tables 5 and 7. In this case, the accuracy and recall of the K-means-SMOTE-XGBoost model in this research reached 0.9149, with precision at 0.9362 and an F1 score of 0.9103. The results in Table 6 not only validate the effectiveness of selecting meteorological factors but also demonstrate that focusing on key meteorological factors can significantly improve prediction accuracy and reliability while maintaining the simplicity of the model.

Furthermore, when concentrating on key meteorological factors during critical periods affecting wheat FHB, namely AWS_4.24−4.28, RAIN_4.5−4.19, ARH_4.12−4.16, and LT_4.19−4.23, the model shows slightly lower performance in various metrics compared to Table 6. However, compared to Table 5, the accuracy and recall remain unchanged at 0.8936, the F1 score increased from 0.8851 to 0.8898, and precision decreased from 0.9249 to 0.9058, still outperforming other models. Moreover, the selected key meteorological factors can predict wheat FHB severity levels at an earlier stage, providing a more efficient and practical model optimization strategy for accurately predicting and managing wheat FHB. For ease of comparison, line charts of predicted versus observed values and heatmaps of the confusion matrix for the K-means-SMOTE-XGBoost model before and after selecting the four key meteorological factors are provided, as shown in Figs. 7 and 8.

The significance results from the t-test comparing the K-means-SMOTE-XGBoost method with other methods in terms of accuracy, precision, recall, and F1 score are presented in Table 8. Ten experiments were conducted for both the K-means-SMOTE-XGBoost model and the comparison methods under different random seeds. A P-value below 0.05 indicates a statistically significant difference between K-means-SMOTE-XGBoost and the other methods. As shown in the table, significant differences are observed across all groups for accuracy, precision, recall, and F1 score ( $P<0.05$).

Based on the experimental results, it can be observed that the K-means-SMOTE-XGBoost method adopted in this study demonstrates excellent performance. Specifically, this study aims to develop a new wheat FHB severity levels prediction model by integrating the K-means-SMOTE and XGBoost models. By introducing the K-means-SMOTE technique, this study effectively addresses the data imbalance issue, enhancing the model’s prediction capability for high-risk FHB levels. Meanwhile, the use of the XGBoost algorithm not only improves prediction accuracy but also optimizes the feature extraction process, enabling the model to learn more decisive information from complex climate and crop health data.