The study area is located in the CCB Plantation Base in Xinhuang County, Huaihua City, Hunan Province. Xinhuang Dong Autonomous County (hereinafter referred to as Xinhuang County) belongs to the mid-subtropical monsoon humid climate, with distinct four seasons, warm and humid, short cold period, long frost-free period, abundant rainfall, with an average annual temperature of 16.6°C, an average annual precipitation of 1160.7 mm, and a total sunshine between 1014.5 and 1590.2 hours. The annual frost-free period is 297.4 days. The geographical location is between 108°47’13” - 109°26’45” east longitude and 27°4’16” - 27°29’58” north latitude. It is located in the western part of southwestern Hunan, under the jurisdiction of Huaihua City, Hunan Province, at the end of the extension of the Miaoling Mountains in the Yunnan-Guizhou Plateau, mainly with mountainous topography, with a forest coverage rate of 67.5%, of which the planting area of CCB trees reaches 18 km, which is the leading variety of traditional Chinese medicine in the county. The origin is in the gentle slope land below 800 meters above sea level and with a slope of less than 25°; the soil layer is loose, and the soil type is yellow soil or red soil. The pH value is 5.5 to 7.0, the soil depth is more than 50 cm, the drainage and air permeability are good, and the organic matter content is ≥1.0%. The superior geographical and climatic conditions, suitable temperature and precipitation, and vast mountain resources provide an ideal growth environment for the growth of CCB trees, and provide sufficient raw materials for the production and development of Xinhuang CCB.

The data collection in this study was carried out using a DJI Mavic 3M UAV, as depicted in Figure 1 .

This professional UAV is equipped with a 20-megapixel visible light camera and four 5-megapixel multispectral cameras, facilitating high-precision aerial surveys, crop growth monitoring, and natural resource investigations. The UAV also utilizes multispectral sensors to capture solar radiation data, which is stored in image files, improving the accuracy and consistency of NDVI (Normalized Difference Vegetation Index) results. Equipped with RTK (Real-Time Kinematic) technology, it provides centimeter-level precision positioning. The UAV’s specifications include a bare weight of 951 grams (including propellers and RTK module), a maximum takeoff weight of 1050 grams, dimensions of 347.5mm * 283mm * 139.6mm. Its flight time is up to 43 minutes, and the longest hovering time is 37 minutes, allowing a single flight to cover up to 3000 acres for mapping tasks. The UAV offers adjustable flight speed and stable flight, with the capability to hover precisely, making it suitable for repeated, multi-scale, and high-resolution data acquisition of crop pests and diseases stress at specific points.

In Xinhuang County, a severely pests and diseases affected area within the CCB plantation base was chosen for UAV multispectral and RGB visible light image acquisition. The aerial imaging mission took place on July 28, 2023, at 3:00 PM local time, with ideal weather and a soft breeze. The UAV operated in a constant-altitude patrol mode, maintaining a downward-facing camera lens. During the flight, the conditions were favorable, with an altitude of approximately 80 meters and intermittent hovering for image acquisition. Due to relaxed specifications for spectral range and pixel center alignment, no radiometric calibration was applied to the images. The captured visible light images consisted of three fundamental bands: red, blue, and green. Following preprocessing with DJI-Terra software, the RGB data generated a seamless mosaic of the entire region, as depicted in Figure 2 , where darker shadow areas denote the mountain’s rear slopes.

As described in following Figure 3 , three sub-regions within the image are chosen for algorithmic study, which clearly show plants damaged by pests and diseases. By manually delineating the exact locations of the affected plants, the gray scale values for these regions are set to 0, whereas the gray scale values for the healthy plant areas are set to 255. As a result, a binary ground truth image is created, which can be used for both training the algorithm and assessing accuracy metrics.

Based on a variety of RGB-derived vegetation indices, this paper proposes a new vegetation index based on BWO and a new ensemble learning strategy to achieve precise detection of trees with pests and diseases. The specific technical route is shown in Figure 4 below.

To begin with, high-resolution RGB image data was collected using an UAV and underwent essential preprocessing for quality assurance, including image matching, image mosaic, image dodging and so on. Following this, various vegetation indices were extracted from the images, including Normalized Green-Red Difference Index (NGRDI), Green Leaf Index (GLI), Red-Green Ratio Index (RGRI), and Excess Green Index (ExG), along with others, to describe vegetation coverage and health status. These computed indices were subsequently optimized using random forest variable importance and correlation coefficients iterative analysis. The BWO algorithm was employed in a small region of the image, with the Critical Success Index (CSI)s index after single-threshold segmentation set as the optimization objective, to construct a new vegetation index based on the optimization algorithm. The optimized weights were subsequently applied to the entire image. Ultimately, based on the ensemble threshold approach, a new ensemble strategy based on BWO was proposed, generating the final pests and diseases detection results.

Currently, vegetation indices are mainly based on visible and near-infrared wavelengths, including the NDVI and the Ratio Vegetation Index (RVI). The remote sensing imagery necessary for these indices is usually characterized by high procurement expenses, limited real-time capabilities, and a coarse spatial resolution. In this study, 24 different vegetation indices were computed from visible light imagery obtained via UAV. By combining these indices with the original red, green, and blue (R, G, B) images, their potential application was investigated in the surveillance of pests and diseases affecting CCB trees.

The mathematical formulas for calculating indices within the visible light bands are detailed in Table 1 .

where the red, green, and blue bands (R, G, B) correspond to pixel values of their respective wavebands, while r, g, and b signify normalized values for each.

Each vegetation index represents a distinct plant response to specific environmental conditions. Due to the use of normalized RGB bands in certain indices, leading to inconsistent units, this paper normalized the 24 vegetation indices derived from RGB images and created grayscale images for uniformity. All subsequent analyses were carried out on these normalized images, as depicted in Figure 5 . The pseudo-color image results reveal that indices such as the GLI, RGRI, and EGRBDI tend to be more sensitive to plants suffering from pests and diseases. The ExGR stands out in identifying healthy vegetation. On the other hand, indices like the IKAW and WI may not be as efficient in reflecting plant health. By integrating the responses of these indices to the healthy and unhealthy plants, the accuracy of pests and diseases monitoring can be significantly improved. This holistic approach leverages the complementary nature of diverse vegetation indices, facilitating more precise detection and evaluation of pests and diseases impacts in remote sensing surveillance.

Threshold segmentation is a region-based image segmentation technique that categorizes image pixels into various groups. This technique, known as image threshold segmentation, is a traditional and frequently used method in image segmentation, and it has become the most basic and widely applied segmentation technique due to its simplicity, low computational cost, and stable performance. The objective of image threshold is to partition the set of pixels according to grayscale levels, resulting in subsets that form regions corresponding to real-world objects. These regions have consistent properties within them, while adjacent regions do not share such consistency. This partitioning is achieved by selecting one or more thresholds based on the grayscale levels.

The fundamental approach of threshold entails initially setting a threshold, after which all pixels are dichotomized into two classes based on the relationship between their feature values and the established threshold. If a pixel’s feature value surpasses the threshold, it is labeled as the object class; otherwise, it’s categorized as the background. By judiciously choosing the threshold, one can effectively isolate the image’s subject from its surroundings. Given an original image denoted by f(x,y), a feature value T is identified within the image, leading to the segmentation of the original into two distinct components, forming the resulting segmented image.

If b 0 = 0 , b 1 = 1 , this refers to the image binarization process.

This paper utilizes threshold techniques for the automatic segmentation of various vegetation index images to extract pests and diseases regions. The following nine threshold methods are employed.

Random Forest is a popular ensemble learning technique extensively applied within the field of machine learning. It addresses a given problem by building numerous decision trees and arrives at the ultimate classification outcome through a voting mechanism. In contrast to alternative approaches, Random Forest stands out for its user-friendly nature, high robustness, and its ability to avoid overfitting.

Feature importance evaluation is used to calculate the importance of sample features, quantitatively describing the contribution of features to classification or regression. Random forests can be used to assess feature importance, which, from another perspective, is a built-in tool of random forests, mainly divided into two methods: (1) Mean Decrease Impurity (MDI), which measures the importance of a node by statistically calculating the decrease in impurity when the node is split; (2) Mean Decrease Accuracy (MDA), which involves randomly permuting the values of a feature in the out-of-bag (OOB) data set and then repredicting, calculating the importance of the feature by measuring the degree of decrease in classification/regression accuracy. MDI uses training set data and can directly obtain MDI feature importance assessment values after RF training is completed; MDA uses OOB data and requires running a separate feature importance evaluation program after RF training is completed. Since MDI uses training data exclusively, it may affect the accuracy of the assessment and is more inclined to increase the weight of high cardinality features. In comparison, the results of MDA are more accurate, which is used in this paper.

Feature selection plays a pivotal role in the Random Forest algorithm, involving the evaluation and ranking of the significance of various features. Let’s consider a dataset S that includes m samples and P feature variables, with y being the class label associated with each sample. The Random Forest algorithm constructs T decision trees through T iterations of bootstrap sampling. Due to this sampling approach, not all samples are used in the building of each tree; the ones that are not used are referred to as OOB samples. The OOB error e is determined by validating against the OOB samples, and this error is recalculated after randomly altering a particular feature. The degree of change in the OOB error, both before and after the feature modification, reflects the importance of that feature. The importance metric J for feature x can be formulated as follows:

where the predicted label output by the model is denoted by h k ( i ) , the genuine label from the validation set is represented by y i , and any changes to the feature x j lead to the modified prediction indicated by h k x ¯ j ( i ) .

In this experiment, a 300 x 300 pixel area was chosen from a 1000 x 1000 pixel image to record the values of 27 indices. These 27 indices act as independent variables, while the classification outcomes of the ground truth map, denoted by 0 or 1, serve as the dependent variable. Utilizing the Random Forest algorithm, the importance score for each of these features was determined and subsequently the 27 features were ranked in order of decreasing importance.

In dealing with the issue of multicollinearity, the CCIA provides an effective and intuitive solution. The core idea of this method is to identify and remove highly correlated feature pairs within the feature set. By iteratively selecting and eliminating variables with high correlation, it is possible to gradually construct a feature set with low multicollinearity, thereby enhancing the interpretability and predictive performance of the model.

The step-by-step iterative analysis can be detailed as follows: Initially, a correlation coefficient matrix for all features is generated to evaluate their linear associations. The matrix then highlights feature pairs with correlation coefficients exceeding a set threshold. For each pair of highly correlated features, one feature is chosen to keep while the other is eliminated to reduce collinearity. This iterative elimination continues until the correlation coefficients of all feature pairs fall below the threshold. After the removal of features with high correlation, the remaining ones are used to construct a multiple linear regression model that minimizes collinearity.

Mathematically, the correlation coefficient between feature i and feature j can be calculated using the following formula:

where x i k and x j k signify the values of the i-th and j-th feature for the k-th data point, whereas x ¯ i and x ¯ j denote their corresponding mean values. The computation of correlation serves to manage the issue of multicollinearity among features effectively during feature selection. By using this method, it can be ensured that multicollinearity among features is effectively controlled during the feature selection process, thereby laying the foundation for constructing a robust regression model.

The BWO (Zhong et al., 2022) is inspired by the behavioral traits of beluga whales. It models the whales’ activities of swimming, foraging, and diving through a three-phase framework: exploration, exploitation, and whale diving. Beluga whales are operational as search agents, navigating the search space by adjusting their position vectors. During the exploration phase, random selection of whales ensures a comprehensive search across the scanning area. In the exploitation phase, the algorithm focuses on local search within the scanning space, utilizing the charged flight technique to improve convergence. Following the completion of exploration and exploitation in each iteration, the algorithm transitions to the whale diving stage of optimization.

Since the BWO algorithm is based on a population mechanism, beluga whales are regarded as search agents here. Each beluga whale acts as a candidate solution during the optimization process and is constantly updated. The position matrix of the search agents is modeled as follows:

where the number of beluga whales is denoted by n, and d represents the dimensionality of the problem variables. The fitness values corresponding to all beluga whales are stored as follows:

The BWO algorithm adjusts the transition from exploration to exploitation using the balance factor Bf, as defined by the equation:

where t represents the current iteration count, T denotes the total number of iterations, and B 0 is a random value between -5 and 5, which varies at each iteration. When the balance factor B f is greater than 0.5, this indicates an exploration phase, while a B f value less than or equal to 0.5 indicates a development phase.

(1) Exploration Phase

The exploration phase of the BWO algorithm is inspired by the swimming behavior of beluga whales. The position of the search agent is determined by the cooperative swimming of the beluga whales, and the position update of the beluga whales is as follows:

where x i , j t + 1 denotes the position of the i-th individual in the j-th dimension at the next iteration; if the problem dimension is assumed to be D, then p represents a random integer within the range [1, D], thus, x i , j t denotes the position of the i-th individual in the randomly selected dimension p under the current iteration; r is also a random integer, where the population size is assumed to be N, and r is a random integer within the range [1, N], thus, x r , p t denotes the position of the randomly selected individual r in the randomly selected dimension p under the current iteration; and r1 and r2 are both random numbers between (0, 1).

(2) Development Phase

The inspiration for the development phase of the BWO algorithm is derived from the feeding behavior of beluga whales. In order to improve the algorithm’s convergence, the Levy flight strategy is incorporated in the development phase, assuming that beluga whales utilize this flight strategy to capture prey. The mathematical model for this strategy is represented as

Where: x i t and x r t are the current positions of the i-th beluga whale and a random beluga whale, respectively; x i t + 1 is the new position of the i-th beluga whale; x b e s t t is the best position of the beluga whales; r 3 and r 4 are random numbers between 0 and 1; C 1 = 2 × r 4 ( 1 − t T ) is the random jump intensity; and L F is a random number following the Levy distribution, which is calculated as follows:

where both μ and V represent random numbers that follow a normal distribution, while β is a constant with a value of 1.5.

(3) Whale Falling Phase:

In order to maintain a constant population size, a position update formula is created by utilizing the current location of the beluga whale and the step size of the whale’s descent.

where r 5 , r 6 and r 7 are random numbers between (0, 1), x s t e p denotes the step length of the whale’s fall, C 2 = 2 W f × n is a step factor related to the probability of the whale’s descent and the population size; U b and L b are the upper and lower bounds for the variable. It is evident that the step length of the whale’s descent is influenced by the problem variable boundaries, the current iteration, and the maximum iteration count.

The probability of the whale falling is designed as a linear function.

More detailed implementation information could be found in Ref (Zhong et al., 2022).

After RFVI or CCIA, a certain number of vegetation indices are selected as input for the subsequent BWO. Assuming the selected vegetation indices are denoted as < V I 1 , V I 2 , … , V I n > , training region images and their corresponding ground truth value regions are chosen, and the BWO algorithm is applied to the n vegetation indices with weighting. For this experiment, the value of n can be 5, 10, 15, 20, or 27. The population size is set to PS, and the initial position of each whale is set to < w 1 , w 2 , … , w n > . The new vegetation index is then defined as:

After acquiring PS new vegetation indices, a specific threshold segmentation algorithm, for instance, OTSU, is chosen to segment these indices. The outcomes of this segmentation are then matched against the corresponding ground truth images to compute the accuracy assessment metric CSI of training area, generating PS CSI values. The CSI values act as the fitness measure, guiding the beluga whales’ behavior to shift between exploration and exploitation modes based on the balance factor B_f . The whales’ actions are fine-tuned, and the optimal candidate is selected based on the descent probability W_f . This process is iterated until the set number of cycles is completed, pinpointing the most suitable candidate as the definitive weight. This weight is then applied to the aggregate images of the n vegetation indices to derive the final new vegetation index, following the formula for the new index. It’s crucial to note that different threshold segmentation algorithms will result in different weights and new vegetation indices, providing essential variables for subsequent ensemble learning methods.

Since it is not possible to determine in advance how many variables to choose for better suitability, following the approach of most studies, multiple variable combinations are selected to analyze the impact of the number of variables on accuracy. Specifically, this paper sets the number of beluga whales to 50, with the dimensionality of the individual whale position vector being 5, 10, 15, 20, and 27, corresponding to the number of selected vegetation index variables. The search space range is from -10 to 10, and the number of iterations is fixed at 50. Following this, nine threshold segmentation techniques are utilized to refine the BWO of the n new vegetation indices, which are then integrated to facilitate the precise monitoring of trees affected by pests and diseases.

Ensemble learning involves training multiple fundamental classifiers and combining them either sequentially or in parallel. This is followed by the application of a particular ensemble strategy to accomplish the learning task, with the goal of minimizing variance, bias, and enhancing prediction accuracy. The main challenge in ensemble learning is the choice of the ensemble strategy, with common approaches being averaging, voting, and learning methods. Averaging is appropriate for numerical regression prediction issues, whereas voting and learning methods are frequently applied to classification tasks. Voting, a strategy often employed in ensemble learning recognition tasks, adheres to the “majority rules” principle. It encompasses absolute majority voting, relative majority voting, and weighted voting. For a given task, an integrated model F ( x i ) is constructed based on the voting mechanism, with the assumption that the ensemble consists of M basic classification outcomes { f 1 , f 2 , … f M } . The final classification category for each pixel is determined by selecting the category that receives the most votes, which constitutes the “majority s voting” strategy.

Within the scope of this paper, the fundamental classification outcomes are the detection images derived from the threshold segmentation technique. Given the varied performance of these fundamental classification outcomes, this paper proposes a BWO-weighted voting approach. By utilizing BWO to assign weights to each fundamental classification outcome and optimizing these weights through a fitness function, the ultimate voting result is ascertained by employing a majority voting strategy.

Here are the detailed steps: Nine distinct threshold segmentation techniques are applied to the new vegetation index, generating nine foundational classification results. The nine base classifications from the sub-region employed in the creation of the new vegetation index were taken as inputs for the BWO algorithm, along with parameters like the balance factor B f , fall probability W f , population size, and iteration count. The fitness function is defined by the CSI values. With each iteration, the weights for the base classifications are adjusted based on the CSI values. Unlike traditional BWO, the weights in this study are integers ranging from 0 to 100. For each beluga whale, a number of copies of each base classification equal to its weight are made; for example, a classification with a weight of 50 is replicated 50 times. After all classifications are duplicated, a comprehensive set ∑ I = 1 9 w i including all classifications is formed, and a majority voting mechanism is applied to each pixel within this set to produce the final ensemble outcome. The CSI value between this ensemble outcome and the actual reference data is calculated to assess the fitness of the individual. Depending on the balance factor, the algorithm advances to either an exploration or exploitation phase, experiences the fall process, and ends after the maximum number of iterations is reached. The best individual and its corresponding weights are then chosen as the final weights. These final weights are then applied to the large-scale image made up of the base classifications, and the ultimate pest and disease detection result is obtained through majority voting for each pixel.

In the assessment of classification outcomes, the prevalent evaluation technique is the confusion matrix (Friedl et al., 2010), often referred to as the Error Matrix based on Sample. It is derived from the counts of correctly classified pixels per class, misclassified pixels to other classes, and undetected instances. This matrix functions as a benchmark to evaluate the precision of pixel classification against actual ground-truth categories. This study employs the confusion matrix for validation and performance evaluation, with a thorough examination of errors in pests and diseases classification carried out through metrics such as: Probability of Detection (POD), False Alarm Rate (FAR), CSI, Overall Accuracy (OA), Kappa Coefficient. Given that the pests and diseases detection in the present investigation is a binary classification, the corresponding confusion matrix is presented in Table 2 .

(1) CSI, also known as Accuracy, is a measure used to evaluate the precision of identifying positive cases. CSI denotes the ratio of correctly classified positive cases to all positive cases:

(2) POD, alternatively known as Sensitivity or True Positive Rate (TPR), evaluates the capability to accurately recognize positive cases. POD signifies the ratio of samples accurately classified as positive within the total population of actual positive samples:

(3) FAR, also known as the Miss Rate or False Positive Rate (FPR), is a metric used to evaluate the error rate in identifying negative instances. It indicates the percentage of negative samples that are incorrectly labeled as positive relative to the total number of actual negative samples:

(4) The OA serves as a quantitative measure representing the extent to which every individual sample is accurately assigned to its corresponding category:

(5) The Kappa coefficient is a metric of the accuracy of the overall image classification result, with a maximum value of 1. A larger Kappa coefficient indicates a greater degree of consistency between the classified results and the ground truth.

where,

The correlation analysis is a quantitative technique that investigates the interdependence of data, evaluates the multicollinearity among indices, and furnishes a theoretical foundation for feature extraction. This paper computed the correlation coefficients for the 27 normalized attributes (comprising 24 vegetation indices and the three RGB bands), resulting in heatmaps displayed in Figure 6 for visual representation of these relationships.

The red color in the heatmap indicates a higher positive correlation, while the blue color indicates a higher negative correlation, and green represents a lower correlation. 5, 10, 15, and 20 features are selected for further iterative optimization, and the generated heatmap are shown in Figure 7 .

To assess the performance of 27 variables (comprising 24 vegetation indices and the R, G, B bands) in monitoring pests and diseases of CCB, a dataset was assembled from 300x300 sub-regions within the study area. Employing a random forest model, the relative significance of these 27 variables was determined. Initially, 27 distinct feature images were generated for this small region as independent variable, with the disease map serving as the dependent variable, represented by binary values (0 and 1). The random forest algorithm was employed to calculate the importance score for each variable, and the features were subsequently arranged in descending order of their importance. The results are illustrated in Figure 8 .

According to Figure 8 , ExGR and COM1 are the top two important features, considerably surpassing the other indices, which implies that these two features are more crucial for classification. Close behind are CIVE and VEG, with eight other indices, while B, G, and 15 other indices have relatively low importance scores, indicating their limited influence on accurate classification.

To verify the efficacy of the proposed BWO-weighted new vegetation index, the RFVI was utilized to select the top 10 most important indices and the CCIA was used to choose the 10 least correlated indices. By employing the OTSU algorithm, binary classification maps were generated for these 10 indices, representing pests and diseases detection. The BWO was then applied to adaptively weight these 10 features, and the resulting new vegetation index images were segmented using OTSU adaptively to produce the final pests and diseases detection map.

Figures 9 , 10 illustrate the OTSU threshold segmentation outcomes for the sets of 10 indices chosen by RFVI and CCIA, respectively. The black areas in these figures represent detected pests and diseases plants. It is apparent that, when compared to the ground truth, the direct segmentation of the original vegetation index yields unsatisfactory results, with significant misclassification and false negatives. Conversely, the segmentation method using the 10 weighted indices through BWO (i.e. BWO-weighted new vegetation index) surpasses the single vegetation index threshold, as demonstrated by the substantial reduction in false negatives and a closer distribution and positioning of pests and diseases trees, as shown in Figures 9L , 10L .

Tables 3 , 4 display the calculated accuracy metrics for the detection results. The direct OTSU segmentation of the indices is not satisfactory, as shown by a low CSI score, elevated POD and FAR values, indicating a propensity for over-segmenting pests and diseases pixels, signified by a low Kappa coefficient. The OTSU threshold segmentation of BWO-weighted new vegetation index outperforms the single vegetation index approach across all evaluation metrics, demonstrating that BWO successfully amalgamates valuable information from each index, generating new index capable of distinguishing pests and diseases areas from healthy ones. The distinctiveness of the indices is crucial, as each index captures a unique attribute, leading to numerous false positives in single-index OTSU segmentation. The results also underscore the advantages of BWO-weighted new vegetation index segmentation. The RFVI-selected indices exhibit higher precision in both OTSU segmentation and BWO-weighted new vegetation index processing compared to the ten chosen by the CCIA, suggesting that RFVI better reflects the relative performance among multiple indices.

In order to evaluate the detection accuracy of disease trees of multiple threshold segmentation algorithms under multiple indices conditions, based on the 10 indices selected respectively by RFVI and CCIA method in this paper, 9 threshold segmentation methods were used for automatic segmentation to obtain the detection results of pests and diseases trees for each index under each threshold segmentation method. Then for each threshold segmentation method, the average value of the detection accuracy of the 10 indices under this method was calculated as the average detection accuracy of this threshold segmentation method. At the same time, based on the 10 indices, the detection accuracy of the BWO-weighted new vegetation index by each threshold segmentation method was calculated.

The Figures 11 , 12 present the results of single-index and single-threshold segmentation for individual indices, as well as the segmentation outcomes of the BWO-weighted new vegetation index. The individual indices are chosen from the top 10 indices derived from RFVI and CCIA. From the figures, it can be observed that, irrespective of whether the indices are selected by RFVI or CCIA method, using any single-threshold segmentation method will result in severe noise and false negatives, primarily manifested in an excessive number of black areas (i.e., the pests and diseases trees) and a significant difference from the ground truth. However, the BWO-weighted new vegetation index, regardless of the single-threshold segmentation method used, can effectively segment the pests and diseases trees with a position similar to the ground truth, although there is still some noise, but its intensity is relatively low.

The quantitative evaluation results are shown in Tables 5 , 6 as follows. The average detection accuracy of 9 threshold segmentation methods on 10 indices selected by the RFVI and the CCIA was calculated and expressed as “Th_Avg”, as well as the detection results on the BWO-weighted new vegetation index, represented as “Th_BWO”. It can be seen from the table that it is consistent with the subjective evaluation results in Figures 11 , 12 . For the 10 indices selected by the RFVI, the average accuracy of each threshold method on all indices is relatively similar. The CSI index value is distributed around 0.30. The lowest is the LDT method, only 0.2579, and the highest is WAT, reaching 0.3974. Except for WAT, the POD values of all other methods exceed 0.95 and even reach 0.99, while the FAR also exceeds 0.6, indicating that there are a large number of false alarms in the detection results. The Th_BWO results are significantly better than Th_Avg, the CSI value exceeds 0.55, the OA exceeds 0.85, and the FAR remains at about 0.3, indicating that there are still certain false alarms, but the degree is much lighter than Th_Avg. For the 10 indices selected by the CCIA method, the overall trend is similar to the indices selected by the RFVI, but on each accuracy metric, the indices selected by the CCIA method are worse than those selected by the RFVI.

In the previous experiments, 10 indices were selected for threshold segmentation and BWO-weighted new vegetation index respectively by using RFVI and CCIA method. How will the selection of different numbers of indices affect the detection results? To solve this problem, in this section, the number of indices was set to 5, 10, 15, 20, and all features, and 9 threshold segmentation methods were used respectively to calculate the average value of the detection accuracy in each case of the number of indices, denoted as “Avg”. Then, in the same case of the number of indices, BWO-weighted new vegetation index was calculated, and 9 threshold segmentation methods were used to calculate the average value of the detection accuracy of the new index, denoted as “BWO”. The results are shown in Figure 13 below:

The above figure shows that for both RFVI and CCIA methods, the number of indices has some impact, but it is relatively stable. In terms of trends, for “BWO”, the highest value is reached when there are 10 indices for both RFVI and CCIA methods, and the value decreases as the number of indices increases. This indicates that on this dataset, 10 indices already contain information that can better reflect the characteristics of pests and diseases, and more indices may introduce some redundancy and conflict, leading to a decrease in detection accuracy. For “Avg”, with the increase in the number of indices based on RFVI, the detection accuracy first increases and then decreases, while for the indices based on CCIA method, the detection accuracy increases with the number of indices. However, whether the trend is increasing or decreasing, the detection accuracy of “Avg” always remains at a relatively low level, far inferior to “BWO”, which also demonstrates the significant advantage of BWO.

Ensemble learning is capable of synthesizing information from various components to achieve better results. In previous experiments, this paper has demonstrated that the BWO-weighted new vegetation index can significantly outperform a single index in terms of detection accuracy when applied to any adaptive threshold segmentation method. In this experiment, the performance differences between the conventional majority voting strategy and the BWO weighted voting strategy are examined. For multiple indices selected through RFVI and CCIA, each index is first segmented using nine threshold segmentation methods, resulting in a series of segmentation outcomes. For example, 10 indices yield 90 segmentation results, denoted with the suffix “_Ths”. These segmentation results are then ensembled using the majority voting strategy and the BWO weighted voting ensemble strategy to obtain the final detection results. As for the BWO-weighted new vegetation index, for a certain number of indices, a new vegetation index is obtained, which is segmented using nine threshold methods, resulting in nine segmentation outcomes, denoted with the suffix “_BWO_Ths”. These are then ensembled using the majority voting strategy and the BWO weighted voting ensemble strategy to achieve the final detection results. The accuracy of various strategies is evaluated using CSI, OA, and Kappa, with the results shown in Figure 14 below:

It can be seen from the figure that, on the whole, whether it is CCIA or RFVI, the BWO weighted majority voting strategies (d)-(e) and (j)-(l) can have higher accuracy than the direct majority voting strategies (a)-(c) and (g)-(i) in various numbers of indices. Different numbers of indices have a significant impact on the final accuracy. For CCIA and RFVI, the highest accuracy of direct index segmentation occurs in the first 5 indices, while the highest accuracy of segmentation using the BWO-weighted new vegetation index occurs in the first 10 indices, meaning that selecting a few indices with the least correlation or the greatest importance often achieves higher detection accuracy than using all indices. By comparing (a) and (d), or (g) and (j), it can be seen that using the BWO-weighted new vegetation index can achieve significantly higher accuracy than the direct segmentation of indices, indicating the obvious advantage of the BWO-weighted new vegetation index proposed in this paper. In terms of the ensemble strategy, the BWO weighted majority voting strategy is superior to the direct majority voting strategy in most index selection methods and quantities, showing certain advantages. It should be pointed out that the optimization goal of BWO in this paper is the CSI index. The definition of the CSI index does not concern the situation where both the true value and the predicted value are TN, while OA and Kappa do consider it. When CSI reaches the maximum, it is also possible that OA and Kappa are not optimal. Therefore, there are situations where the values of direct majority voting in figures (e)(f)(k)(l) are higher than those of BWO weighted majority voting. However, in general, the three accuracy indicators still show a high degree of consistency.

In order to substantiate the effectiveness and generalizability of the approach outlined within this paper, a rigorous test were conducted by deploying the BWO-weighted newly formulated vegetation index and the BWO-weighted ensemble strategy methods. These methods were applied to a set of 10 indices derived from the RFVI, which had been previously trained in the earlier section of our paper. The purpose of this exercise was to assess how well the previously proposed method would perform when introduced to new and unseen scenarios. The outcomes of this test, which provide a visual representation of the method’s adaptability and accuracy, are depicted in Figure 15 . This figure presents a comprehensive view of the results, allowing for a clear evaluation of the proposed method’s ability to generalize across different settings and conditions. Among them, (a) and (d) are the original RGB images, which show some CCB trees withered due to pests and diseases. (b) and (e) are the fine position ground truth maps of pests and diseases trees manually marked out, and (c) and (f) are the detection results of the method in this paper. It can be seen that, overall, the detection results have a very high similarity with the ground truth maps, especially the details of the pests and diseases trees. Although there is still some noise, the pests and diseases trees can basically be accurately detected.

The five accuracy metrics of CSI, POD, FAR, OA, and Kappa of the two scenarios were calculated based on the confusion matrix, and the results are shown in Table 7 below:

Comprehensive analysis of the data of two different scenarios reveals that the model trained by the previous dataset shows a high-precision recognition ability in other scenarios, with little difference in accuracy from the previous experiments, and can effectively identify the pests and diseases areas in RGB images. It is worth noting that the OA index not only assesses the recognition accuracy of the pests and diseases areas but also considers the non-pests and diseases areas, and the latter usually has a wider range. Therefore, the value of the OA index is very high. Overall, despite the challenges of new scenarios, the model still performs well, showing good generalization ability.