Digital Assessment of Banana Weevil (Cosmopolites sordidus Germar) Resistance as Compared with Expert Visual Assessment

1.0. Experimental Setup

The study was conducted at the International Institute of Tropical Agriculture-Sendusu station in Uganda. A hundred and eighty-five tissue culture-generated plants representing 22 genotypes were evaluated in pot trials using a Partially replicated (p-rep) experimental design (Williams et al., 2014). Each genotype was replicated twice with four pseudo replications per genotype per block. Calcutta 4, a known weevil-resistant genotype and three susceptible landraces to include Mbwazirume (Highland cooking banana), Mchare (Highland cooking banana) and Obino lÉwai (plantain) were used as checks. Infestation with three female and three male weevils was at 60 days post-planting upon attainment of sufficient corm size while data collection was at 60 days post-infestation (dpi) following the cross-section method described in Gold et al. (1994).

1.1. Scoring for Weevil Damage Using Visual Observation

The corm was cut transversely, first at the collar (pseudostem and corm juncture) to score upper cross-sectional damage, and two centimeters below the collar to score lower cross-sectional damage. For each cross-section, weevil damage was assessed separately for the central cylinder and the cortex, estimating the percentage of corm tissue damaged by the weevil in each area. The exposed corm was divided into four sections. The total number of sections with weevil damage is divided by four and multiplied by 100 to calculate the percentage of cross-sectional damage. Parameters scored included (a) percentage of outer upper cross-sectional damage (cortex), (b) inner upper cross-sectional damage (central cylinder), (c) outer lower cross-sectional damage (cortex), and (d) inner lower cross-sectional damage (central cylinder). The mean of the four scores was calculated to generate the overall damage.

1.2. Scoring for Weevil Damage Using Image Analysis

1.2.1. Image Capture

Corm sections were prepared as discussed in section 1.1. High-quality images of each section showing the genotype name, block, and plot number were captured with a green background to increase contrast and optimize quality as shown in Figure 1. A ruler was incorporated as a reference in all images to ensure precise measurements using the ImageJ tool in Figure 1. The resulting images were saved as Joint Photographic Expert Group (JPEG) format. For each plant, two images representing upper and lower cross-sections were captured, totaling 370 images from 185 plants of 22 genotypes (Supplementary Table S1)

2.0. Image Processing Using ImageJ

We describe the approach taken to process the captured image data as a preparatory step before analysis by ImageJ.

2.1.1. Automated Background Removal

To facilitate the analysis of the images, the background was removed by using the rembg Python package. To improve image accuracy, the initial dull green background shown in Figure 2A was replaced with a consistent bright green background with a RGB value of 0, 128, 0 as shown in Figure 2B. RGB is an additive color model system representing red, green and blue colors of light used on a digital display screen to reproduce a broad array of colors for sensing and displaying images in electronic systems. The newly generated RGB images were preserved in their original size and saved as JPEG files with a compression quality of 95, ensuring maximum retention of image quality. Compression quality is the level of perfection in digital images achieved through data minimization without degrading image quality.

2.1.2. Differentiating Healthy and Damaged Regions

The rembg processed images were imported into the ilastik version 1.3.0 tool for labelling to train a custom image classifier. This was to create binary segmentations of the healthy and weevil-damaged tissue as well as background from the corm images. We adopted a two-fold training approach combining manual labeling and automated machine learning algorithms to precisely label healthy and weevil-damaged corm sections plus background on a subset of imported images to complete the segmentation framework and create the labelled dataset.

Using the labelled dataset, ilastik was trained to detect patterns and features characteristic of each region, allowing the custom classifier to accurately categorize and yield binary masks for the entire set of rembg-processed images. For the binary masks, each pixel received a value of 0 or 1 to represent healthy and weevil-damaged sections respectively (Figure 3). Upon completion of binary image segmentation, the trained ilastik model was saved for future analyses.

2.1.3. Quantification of Healthy and Weevil-Damaged Sections

For particle analysis differentiation, the imageJ threshold was adjusted using “Set Threshold (2,2) for healthy and (1,1) for weevil-damaged corm tissue in the binary image. To quantify healthy and weevil-damaged tissue within the binary image, precise parameters, i.e., size: pixel² 0 - infinity, circularity: 0.00 - 1.00 were set to acquire area measurements and standard deviation. These adjustments were used to analyze all ilastik-generated binary images.

2.2. Image Processing Using Machine Learning

We describe the approach taken to process the captured image data using machine learning techniques.

2.2.1. Annotations

Corm images were segmented using the YOLO (“You Only Look Once”) version 8 (Reis et al., 2024) deep learning model that classifies image pixels into different classes. YOLO is a lightweight neural network that performs object detection and segmentation of different regions in images. The image labeling was completed using Makesense (https://www.makesese.ai), a web-based annotation tool that supports drawing of annotations on images (Figure 4). Irregular polygons were drawn on the training set of images to map out boundaries of irregular regions and damaged corm sections. The resultant annotations were exported in Common Objects in Context format (COCO format) (Lin et al., 2014) and converted to YOLO txt labels for training corm segmentation using YOLOv8 model (Reis et al., 2024). YOLOv8 is implemented by Ultralytics, an open-source library that enables training and implementation of various neural networks (Jocher et al., 2023).

2.2.2. Model Training

The images were split into train, test and validation sets using 20, 5, and 5 images respectively. The ultralytics module was utilized for hyperparameter tuning during model training, with modifications such as setting model version to YOLOv8-small; image size to 640 x 640 pixels; batch size to 4; and patience for early stopping to 20 epochs. Most parameters were left as default, however, these specific adjustments were made to optimize the training process for the pre-trained YOLOv8-small model to attain good performance with a small dataset and 100 epochs used to fine-tune the banana corm dataset.

2.2.3. Image Detection Using Trained Model

Image segmentation was completed using Intersection Over Union (IOU), which measures overlap between predicted area and ground truth as a ratio of area of intersection to area of union between ground truth and predicted regions. IOU ranges from 0 to 1, with higher values indicating better predictions. This metric was used to assess model accuracy when predicting regions in an image belonging to a certain class by comparing the overlap between predicted and ground truth regions (Figure 5). The sum of precisions across different levels of prediction confidence and IOU were used to calculate mean Average Precision (mAP). A value of 0.995 indicated a near perfect segmentation performance on the validation set at mAP0.5 (Figure 6).

2.2.4. Background Removal

To remove the background from the images, a three-step python script was followed. In step 1 an input image was loaded and passed through the trained machine learning model. In step 2 the model analyzed the image and returned a prediction for the mask representing the corm in the image and in step 3 the predicted mask was combined with the input image to retain only the corm and eliminate the background in the image (Figure 7).

2.2.5. Quantification of Healthy and Weevil-Damaged Sections

After the background removal step, the images contained pixels for only the damaged and healthy sections of the corm. These sections were clustered using a pixel-based thresholding technique that used binary classification to determine whether values were larger or less than 120-threshold value, the optimum to distinguish between damaged and healthy pixels in the corm. The input images were converted to grayscale and blurring transformation to improve the detection of different sections. Two image detection thresholds were used to identify damaged sections in the grayscale images. First, the damaged sections were segmented white and healthy as black to obtain healthy section pixels and in step 2 the healthy was segmented white and damaged as black to obtain damaged section pixels. The healthy and damaged pixels were summed to obtain the entire corm pixels. Percentage damage was calculated following equation 1 below.

$$\mathrm{Damage} \mathrm{Percentage}={\displaystyle \frac{damage pixels}{cormpixels}} \times 100—\mathrm{Equation} 1$$

(1)

2.2.6. Model Deployment

The python script developed in 2.2.4 was deployed to automatically analyze all images in the dataset producing results as comma-separated values. (csv) excel file format.

2.3. Data Analysis

2.3.1. Statistical Analysis

Statistical analyses of the data were conducted using Pandas (McKinney, 2012) and Matplotlib (Hunter, 2007) python libraries. The Pandas library was used to compute means, standard deviations, and mean difference (bias value) using the formulae below:

The mean of each score method was determined using equation 2 below:

$$\mathrm{Mean} \mathrm{score} \mathrm{of} \mathrm{method} =\sum (Weevil damage values)/(Number of data points)$$

(2)

where, ∑Weevil damage values representing the sum of all weevil damage values in a particular weevil damage scoring method, either the visual or image analysis methods.

The bias value or mean difference was calculated using equation 3 below

$$\mathrm{Mean} \mathrm{difference} \left(\mathrm{Bias}\right)=mean of method A-mean of method B$$

(3)

where method A and B are either visual score or image analysis methods.

The standard deviation was computed using equation 4 below:

$$\mathrm{Standard} \mathrm{deviation} = {\displaystyle \frac{\sqrt{\left({\Sigma }_{\left(i=1\right)}^n{\left(x_i-\bar{\mathit{x}}\right)}^2\right)}}{\left(n-1\right)}}$$

(4)

where ∑ = summation, n = number of data points in the sample, x_i = represents each data point, and x̄ = the sample mean.

2.3.2. Assessing Agreement Among Score Methods

The Bland-Altman plot was used to visualize and assess agreement among the three methods. The upper and lower limits of agreement (LoA) were set at ±1.96 standard deviations to determine the range within which differences among the methods were expected to lie for 95% of observations from the mean difference (Bland and Altman, 1986). The precision of the Bland-Altman LoA is measured in terms of approximate interval width which is twice the value of margin of error or the difference between the upper and lower bounds of confidence interval (Bland and Altman, 1986).

Matplotlib library was used to generate a Bland-Altman plots (Bland and Altman, 1986) with limits of agreement (LoA) computed using equations 3 and 4 below.

Equation 3; Upper limit of agreement (Upper LoA) =$Bias + \left(1.96 * SD\_diff\right)$.

Equation 4; Lower limit of agreement (Lower LoA) =$Bias - (1.96 * SD\_diff).$Where SD_diff = Standard difference

The Numpy library was used to compute the correlation coefficient using equation 5 below:

$$\mathrm{Correlation} \mathrm{coefficient} = {\displaystyle \frac{\Sigma \left(x-\bar{\mathit{x}}\right)\left(y-ȳ\right)}{\sqrt{\Sigma \left(x-\bar{\mathit{x}}\right)2 \Sigma \left(y-ȳ\right)2}}}$$

(5)

where X and Y represent values of either visual score or Image analysis score methods; x̅ and ȳ are the means of X and Y, respectively; Σ = the sum of the products and squares over all data points.

Stats model python library was used to compute Lin’s concordance correlation coefficient (CCC) using equation 6 below (Lin, 1989). Lin’s CCC measures both precision and accuracy and ranges from 0 to + 1.

$$\mathrm{CCC}={\displaystyle \frac{\left(2\mathit{*}\mathit{r}\mathit{*}\mathit{\sigma }\mathit{x}\mathit{*}\mathit{\sigma }\mathit{y}\right)}{(\mathit{\sigma }\mathit{x}^2+\mathit{\sigma }\mathit{y}^2+(\mathit{\mu }\mathit{x}–\mathit{\mu }\mathit{y})^2)}}$$

(6)

W here r = Pearson correlation coefficient; σx and σy are standard deviations of either visual and ImageJ scoring methods, visual and machine learning scoring methods or ImageJ and machine learning scoring methods respectively; μx and μy means of either visual and ImageJ scoring methods, visual and machine learning scoring methods, or ImageJ and machine learning scoring methods respectively. Interpretation of Lin’s CCC according to McBride et al. (2005) (Table 1)

Matplotlib library was used to plot concordance correlation coefficient scatter plot (Hunter, 2007).