Disease diagnosis in banana leaves: a review on AI powered techniques

How effective is IP in detecting and classifying diseases in banana leaves?

In Jianqing et al. (2022a) an image processing technique was demonstrated to identify sigatoka and banana gray leaf spot. The technique retrieved color features using color segmentation, YUV conversion (Y-Luminance, U-Blue chrominance, V-Red chrominance), and Otsu thresholding. Using a Euclidean distance classifier, it obtained 90% accuracy for sigatoka and 91.7% accuracy for gray leaf spot. Jianqing et al. (2022a) proposed a method for segmenting images that combines an area threshold approach, Otsu segmentation, and color segmentation. This approach processes the RGB images that comprise the U component using an “AND” operation. With a private image dataset of banana leaf disease, the method achieved a 2.3% error rate and 97% segmentation accuracy. Correa et al. (2021) utilized an IP method for diagnosing the diseases in banana leaves. The images of both infected and non-infected leaves with conditions like bacterial wilt and black sigatoka were taken from a public source. The research covers the areas of segmentation, the Otsu method, histogram equalization, masking, and classification. The model achieved a recognition rate of 96.26% using the techniques. Deenan, Janakiraman & Nagachandrabose (2020) used various image segmentation algorithms, including grayscale conversion, image smoothing, local gradient operator, and edge localization. The outcomes of these algorithms were also computed based on Mean squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM). The diseases found are the pseudostem borer, panama disease, yellow sigatoka, rhizome rot, banana bract mosaic virus, and banana bunchy top virus. Figure 8 determines the workflow of IP techniques.

The study (Prabhakar & Sudha, 2024) proposes a filtering algorithm for disease detection in banana plant leaves. This involves many pre-processing methods, various segmentation techniques, and different feature extraction methods. The filtering algorithms used on the images are the Canny filtering algorithm, the Prewitt filtering algorithm, and the Roberts filtering algorithm. Other IP tools applied to the study are MATLAB and Python. The diseases focused are cordana, sigatoka, and pestalotiopsis.

Bhamare & Kulkarni (2013) concentrated on identifying the black sigatoka disease in banana trees through IP methods. The captured images underwent processing using Laplacian, median filtering, Gaussian, and linear. Image pre-processing involves operations like resizing, filtering, segmentation, and morphological transformations. Furthermore, for object extraction, techniques such as background subtraction, filtering, and segmentation were utilized. The study compared (Abisha & Bharathi, 2023) several models for image segmentation, including morphological transformation, contour detection, sobel filtering, k-means clustering, edge detection, canny edge detection, and thresholding, to determine the best technique. A collection of 1,600 images, such as those from the Plant Village dataset, was used for this analysis. Additionally, Gaussian blur was employed as a technique in IP. The results achieved were an RMSE value of 0.0013, a PSNR value of 57.41, and an SSIM value of 0.9982. Research by Patil et al. (2024) introduced an image-processing method for identifying the leaves of banana plants. Diseased leaves like xanthomonas, fusarium, and sigatoka were gathered. Features were extracted using multiple layers of CNN. Additionally, an Android app was also developed to help the farmers. The study referenced in Ibarra, Rivera & Manlises (2023a) examines the proportion of leaf degeneration in banana leaves affected by Panama disease. It utilizes color space transformations and OpenCV tools to segment and analyze images of the leaves. Through image processing, this method quantifies the extent of degradation by calculating the percentage of diseased areas. The researcher (Lin et al., 2021) introduces EM-ERNet, a new neural network specifically designed for banana disease recognition. The network is built on the ResNet backbone and uses dilated and multi-scale convolutions to enhance feature extraction. Innovations include batch normalization and a fusion mechanism powered by an optimized Extreme Learning Machine (ELM) algorithm. The method achieved high accuracy rates of up to 96.39% for the different banana diseases, and with faster processing times than in traditional models such as ResNet50. Non-Subsampled Dual-Tree Quaternion Wavelet Transform (NDQWT) in tandem with an Neighborhood Threshold Pattern (NTP) to classify banana leaf diseases is applied (Mathew, Kumar & Cherian, 2023a). In NDQWT, features were extracted while retaining direction and spatial form after local textures were recorded via NTP. The suggestion is that this technique will improve the accuracy of diagnosing different neglects of banana leaves. The study by Elinisa et al. (2025) address diseases like fusarium wilt and black sigatoka by utilizing 18,240 photographs of banana leaves and stalks taken by mobile phone cameras (Elinisa et al., 2025). The model uses the CNN architecture for the U-Net to segment images of banana plants to mark infected regions for early diagnosis. The U-Net model achieved a dice coefficient of 96.45% and IoU of 93.23%. The study by Mahendran & Seetharaman (2022) used Gray-Level Co-occurrence Matrix (GLCM) for texture feature extraction and classification utilizing Deep CNN. The dataset consists of 10 samples collected from a prehistoric cultivation field. Table 2 provides an overview of IP.

How does the practice of using ML models enhance the process of banana leaf disease diagnosis?

In 2020 Selvaraj et al. (2020), used UAV and multispectral satellite imagery for banana classification based on pixels. Using the MicaSense RedEdge multi-spectral camera and the Pix4Dcapture app, UAV-RGB (Red Green, Blue) photos were obtained for the localization of bananas and disease diagnosis. DL models such as RetinaNet and ResNet are used to locate and identify banana plants. The accuracy of the performance metric was 97%, while the recall results were 0.99, 0.95, 0.92, and 0.90. Three distinct models for diagnosing illnesses in banana leaves are shown by Vidhya & Priya (2022) utilizing both DL techniques, Alexnet, and ML techniques, K-Nearest Neighbors (KNN) and SVM. In particular, sigatoka and leafspot is manily focused on this study. A total of 12,544 images using basic photography equipment, photographs of banana leaves were taken from different banana fields in Kerala, India. These include augmentation methods, including clipping, zooming, and rotation. It yields an accuracy level of 96.73% with AlexNet, 84.86% with SVM, and 76.49% with KNN. Figure 9 explains the workflow of banana leaf diagnosis using ML.

The study by Chaudhari & Patil (2020) exhibited the design of a SVM classifier to identify banana leaf diseases by visualizing the symptoms, especially focusing on sigatoka, cucumber mosaic virus, bacterial wilt, and panama diseases. A digital camera has been utilized for capturing images. Preprocessing techniques for denoise removal was utilized. K-means clustering is used during the segmentation process in identifying the diseased regions while feature extraction, colors, shapes, texture features are extracted to classify the type of disease and those texture features are evaluated by the GLCM. It is shown the classification of the four classes of diseases is correct in 85% of cases. The author (Freire et al., 2023) targeted detection of sigatoka disease on the banana foliage and proposed a method integrating CNNs with Shapley values for developing an explainable diagnostic system for banana leaf sigatoka disease. The dataset contains 155 images of healthy plants, 320 images infected with sigatoka, and 814 images infected with xanthomonas. The SHapley Additive exPlanations (SHAP) SHAP VGG 16 model was found to be working better than other models. The study by Saranya et al. (2020) presented a system that utilizes image processing techniques to detect illnesses early, along with an artificial neural network (ANN) algorithm for classification. The process begins with pre-processing techniques, such as scaling and filtering, including both high-pass and low-pass filters. Ten fruit and leaf samples are used for disease detection. Segmentation methods, such as fuzzy c-means clustering and thresholding, are employed. The diseases targeted in this study include black Sigatoka, freckle-affected leaves, and anthracnose-affected fruits. Yathurshan et al. (2023) introduces “Leaf Guard” a Smartphone application designed especially for the agricultural environment in Sri Lanka, aimed at detecting diseases in the banana crop and treating the same. In this, field visits, agricultural journal records, and expert talks were used to collect 6,000 pictures of leaves of banana plant. Pictures were resized before processing the images using Histogram-based feature extraction methods. In this application, Xanthomonas and Sigatoka- related issues are dealt. It provides a cure percentage analysis along with a chemical suggestion system. More than 95% accuracy was achieved in total. The article reviews (Raja & Rajendran, 2022) banana leaf disease detection methods, highlighting that traditional ML models like SVM, KNN, and ANN achieve around 89–98% accuracy but rely on manual feature extraction and struggle with complex images. In contrast, DL models (CNN, DCNN, MobileNet, LeNet) automatically learn features and perform better, reaching up to 99% accuracy even under variations in lighting, scale, and background. The study emphasizes that DL is the most reliable approach for early, automated, and real-time detection, though challenges remain in terms of limited datasets, detection of tiny symptoms, and lack of severity estimation.

Thiagarajan et al. (2024) analyzed banana plant health using image capturing and preprocessing. It employed k-means clustering for segmentation and used SVM and CNN for classification, focusing on diseases like panama wilt and black sigatoka. A filtering method is proposed by Prabhakar & Sudha (2024) for the detection of illnesses in the leaves of banana plants. Numerous preprocessing approaches, multiple segmentation strategies, and numerous feature extraction strategies are all used. Three different filtering algorithms were applied to the images: the Roberts, Prewitt, and Canny filters. MATLAB and Python are two other image-processing tools used in the study. This article discusses three diseases: pestalotiopsis, sigatoka, and cordana. However Pai, Naik & BR (2020) introduced a method of detecting banana leaf diseases with a combination of a CNN algorithm with an ML model. There are four disease types with 1,200 images used here. The web interface allows users to upload an image, and the application will identify the type of disease present. In this study, a method for detecting sigatoka disease in banana leaves was proposed (Tuazon, Duran & Villaverde, 2021). A total of fifty banana leaves affected by the disease were collected, and each leaf was photographed twice and subsequently stitched together. To eliminate the background, Otsu’s thresholding technique was applied. A confusion matrix was then utilized to analyze the gathered data, resulting in an accuracy of 96.15%. Aruraj et al. (2019) introduces a method to identify and classify diseases on banana plants using texture patterns. The uses two steps approach such as texture feature extraction with Local Binary Patterns (LBP) and disease classification with SVM and KNN. The technique was tested on the public dataset, which included two cases: healthy and black sigatoka and healthy and cordana leaf spot. In total, 123 images were collected. The SVM classifier had an accuracy of 89.1% and 90.9%.

A method of disease recognition for banana leaves by using LBP along with GLCM was presented for extracting features (Chaudhari, Dawoodi & Patil, 2022). These methods contrast with color-based feature extraction techniques and use SVM and K-NN classifiers. They diagnose diseases such as banana yellow sigatoka, bunchy top, and cucumber mosaic virus. Digital cameras, as well as smartphones, were used for capturing images in a natural agricultural environment. Here Mary, Singh & Athisayamani (2021) automatically classifying banana leaf diseases through Gabor feature descriptors technique is proposed. It uses Gabor filters to extract texture features from images of diseased banana leaves, which will then be analyzed and classified by ML algorithms.

The Compared ShuffleNet V2 CNN model was compared with SVM and KNN for image classification of the black Sigatoka infection in banana leaves is proposed (Yumang et al., 2023). Gathering samples from 30 healthy leaves and 30 black sigatoka-infected leaves. This CNN model gave 95% accuracy in classifying raw and augmented images obtaining 96.67% sensitivity rate and 93.33% specificity rate. The model was deployed on a Raspberry Pi prototype for testing, and its performance was tested against SVM and KNN; it proved to be effective in anomaly detection in agriculture.

In Syihad et al. (2023) CNNs and ML are used to classify banana plant illnesses based on images of the leaves. The VGG-19 and ResNet50 models are utilized, with the former achieving 91% accuracy, 88% precision, 91% recall, and 89% F1-score. The dataset comprises 936 photos. Banana leaves are classified into four groups: pestalotiopsis, sigatoka, cordana, and healthy. This dataset was sourced from the Kaggle website.

The proposed system (Chaudhari & Patil, 2023) uses ML to detect diseases in banana leaves, which involves image acquisition, preprocessing (contrast stretching and filtering), and picture segmentation using a genetic algorithm. Features are extracted using LBP. The technique utilizes the strength of several classifiers by combining their outputs using an ensemble method and applying a genetic algorithm for segmentation in a new way. The images were taken with 48 MP cameras on mobile phones. The image is 640 × 480 pixels in size. The ensemble model outperformed single classifiers like SVM, Naïve Bayes, and others with an accuracy of over 92%.

In Nayak et al. (2024) the RF-XG-ResNet50 ensemble DL classifier is introduced to automate banana leaf disease classification with a hybrid architecture, where an ensemble multi-architecture is combined to enhance the accuracy. Data augmentation has been used for better training and generalization. It has been tested with an accuracy of 95% against DT and SVM. The Kaggle public dataset is used for training. Table 3 explains the overview of ML-based techniques.

In what ways does DL add to the complexity of identifying and classifying banana leaf diseases?

This method employs image segmentation to automatically detect diseases in banana plant leaves. It utilizes a collection of 9,000 images from the International Center for Tropical Agriculture (CIAT) banana image library to identify Black Sigatoka, Yellow Sigatoka, banana bacterial wilt, and healthy plants through the integration of hybrid fuzzy C-means clustering. For the preprocessing stage, median filtering and image resizing were applied. The Total Generalized Variation Fuzzy C-means clustering (TGVFCM) technique merges fuzzy C-means with Total Generalized Variation for effective segmentation. The TGVFCM, combined with a CNN, achieved an accuracy of 93.45%, specificity of 96.38%, and sensitivity of 89.04% is addressed (Gokula Krishnan et al., 2022). The study by Narayanan et al. (2022) used a hybrid CNN to identify illnesses in banana plant leaves. The dataset, collected from Tamil Nadu and the Tamil Nadu-Kerala border, comprised 3,500 images of healthy and damaged banana plants. Pre-processing techniques included median, low-pass, and high-pass filtering. SVM served as the classifier, while CNN was used for feature extraction.

The study (Ridhovan, Suharso & Rozikin, 2022) used 936 images of healthy and disease-infected banana leaves from the Kaggle website, collected by the remote sensing lab using a smartphone camera. Data preprocessing involved under-sampling, over-sampling, and data augmentation. DenseNet and Inception techniques were applied, achieving an F1-score of 84.62%, a recall of 84.73%, an accuracy of 84.73%, and a precision of 84.80%. Seetharaman & Mahendran (2022) developed an image-processing method for diagnosing diseases in banana leaves using 1,875 images from online and Indian farms. Pre-processed images were segmented, and features were extracted using Gabor-based patterns and a convolutional recurrent neural network (CRNN). The CRNN-RCNN model achieved a 98% accuracy, outperforming DCNN (88.9%), SVM (92.63%), KNN (79.56%), and CNN (87.6%).

Unveiled a novel agro-DL model (Sangeetha et al., 2023) aimed at recognizing the panama wilt disease. Utilized a set of 1,000 JPEG images. Pre-processing included rescaling and transforming the image. For segmentation, max pooling was implemented, for augmentation process cropping, flipping, rotating, and zooming the images was used. The proposed method attained a recall of 88.56%, an accuracy of 91.56%, a precision of 91.61%, and an F1-score of 81.56%. CNN architecture was presented (Bhuiyan et al., 2023) that has been improved using Bayesian optimization called BananaSqueezeNet. This architecture can identify three types of diseases in banana leaves. Moreover, the model can diagnose diseases affecting banana plants. A smartphone application was also developed for the identification of diseases. Data was collected from the experimental field as well as from various banana fields in Bangladesh. The 937 pictures have four categories: cordana, sigatoka, pestalotiopsis leaf blight disease, and healthy leaves. Augmentation techniques, including the Gaussian blur, horizontal flipping, image cropping, linear contrast modification, shearing, translation, and rotation, were used to address the problem of unbalanced data. CNN was used in the classification process. There were two different types of layers in the CNN: feature extraction layers and classification layers. The neural networks that use Bayesian optimization like SqueezeNet, MobileNets, EfficientNets, and ResNet, are used for image classification and object detection tasks. With a recall of 96.25%, precision of 96.53%, F1-score of 96.17%, specificity of 98.75%, and MCC of 95.13%, the model attains an overall accuracy of 96.25%.

The Ghost ResNeSt-Attention RReLU-Swish Net, based on ResNet50, Deng et al. (2024) uses a dataset of 13,021 banana leaf images to detect diseases. It achieved high detection rates: 96.42% for Sigatoka, 97.54% for cordana, and 95.58% for pestalotiopsis, with an overall accuracy of 98.38%. An application for smartphones for the identification of diseases in banana leaves, like fusarium wilt and black sigatoka, is outlined by Elinisa & Mduma (2024a). The 27,360 pictures were divided into three groups: black sigatoka, fusarium wilt, and healthy banana leaves. Mobile cameras were used to gather this image dataset from Tanzania. The programs VisiPics and Duplicate Photo Finder were utilized to identify and eliminate duplicate photos. In a matter of seconds, the suggested program could accurately detect duplicate images with 90% accuracy.

The study by Mathew, Kumar & Cherian (2023b) developed an automated approach by DL for classifying three major banana leaf spots: sigatoka, cordana, and deightonella. A Canon EOS 750D camera captures photography at the farms at Kerala Agricultural University (KAU). The third author, being a plant pathologist, supervised the data collection and labeling processes. The accuracy attained for DenseNet 121 was 91.7%.

This study (Banerjee et al., 2023) combined a mixture of CNN and SVM to help in disease diagnosis on banana fruit leaves. Using the photographs from both primary and secondary sources, the authors created a dataset of 1,100 images, which fall into four groups. The preprocessing steps come denoising and feature extraction using CNN, followed by disease classification of the leaves in SVM. This hybrid approach of CNN and SVM achieves performance accuracy of 94%. Authors in Tanwar, Sharma & Aanand (2023) used CNN and SVM techniques in network technology for the purpose of banana fruit leaf disease detection. It has utilized a big real-world data set of 1,100 images gathered from farms and online repositories. It targeted three types of banana disease: black sigatoka, cordana, and speckle disease. These types of disease classification were achieved by using SVM. The implemented methodology reached an accuracy across four different experimental scenarios of 90%, 95%, 89%, and 88%

An eight convolutional layered DCNN was proposed (Devi et al., 2022) to predict banana leaf disease. The public dataset Kaggle ML repository was utilized, consisting of 937 photos from the cordana, healthy, pestalotiopsis, and sigatoka disease categories. The training dataset was used by eight convolutional layered DCNNs. The same dataset was tested on models including MobileNetV2, DenseNet201, VGG19Net, InceptionV3Net, ResNet152, and NASNet. Large to determine performance metrics. 98.75% accuracy, 98.75% recall, 96.42% precision, and a 97.57% F1-score were attained.

In Singh, Guleria & Sharma (2024) was developed and deployed a CNN model that will identify and locate the banana leaves in images accurately. The images of bananas were collected from the online Kaggle repository. Preprocessing techniques like scaling, normalization, and augmentation were utillzed, and the model has convolutional layers to extract the features accurately and classify them. This method accurately determines diseases in banana leaves correctly for a success rate of 98.24% for 20 epochs.

The study by Bharathi Raja & Rajendran (2023) discusses a TL approach that incorporates a Faster R-CNN to support detection. The method uses all three different architectures: ResNet 50, InceptionV2, and MobileNet1. Haq et al. (2024) proposed a hybrid DL architecture named DenseMobilenetV2 that integrates the strong and reliable architecture of DenseNet121 with the efficient and accurate MobilenetV2. Data were gathered from Bangladesh, which consist of JPG images of 937 bananas, including samples of those infected by sigatoka, cordana, pestalotiopsis, and healthy samples. Color inversion, along with the data augmentation technique, has been used in pre-processing. The method adopted for using an outlier has been the IQR. And, the precision accuracy score generated was 100% on the model. F1-scores, along with recall accuracy, scored 0.9987 and 0.9974, respectively; those have been considered the optimum scores.

Faster RCNN, accompanied by Mohanraj et al. (2023) region proposal network in the detection of diseases for banana leaves. CNN has been used to extract feature extraction. The dataset for the research was sourced online through the website Kaggle.com. It contained both publicly available images and real-life photos. The diseases present are bacterial wilt, sigatoka, and pestalotiopsis. In addition, techniques involving image scaling and augmentation, like flipping, rotation, cropping, as well as scaling, are used. The accuracy reached was 0.9463.

The study by Sanga et al. (2020) developed a mobile application that uses five DL architecture models: Resnet18, Resnet152, Vgg16, Resnet50, and InceptionV3 are intended to detect illnesses in banana leaves. To implement the mobile version of this study, InceptionV3 was used in the mobile deployment. FUSI (Fusarium Sigatoka) Scanner is the name of the developed application. The models’ performances are as follows: InceptionV3 performed at 95.4%, Resnet18 attained 98.8%, Resnet50 likewise attained 98.8%, Resnet152 recorded 99.2%, and VGG16 attained 98.4% (Mora et al., 2024). A RedEdge MS camera, drone, and UAV imagery were used to photograph landscapes in Eastern Congo, facing challenges like low altitude and tall trees. Techniques for image augmentation addressed data imbalance, and DL algorithms like Yolo-V8 and Faster R-CNN detected banana stems infected with banana xanthonomas wilt. Sujithra & Ukrit (2022) published comparing the performance of banana and sugarcane leaf disease. Images were taken from the farm Acharapakkam of Chengalpattu District, Tamil Nadu. The camera used here is a Canon EOS 3000D, with a resolution of (512 * 512). This article used CNN, Feed Forward Neural Network (FFNN), and Radial Basis Neural Network (RBNN) for analysis. In addition, preprocessing techniques and classification models, which include CNN, were employed. The obtained performance metrics for banana and sugarcane were 97% accuracy, 89%, and 90% in sensitivity and specificity at 98%, 87%, 91%, 96%, 86%, and 92%, respectively. For the same data, accuracy was obtained at 95%, 88%, and 91% in sensitivity and specificity at 94%, 86%, 92% and 96%, 87%, and 94%, respectively. A lightweight CNN for fruit crop leaf disease detection was proposed (Hari & Singh, 2023) in, focusing on banana, guava, and mango leaves. With a dataset of 8,010 images, it achieved accuracies of 98.82%, 99.71%, and 97.41%. Also, the other study (Aliff et al., 2024) focused on the utilization of DL methods and aerial photography to detect banana plant diseases. Preprocessing, segmentation, feature extraction, classification, and picture acquisition are also included in this. 80.36% and 79.24% accuracy were achieved by the CNN. Figure 10 estimates the workflow of DL.

DL technique Heap Auto Encoders (HAEs) (Mary, Robert Singh & Athisayamani, 2020) proposed, K-means algorithm used for segmenting the banana leaf, the Expectation-Maximization (EM) algorithm is also used. The two HAE functions utilized the dropout method and the ReLU activation function, utilizing the publicly available dataset. The diseases focused on are bunchy top, bacterial bright spot, and banana bacterial wilt. The accuracy achieved is 99.35% for real data. The study (Criollo et al., 2020) proposed a CNN to detect the diseases concerning banana leaves. The data had 623 images, which were stored in three classes which were black sigatoka, bacterial wilt, and healthy, in an online repository. The architecture started with AlexNet, which includes convolutional, max pooling, and dense layers.

The dataset is around 8,000 to 9,000 images, gathered from the repositories Mendeley and Kaggle (Vinta et al., 2025). There were three classes of images included Kaggle dataset sigatoka, xanthomonas, and green leaf. Panama diseases, including black and yellow Panama, are also mentioned in the Mendeley.

A multitude of modifications—crop rotation, reflection, scale, crop, color saturation, and noise—were performed. In this work, the You only look once (YOLO) algorithm is used to diagnose illnesses. With 900 gathered image samples, Jain et al. (2024) classified banana plant diseases into five severity levels using a technique based on ML. In this work, the CNN and DT were combined for training and evaluation purposes. Levels 0 through 5 also represented the non-symptomatic to highly affected stages of the disease’s severity. The overall performance reached an accuracy of 96.04%.

Here, a collection of 18,000 field images was gathered (Selvaraj et al., 2019) through a network of banana experts from Tamil Nadu Agricultural University (TNAU) in Southern India and Africa. The proposed system experimented with architectures such as ResNet50, InceptionV2, and MobileNetV1. These models were trained to identify various types of diseases and pests affecting banana plants. The research achieved an impressive classification accuracy ranging from 70% to 99%.

A DL method (Neupane, Horanont & Hung, 2019) identified and counted banana plants from UAV RGB images, achieving 99% recall despite challenges like lighting and overlapping leaves. Techniques used included the Triangular Greenness Index. Saleem, Potgieter & Arif (2019) analyzes the use of DL models, specifically CNN, for disease diagnosis. The model utilizes a dataset from Plant Village for training. The workflow comprises data collection, data cleansing, and CNN architecture model training, followed by the evaluation phase that gauges the model’s performance. The findings show that DL approaches perform better than conventional approaches.

The study (Baguisi, Buenaventura & Yumang, 2024) works with the CNN algorithm to detect black sigatoka infections in the banana leaf using IP. This analysis was done using the raw dataset for augmentation. The performance attained accuracy of 95%, specificity of 93.33%, and a sensitivity of 96.67% (Elinisa & Mduma, 2024b). Implemented the Mask R-CNN DL technique to handle disease detection and segmentation on banana plants. This method utilizes the model’s capability of instance segmentation, whereby the pixel level of the area that is afflicted with disease is accurately pinpointed. This is a way to mitigate the accurate and efficient automated methods of disease detection, which tend to be labor-intensive. The proposed technique attained a mean precision of 0.04529 while utilizing an image dataset of 6,000 stalks and banana leaves obtained from the field, and was able to achieve segmentation on the afflicted regions proficiently.

On the other hand, Altabaji et al. (2023) proposed a disease diagnosis method for banana leaf diseases—cordana, pestalotiopsis, and sigatoka using CNN through a set of images of 2,000 banana leaves. The best CNN architectures for feature excitation and disease identification are chosen based on the performance of three models namely MobileNet, Xception, and InceptionV3. The proposed method produced results that should be further explored, as it showed a validation accuracy rate of 96.24%, making it very plausible for exact and effective disease classification and detection of banana leaves. Sigatoka and xanthomonas are the banana leaf diseases that are classified using DL models, namely the ResNet-18 and ResNet-34 architectures (Anushya & Shiwani, 2024). Images of healthy and sick banana leaves with xanthomonas and sigatoka infections. The images were gathered from various fields of agriculture. ResNet-34 achieved 98.32% accuracy. This work (Nieto et al., 2024) compares two methods: combination of the Green Leaf Index, K-means clustering, and YOLOv5m DL model. The YOLOv5m model performed precision of 86% and recall of 90%. In the task of coverage estimation, U-Net architecture was used, which had an IoU of 52.3%.

The study (Jadhaw & Bhandari, 2024) employed DL models to detect banana leaf diseases early, using 1,000 images from Madhya Pradesh and Maharashtra. It focused on diseases like banana bacterial wilt and sigatoka disease. The YOLOv4 algorithm (Ibarra, Rivera & Manlises, 2023b) has been applied for detecting panama disease in banana leaves. Besides, the annotated dataset of 500 images of banana was collected from the Real-world plantations. It identifies the disease and locates with high accuracy and precision using the YOLOv4 DL model. The focused disease is panama disease. The primary algorithm (Rajalakshmi et al., 2025) using CNN, trained to recognize images of banana crops with or without disease was applied. Rectified Linear Unit (ReLU) activation functions are applied in all CNN layers. Cross-entropy loss and optimization algorithms by applied during the training process. The best performance of the model delivers 88.32% accuracy. The GLCM for feature extraction from texture and classification by using DCNN (Mahendran & Seetharaman, 2022) was applied.

The ResNext50 DL model to create an automated framework for the classification of the severity of sigatoka disease in banana leaves with 10,000 annotated images in the dataset Singh et al. (2024) was used. This has been done at 95.53 accuracy with rigorous preprocessing and fine-tuning. It has been validated using recall and F1-score metrics. The confusion matrix and graphical comparisons explain the model’s effective performance, forming a foundation for precision agriculture and sustainable crop management through automated disease evaluation.

A hybrid model Kumar & Kumar (2024) that predicts the severity of banana leaf wilt using CNN and SVMs: Using DL techniques, which are effective in providing an accurate diagnosis of the condition, the model is created using 10,000 image-labeled sets of banana leaves; data augmentation improves the feature with CNN that has already been trained. It was able to provide a thorough assessment of the disease’s severity with an accuracy of 94.77% using precision, recall, and F1-score metrics. When compared to other top models and current techniques, it fared better. The lightweight DL algorithms to identify banana leaf diseases from images collected in Arusha, Northern Tanzania, with a smartphone, producing 16,092 images (Nassor, Mushthofa & Priandana, 2024) was used. Out of the models experimented: MobileNetv2, MobileNetv3-small, ShuffleNetv2, and SqueezeNet, SqueezeNet gave the better performance by recording 97.12% accuracy, precision, recall, and F1-score. MobileNetv3-small has a faster training time; MobileNetv2 was the lightest of the algorithms.

Senthilraj & Parameswari (2022) using CNN and VGG16 models, create a DL model to recognize and categorize the stages of black sigatoka disease in banana leaves. The CNN model showed better results than VGG16. Table 4 give the DL method used. CNN and M-SVM are used by Yan, Shisher & Sun (2023) to diagnose banana diseases in one category, specifically xanthonomas, fusarium wilt, bunchy top virus, black sigatoka. A total of 3,500 leaf photos taken with cell phones equipped with VGA cameras from various locations in South India, particularly Tamil Nadu and the Kerala border, made up the dataset. During preprocessing, median filtering was used, and the system achieved 99% accuracy, outperforming comparable techniques in disease prognosis.

What kind of impact does TL have on the process of detecting and classifying banana leaf diseases?

Deep CNN combined with TL to detect the fusarium wilt of banana plants (Derafi, Razak & Sayeed, 2024) was proposed. In this regard, ResNet-50 was used, and a TL is applied in it to enhance its precision as well as efficiency. A total of 500 test images consisting of healthy and diseased parts were used. The accuracy is recorded at 0.98 and the F1-score is at 0.98.

Salehin, Islam & Alam (2024) applied TL for the disease identification of the banana leaf. Besides that, pre-trained models, EfficientNet, AlexNet, VGG, and ResNet, were also tested with the diseases. The dataset is obtained from the Plant Village, Nelson Mandela dataset, with 17,068 images of banana plants, and the PSDF- Musa. A total of 54,303 images of 14 types of crops were used. Among them, the data preprocessing techniques involve methods such as torchvision. transforms.resize for resizing and e-transforms. ToTensor() for normalization of pixel values, with transforms. Normalize normalizing data. The accuracy and the F1-score are also more than 90% and 0.90%.

TL-based (Genet et al., 2024) is proposed to automatically identify banana diseases. A dataset of 17,068 labeled images was compiled using the AdSurv mobile application on Samsung phones for taking pictures of banana leaves and stems. Researchers from the African Institution collaborated with the Tanzania Agricultural Research Institute. This comparison study on the performances of such models as VGG19, NasNetLarge, Inception ResNet V2, and Xception found that the model Xception is the best since it achieved a classification accuracy of 98.062%. A robust disease detection model is developed (Priya et al., 2024) using CNN algorithms to classify and identify diseases in banana plants. The dataset comprises 1,982 photos, all of which have been affected by banana diseases. The images are categorized into three groups: healthy, xanthomonas wilt infected, and sigatoka leaf spot infected. Using data augmentation and TL techniques that are high-end, the architecture of models was fine-tuned using MobileNet, EfficientNet, VGG16, VGG19, and InceptionV3. Amongst them, the VGG16 model performed excellently well and successfully withstood independent test datasets with 81.53% accuracy. A DenseNet201Plus model was created (Mazumder et al., 2024) to identify banana leaf diseases, achieving 0.9012 accuracy and precision, 0.9012 recall, and 0.9716 AUC. It was trained on healthy, sigatoka, and xanthomonas leaf classes, using data from Mendeley Data and enhanced through rotation, flipping, and scaling.

A fused nine pre-trained CNNs (Shafik et al., 2024), including ResNet, DenseNet, and EfficientNet, with deep feature extraction to build Plant Disease Detection Network—AutoEncoder (Early Fusion (PDDNet-AE)) and Plant Disease Detection Network—Lead Voting Ensemble (PDDNet-LVE) models. The experiments on the PlantVillage dataset, containing 54,305 images organized in 15 categories, validated their claims of superior performance. PDDNet-AE reached an accuracy of 96.74%, and PDDNet-LVE outperformed it at 97.79%, surpassing other CNN models.

The author (Huang et al., 2024) uses DL to enhance the detection of banana leaf diseases to improve agricultural productivity and banana quality. It employs a multi-layer self-attention mechanism on top of the VGG16 model, which acts as the base architecture. The dataset image was captured using a smartphone camera at Bangabandhu Sheikh Mujibur Rahman Agricultural University and surrounding banana farms. Which contains images of both diseased and healthy banana leaves, along with three prominent leaf spot diseases: cortana, pestalotiopsis, and sigatoka, all annotated by plant pathologists with 400 images added to each class for comprehensive analysis. By adding the self-attention mechanism layer, the model accuracy was increased from 89% to 92%. A CNN was used (Arifin, 2025) to classify banana leaf diseases, healthy, fusarium wilt, sigatoka from 1,289 images. Data augmentation techniques improved generalization. GoogleNet achieved accuracies of 90.59% for Healthy Leaves, 90.30% for Fusarium Wilt, and 90.49% for Sigatoka.

Nasra & Gupta (2024) uses a CNN model based on ResNet50 to identify banana leaf diseases, achieving 96% accuracy across 937 images. The model effectively classifies cordana, healthy, pestalotiopsis, and sigatoka, with strong performance metrics for each category. Tiwari (2024), suggested the ViT techniques, such as ViT-b16 and ViT-l16, for classifying banana leaf diseases. These models were trained and tested with 408 banana leaf images which were categorized into seven classes of diseases which included: healthy leaf, black sigatoka, bract mosaic virus disease, insect pest diseases, panama disease, moko disease, and yellow sigatoka. Also, the accuracy obtained 96.88% arose from ViT-b16 outperforming all other deep architectures including GCN, EfficientNet-B3, MobileNet-V2, and ResNet-cut.

In Saini (2024) image classification settles on the Vision Transformer (ViT) model to classify various diseases afflicting the banana leaf and achieves an astounding accuracy of 96.88% at epoch 100 is noted. The dataset was collected through Kaggle and comprised of 408 images from farmers’ fields, which was later expanded to 2,856 images using flip, zoom and brightness increase techniques. The model accurately identified the damages done due to mosaic virus, black sigatoka, moko disease, panama disease and yellow sigatoka with very few instances of misclassification.

In this study, Sharma & Vishwakarma (2024) suggests a ViT-based approach to classify the leaves of banana plants diseases. Unlike traditional plant monitoring methods, which are laborious and human-dependent, computer vision techniques provide an efficient solution. Vision transformers tend to perform better on real-world crop field images by looking at global patterns rather than only local regions. The proposed model achieved 90% accuracy on a dataset from Maland University, Karnataka, India, which comprises over 7,000 images of banana plant leaves affected by deficiencies in eight unique nutrients. The Convolutional Swin Transformer (CST) model, proposed in Guo, Lan & Chen (2022) 2022 for detecting plant diseases, achieved high accuracy (up to 0.982) even in noisy conditions. It used a dataset of 1,288 banana leaf images and employed data augmentation to prevent overfitting, outperforming other models in real-life agricultural scenarios. Table 5 illustrates the TL-based approaches.

Discuss the model limitations and failure cases?

To identify banana black sigatoka disease, the study (Kayanja et al., 2024) developed a CNN-based approach by applying techniques such as saliency maps, SHAP, and Gradient-weighted Class Activation Mapping (Grad-CAM). The decision-making model can be employed to clarify the results observed. This methodology leverages the ResNet-18 model along with a unique CNN-based model, which surpasses existing techniques and promotes responsible AI in disease detection.

Bharath, Hemalatha & Sreekumar (2025) uses Global Average Pooling (GAP)-based CNN model on the Banana LSD dataset for the identification of diseases in banana leaves. The model achieved a test accuracy of 74.71%. XAI strategies such as Grad-CAM and SoftMax confidence analysis have revealed important aspects of model explainability. As observed in the Grad-CAM heatmaps, CNNs tend to fixate on biologically meaningful spots and dark patches associated with the disease, which reveals a degree of misclassification. Moreover, SoftMax outputs showed a lack of confidence between classes such as Healthy and Pestalotiopsis, which are visually analogous to each other.

Ünal (2024) introduces the Banana Leaf Spot Diseases (BananaLSD) Dataset as a response to the growing concern of banana leaf spot diseases and its consequences on food security, a systematically assembled dataset that consists of 1,600 augmented and annotated images of the most common three banana leaf spot diseases—sigatoka, cordana, and pestalotiopsis—along with healthy banana leaves and focuses on DL. The classification accuracy of DL models DenseNet-201, EfficientNet-b0, and VGG16 was evaluated through accuracy, precision, recall, F1-score, and confusion matrix calculations. The performance attained by DenseNet-201 is 98.12% accuracy, followed by 97.81% VGG16, and then EfficientNet-b0 with 87.81% accuracy. The use of Grad-CAM visualizations is used to validate model predictions and highlights the diseased regions on the images.

Model interpretability and explainability are essential for creating trustworthy and actionable AI solutions in agriculture, although minimal research has focused on these aspects concerning banana leaf disease detection. The transparency of model decision-making has been somewhat neglected in academic discussions, especially when compared to the focus on improving DL architectures and increasing classification accuracy. To illustrate feature significance and identify disease-impacted regions on banana leaves, several key advancements have utilized XAI methods such as Grad-CAM, saliency maps, SHAP, and Local Interpretable Model-agnostic Explanations (LIME). Although they are currently uncommon, these initiatives are essential for enhancing the transparency, reliability, and effectiveness of AI-based diagnostics for banana plant leaf diseases. These studies provide a valuable understanding for building user trust and aiding informed decision-making in precision agriculture by emphasizing interpretability alongside performance metrics. The study by (Jiménez et al., 2025) analyzed EfficientNetB0, ResNet50, and VGG19 on the classification of diseases in banana leaves; here, the DL models had recurrent miscalculation of the images as black sigatoka, cordana, and healthy leaves. ResNet50, for example, miscalculated 16 of 60 images processed as black sigatoka leaves. Despite the models achieving above 88% accuracy, there was severe miscalculation, especially between Cordana and Black Sigatoka. MobileNet and AlexNet were other models attempted, but because of overfitting, they only achieved 33% accuracy and were discarded. The researcher (Helmawati & Utami, 2025) trained MobileNetV2 on classifying four classes of banana leaf diseases, achieving 90.62% accuracy. The dataset imbalance leads to low learning in (45.26% accuracy at epoch 1), performance on images taken in different lighting. Due to the absence of the confusion matrix, class-wise evaluation, and class accuracy evaluation, results, the dataset was deemed unreliable for application in multiclass situations. The dataset used in the research (Escudero, Calvo & Bejarano, 2022) contained 937 images, which were imbalanced because they contained mostly images of sigatoka, leading to poor accuracy in less represented classes such as cordana and healthy. The model accuracy attained 90.62%, but no other metrics, such as precision, recall, or F1-score, were used, which makes the performance interpretation limited. The model is built on MobileNetV2 alone, which is a sharp contrast to other DL models like ResNet or even traditional models.

The accuracy of the study (Sholehah & Siti, 2021) is constrained by a very small dataset consisting of just 48 images, which heightens the model’s risk of overfitting and restricts its ability to generalize. The model’s accuracy is also likely overestimated due to the absence of external validation and dependence solely on cross-validation, which ultimately limits the model’s real-world applicability.

The accuracy of classification achieved by the proposed (Deng et al., 2024) method is quite high. The model incorrectly classifies many banana leaf diseases, including Sigatoka, Cordana, and Pestalotiopsis. This is primarily because the yellow halos and certain shaped lesions associated with these diseases are quite common. The model suffers from a lack of real-world robustness to occlusion, motion blur, and low lighting conditions, which is not well-covered in the training data. Reliance on augmented data from within the training set creates a scenario in which the model performs well in a limited set of conditions, but generalizes poorly to episodic field and outdoor unpredictability. Moreover, the model does not automate the delineation of the regions of interest within the leaf, further hampering relevant application to the integrated diseased tissue management. Genet et al. (2024) suffers from two primary issues. The key problems stem from the restricted geographical and spatial range of the dataset, which consists of 1,982 images from specific areas in Ethiopia. This limitation increases overfitting in the model due to insufficient diversity in the training set and affects the model’s ability to generalize. Additionally, the model does not incorporate XAI methods, leading to a lack of clarity in its predictions, which diminishes the reliability of the findings for practical agricultural applications.

Helmawati & Utami (2025) was limited in size and imbalanced, consisting of merely 937 images. This severely restricted the model’s ability to generalize and heightened the chance of bias favoring the dominant classes. Additionally, the model’s practical credibility is empirically compromised by the lack of real-world or external validation. Testing solely with one Kaggle dataset does not suffice; the model’s dependability cannot be asserted across a diverse range of field conditions.

In the study, Chong et al. (2025), two notable failure cases were reported. In the first case, SqueezeNet showed signs of overfitting with high training accuracy but poor validation consistency. This is likely the result of underwhelming available training data combined with overfitting and a lack of model flexibility. In the second case, all three models exhibited low F1-scores in classifying Fusarium wilt. This, most likely, stems from the disease’s high rates of incorrect identification due to a lack of prominent visual features. Alanazi (2025) suffers from class imbalance, where the overrepresented class, potassium deficiency, dominates the dataset, which deteriorates precision and recall for the underrepresented classes, like Panama disease and healthy leaves. Moreover, some diseases with high visual resemblance, for example, black sigatoka and potassium deficiency, tend to be misclassified heavily. Other notable restrictions include the absence of dataset diversity grounded in real-world environments, no comparative analysis with other object detection systems, and the lack of a clearly defined practical field-use deployment plan. Derafi, Razak & Sayeed (2024) described two key limitations. First, all the pre-trained models underwent training for only ten epochs. Moreover, all the models had a constant learning rate of 0.1, which is highly likely to hinder convergence, particularly for more complex models. Additionally, limited tuning of weak hyperparameters stifled the possibility of further optimization. Lastly, models achieving high accuracy in offline tests is not ideal in the absence of evaluation or deployment to real-time platforms such as mobile or edge devices. This practical absence, means the model is not useful in agricultural situations that require immediate action. Numerous limitations have been identified in the reviewed studies.

Confusion matrix

To evaluate a model’s performance in classification tasks, a confusion matrix is used to estimate how closely the simulation predictions align original results. The confusion matrix comprises four distinct elements. Figure 12 illustrates the confusion matrix model.

True positive (TP): Model correctly predicted the positive class.

True negative (TN): Model correctly predicted the negative class.

False positive (FP): Model incorrectly predicted positive when it was actually negative (Type I error).

False negative (FN): Model incorrectly predicted negative when it was actually positive (Type II error).

Accuracy

Accuracy is another metric of performance that is mostly employed in ML, more so in classification problems. This calculates the ratio of the model’s total correct predictions (Narayanan et al., 2022). Equation (1) is the formula used to calculate accuracy.

(1) $$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}+\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s}}.}$$

Recall (Or) Sensitivity

Recall, or Sensitivity, is a special metric used to determine how effectively a model finds all relevant positive cases in the data. The main objective of applying such a measure is to reduce the false negatives to maximize the detection of positive cases (Narayanan et al., 2022). The given Eq. (2) is used to calculate the performance of recall, also known as sensitivity.

(2) $$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{v}\mathrm{e}+\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}.}$$

F1-score

It is a performance measure that is more useful than the accuracy score in cases when classes are imbalanced. It is computed in such a way that it ensures that both precision and recall are captured with the single measure that it returns. This score normally applies where the need to recall all relevant items outweighs the risk of including too many irrelevant items (Ridhovan, Suharso & Rozikin, 2022). The F1-measure is useful in situations where a model’s precision, which refers to the number of selected items that are relevant to the tasks, or recall, which signifies the percentage of relevant items that are selected, has to be optimized. The following Eq. (3) is used for calculating the F1-score performance

(3) $$\mathrm{F}1\text{-}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \frac{2\times \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}.}$$

Selectivity

Selectivity is a measure used primarily for binary classification problems, assessing the probability of a true negative instance being accurately classified as such. This relates to sensitivity, which is often misestimated as the fraction of TP out of the total number of positives. Selectivity, however, is narrower in scope and is broader than specificity as it concentrates on the expected outcome of no positive diagnosis, resulting in avoidance of false positive predictions, making it of value in the context where the problem of false positive diagnosis is significant. Equation (4) is used for calculating the selectivity

(4) $$\mathrm{S}\mathrm{e}\mathrm{l}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}+\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}.}$$

Precision

Precision is especially important in classification tasks where the FP cost is high, as it is considered a performance metric (Narayanan et al., 2022). It represents TP’s ratio to the total instances that the model predicted as positive. Precision measures the accuracy of a model’s positive predictions. Equation (5) provides the formula for calculating precision performance.

(5) $$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}+\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}.}$$

RMSE

This measures regression models’ accuracy by averaging the absolute errors over predicted and actual values. To find the value, calculate the square root of the mean of the squared differences between the two values. Larger errors and outliers are more prevalent, so RMSE is very sensitive to them. Better model accuracy is implied through lower RMSE scores. When predicting the severity of diseases in bananas, the RMSE measures the model accuracy for continuous output predictions (Manavalan, 2020). The equation provided below (Eq. (6)) is the formula for calculating the performance metric known as RMSE.

(6) $$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{\mathrm{n}}\sum _{\mathrm{i}=1}^{\mathrm{n}}(\mathrm{y}\mathrm{i}-{\hat{\mathrm{y}}}_1{)}^2}.$$

• yi = actual value (true value)

• ${\hat{\mathrm{y}}}_1$ = predicted value

• n = number of data points.

SSIM

The SSIM is a complex performance metric that evaluates image quality through image structure comparison. It considers perceived quality based on luminance, contrast, and structure instead of pixel-wise differences, making it very useful for image-related tasks. SSIM is from −1 to 1, where −1 indicates the lowest similarity, while values closer to 1 signify higher similarity.

(7) $$\mathrm{S}\mathrm{S}\mathrm{I}\mathrm{M}(\mathrm{a},\mathrm{b})={\displaystyle \frac{\left(2\mathrm{\mu }\mathrm{a}{\mathrm{\mu }}_{\mathrm{b}}+{\mathrm{C}}_1\right)\left(2{\mathrm{\sigma }}_{\mathrm{a}\mathrm{b}}+{\mathrm{C}}_2\right)}{\left({\mathrm{\mu }}^2\mathrm{a}+{\mathrm{\mu }}^2\mathrm{b}{\mathrm{C}}_1\right)\left({\mathrm{\sigma }}^2\mathrm{a}+{\mathrm{\sigma }}_{\mathrm{b}}^2{\mathrm{C}}_2\right)}}$$

• a and b = Two images being compared

• $\mu a{\mu }_b$= Mean intensities of images a, b

• ${\sigma }_a^2+{\sigma }^{2}b$ = Variances of images a and b

• ${\sigma }_a,{\sigma }_b$ = Covariance between a and b

• C1, Small constants to stabilize the formula and avoid division by zero.

PSNR

The PSNR compares reconstructed or compressed images to the original ones. It measures the maximum possible signal power to the power of noise, with higher PSNR values indicating images of finer quality (Manavalan, 2020). The Eq. (8) provided is the formula used to calculate PSNR.

(8) $$\mathrm{P}\mathrm{S}\mathrm{N}\mathrm{R}=10\cdot {\log }_{10}\left({\displaystyle \frac{{\mathrm{M}\mathrm{A}\mathrm{X}}^2}{\mathrm{M}\mathrm{S}\mathrm{E}}}\right)$$

• MAX: Maximum feasible pixel value of the image

• MSE: Mean squared error.

MSE

The MSE is an error-based metric that is very popular in regression and ML-based tasks. The MSE shows the proximity of the regression line to a set of sample data points. A good mean squared error is 0, which means predictions are closer to the true values. Equation (9) is used to demonstrate the performance of MSE

(9) $$\mathrm{M}\mathrm{S}\mathrm{E}={\displaystyle \frac{1}{n}\sum _{i=1}^n{\left(y_i-{\hat{y}}_1\right)}^2.}$$

• n: aggregate number of data points

• yi: Actual value (ground truth) of the iii-th data point

• ${\hat{y}}_1$ = Predicted value for the iii-th data point

${\left(y_i-{\hat{y}}_1\right)}^2$ = Squared difference between the actual and predicted values.

AUC-ROC curve

The performance of a classification model over various threshold values is shown graphically by the AUC-ROC Curve. It is frequently used to assess models in problems involving binary classification (Manavalan, 2020). The equation provided in Eqs. (10), (11) is used to calculate performance.

(10) $$AUC=\underset{0}{\overset{1}{\int }}TPRd\left(FPR\right)$$

(11) $$\mathrm{R}\mathrm{O}\mathrm{C}=\mathrm{p}\mathrm{l}\mathrm{o}\mathrm{t}\mathrm{o}\mathrm{f}(\mathrm{F}\mathrm{P}\mathrm{R},\mathrm{T}\mathrm{P}\mathrm{R}).$$

• False positive rate

• True positive rate.