The guava plant has achieved viable significance in subtropics and tropics owing to its flexibility to climatic environments, soil conditions and higher human consumption. It is cultivated in vast areas of Asian and Non-Asian countries, including Pakistan. The guava plant is vulnerable to diseases, specifically the leaves and fruit, which result in massive crop and profitability losses. The existing plant leaf disease detection techniques can detect only one disease from a leaf. However, a single leaf may contain symptoms of multiple diseases. This study has proposed a hybrid deep learning-based framework for the real-time detection of multiple diseases from a single guava leaf in several steps. Firstly, Guava Infected Patches Modified MobileNetV2 and U-Net (GIP-MU-NET) has been proposed to segment the infected guava patches. The proposed model consists of modified MobileNetv2 as an encoder, and the U-Net model’s up-sampling layers are used as a decoder part. Secondly, the Guava Leaf Segmentation Model (GLSM) is proposed to segment the healthy and infected leaves. In the final step, the Guava Multiple Leaf Diseases Detection (GMLDD) model based on the YOLOv5 model detects various diseases from a guava leaf. Two self-collected datasets (the Guava Patches Dataset and the Guava Leaf Diseases Dataset) are used for training and validation. The proposed method detected the various defects, including five distinct classes, i.e., anthracnose, insect attack, nutrition deficiency, wilt, and healthy. On average, the GIP-MU-Net model achieved 92.41% accuracy, the GLSM gained 83.40% accuracy, whereas the proposed GMLDD technique achieved 73.3% precision, 73.1% recall, 71.0% mAP@0.5 and 50.3 mAP@0.5:0.95 scores for all the aforesaid classes.
Guava (Psidium guajava L.) has achieved commercial significance in subtropics and tropics due to its broad flexibility to wide climatic and soil conditions and high medicinal and nutritional values. Guava fruits and leaves both have remarkable minerals and nutrition [
Various methods have been developed to diagnose plant diseases [
With the innovations of Computer Vision (CV), Machine Learning (ML) and Artificial Intelligence (AI), advancements have been accomplished in creating computerized models, enabling precise and timely detection of plant leaf diseases. Within a decade, ML and AI innovations have achieved a massive intrigue in the accessibility of several high-speed computing devices and processors that have improved the processing time, reliability and accuracy of the results/output obtained. For automation of plant disease detection, various strategies, including conventional machine learning methods like SVM (Support Vector Machine), KNN (K-nearest neighbour), random forest etc., are used. A significant variation prevails in the accuracies of the methods employed [
In the last few years, Deep Learning (DL) has been primarily used in agriculture [
Dissimilarities in guava diseases prevail globally due to various factors such as the shape of the disease, symptoms, colour of leaves, varieties of guava and environmental factors. No study has proposed detecting and localizing multiple diseases on a single guava leaf image in the literature reported. Real-time guava disease detection is one step ahead and more challenging due to the possibility of multiple diseases on a single leaf. Although a guava leaf and fruit disease dataset has already been developed [ A Guava Infected Patches Modified MobileNetV2 and U-Net (GIP-MU-NET) segmentation technique is developed to segment the infected guava patches. The proposed GIP-MU-NET model consists of modified MobileNetv2 as the encoder and up-sampling layers of U-Net as the decoder. A novel Guava Leaf Segmentation Model (GLSM) is developed to segment the guava leaves from the healthy and infected patches. A real-time Guava Multiple Leaf Disease Detection (GMLDD) model is developed based on YOLOv5 to detect and localize multiple diseases from a single guava leaf. To the best of our knowledge, this is the first-ever model employed to observe various diseases on a sole guava leaf image. The proposed method can precisely detect and localize multiple diseases, including anthracnose, nutrient deficiency, wilt diseases and insect attacks, from a single image. A first-ever Guava Leaf Diseases Dataset (GLDD) is developed to detect the various diseases on a single leaf. The dataset consists of five classes: anthracnose, nutrient deficiency, wilt diseases, insect assault, and healthy.
The rest of the article is organized as the literature review in Section 2, the materials & methods in Section 3, the research results in Section 4, and Section 5 presents the conclusion and future work followed by the references.
Some work has been presented in the literature on guava leaf disease detection, but the methodology exercised focused only on detecting a single disease from a single leaf or fruit. Al Haque et al. [
Leaf segmentation with a complex background was also a challenging and complicated task. Few researchers worked on different plant leaf segmentation, such as [
You Only Learn One (YOLO) is a unique technique in deep learning to detect multiple objects and localize the object. Many types of research have been accomplished on disease detection and other classification using this technique, attaining good results. Kasper-Eulaers et al. [
Different varieties of guava exist in other regions of the world. Therefore, variations in guava diseases prevail worldwide for several reasons, including the shape of the disease, symptoms, colour of leaves, varieties of guava and environmental factors. In literature, all research is done on a single disease on a single leaf, but the earlier works failed to detect the multiple diseases on a single leaf. Most of the crops or leaves were attacked with various diseases on a single leaf. As reported in the literature, real-time disease detection and localization is another problem [
This research proposes a hybrid deep learning model for real-time guava multiple disease detection and localization from a single guava leaf. The proposed method consists of three parts. The first part detects the infected guava patches from the guava tree images using GIP-MU-NET. Secondly, the Guava Leaf Segmentation Model (GLSM) is proposed to segment the guava leaves from images. The final part detects the multiple diseases of a guava leaf using the GMLDD model. The flow chart of the proposed method is shown in
The performance of deep learning models heavily depends on an appropriate and valid dataset. We developed two real-time datasets of guava, i.e., the Guava Patches Dataset and Guava Leaf Diseases Dataset from the Okara district in the Central Punjab region of Pakistan. The detailed description of datasets is as follows:
A real-time dataset was developed in the form of videos and pictures. Different capturing devices, including mobile phones, digital cameras, and drones, were used to ensure variations in the real-time dataset. The capturing distance for the mobile phone and digital cameras was 1 foot to 2 feet, whereas the drone’s capturing distance was 5 to 10 feet. Since drone fanning distorted the videos and images because of plant leaf movement, we maximized the plant and drone distance as much as possible. District Okara in the central Punjab region of Pakistan was selected for the present research work due to the increased guava cultivation. We focused on the varieties of guava from the district Okara, namely Choti Surahi, Bari Surahi, Gola, Golden and Sadabahar. The images and videos were captured under varying conditions, including morning, evening, noon, cloudy, sunny, rainy, and different seasons (summer, winter, spring and autumn), to monitor the environmental variations and their effect on disease incidence and intensity. The images were taken at different resolutions to incorporate the variations in the dataset quality. With the help of Python code, guava plant videos were converted into frames (images). Then, with the assistance of plant disease experts, patches of plant images were annotated into infected ones (
Split | Images |
---|---|
Training | 956 |
Validation | 240 |
Test | 120 |
Using the Guava Patches Dataset, the guava leaf images were extracted using the proposed GLSM model. The leaves were annotated with the help of plant pathologists into five classes: nutrient deficiency, insect attack, anthracnose, wilt, and healthy, as depicted in
Split | Images | Class | Label samples | Total samples |
---|---|---|---|---|
Training | 2346 | Anthracnose | 2180 | 5777 |
Healthy | 447 | |||
Insect attack | 1398 | |||
Nutrient deficiency | 1281 | |||
Wilt | 471 | |||
Validation | 98 | Anthracnose | 53 | 232 |
Healthy | 24 | |||
Insect attack | 67 | |||
Nutrient deficiency | 69 | |||
Wilt | 19 | |||
Test | 98 | Anthracnose | 53 | 232 |
Healthy | 24 | |||
Insect attack | 67 | |||
Nutrient deficiency | 69 | |||
Wilt | 19 | |||
All datasets were split into the training, validation, and test sets used for the proposed model’s training, validation, and testing. An unseen dataset was used afterwards to ensure the model’s fitness. The GPD training set included 1196 images with the same number of masks (targets/labels) for the infected classes, as shown in
The Guava Leaf Diseases Dataset was divided into training, validation and test sets having 2346, 98, 98 images, and 5777, 232, and 232 targets, respectively, as shown in
The first step is to segment the infected guava patches. For this purpose, we developed a novel deep learning model called GIP-MU-Net. The proposed method implements an encoder-decoder U-Net [
This study aims to make the Guava Leaf Segmentation Model (GLSM) from a U-Net-like model to segment guava leaves from the images. The suggested GLSM model’s architecture is depicted in
The YOLO model [
As previously stated, the YOLO method used fewer calculations; thus, it was faster and more accurate. Moreover, there were still certain limitations while detecting a small object like RGs. The varifocal loss was employed rather than a focal loss in the original YOLOv5 to improve the model. The proposed model obtained a significant capacity to locate and categorize galaxies upon training with the help of a dataset.
In the original YOLOv5 neural network model’s structure [
Model name | GMLDD model |
---|---|
Number of classes | 5 (Anth, Healthy, Insect Attack, Nut. Def, and Wilt) |
Image size | 416 × 416 |
Batch size | 32 |
Epochs | 500 |
Training optimizer | Adam |
The proposed model was quantitatively evaluated using the following performance evaluation criteria: precision, recall, f1 score, and mean Average Precision (mAP).
One well-known metric which combines precision and recall, called the f1 score, is defined as:
The proposed model performance was evaluated using mAP to calculate the harmonic average of recall and accuracy.
The proposed models were trained and tested using freely available Google Colab with powerful GPUs without configuration requirements. The freely available LabelMe tool [
The experimental results focused on the following:
The performance of the proposed Guava Infected Patches Modified MobileNetV2 and U-Net (GIP-MU-NET) segmentation techniques was evaluated with the help of the Guava Patches Dataset (GPD). Guava Leaf Segmentation Model (GLSM) was applied to extract the region of interest (leaf) from leaf patches. The proposed Guava Multiple Leaf Diseases Detection Model performance was measured using Guava Leaf Diseases Dataset (GLDD) to identify and localize the guava multiple diseases on a single leaf. We compared the proposed GMLDD model with other YOLO variants.
The experiment used an Adam optimizer and a batch size of 04 with 125 epochs for training. The performance graph of the proposed model observed in
Images | Accuracy | |
---|---|---|
Test Set | 120 | 92.41% |
We suggest the following recommendations to improve the performance of the patch detection model: 1) Enhanced the dataset could improve its performance of the dataset. The deep learning methods required massive data samples for training; otherwise, overfitting occurred. Therefore, the dataset’s enhancement could resolve the overfitting problem [
The experiment used an Adam optimizer and a batch size of 4 with 50 epochs for training. The performance graph of the proposed model can be seen in
Images | Accuracy | |
---|---|---|
Test Set | 119 | 83.40% |
The experiment used an Adam optimizer and a batch size of 32 with 500 epochs for training. The graph of the proposed model can be seen in
The proposed model inference was calculated on unseen data (test set) of five classes, including anthracnose, healthy, insect attack, nutrient deficiency and wilt. The test set contained 98 images with 232 targets (labels) for all classes, as shown in
Class | Images | Targets | Precision | Recall | mAP @0.5 | mAP @0.5:0.95 |
---|---|---|---|---|---|---|
All | 98 | 232 | 73.3 | 73.1 | 71.0 | 50.3 |
Anthracnose | 98 | 53 | 45.4 | 54.7 | 44.3 | 14.0 |
Healthy | 98 | 24 | 93.3 | 100 | 95.9 | 92.0 |
Insect attack | 98 | 67 | 64.6 | 79.1 | 79.8 | 35.3 |
Nutrient def. | 98 | 69 | 66.5 | 31.7 | 35.5 | 11.9 |
Wilt | 98 | 19 | 96.3 | 100 | 99.5 | 98.2 |
The proposed method confusion matrix is presented in
In the p graph, as shown in
The precision-recall curve (PR-curve) represented the balance between precision and recall for various thresholds. A large area under the curve indicated high recall and precision, indicating a low false-positive rate, and high recall indicated a low false-negative rate. The results showed that our model expressed a high area under the curve, meaning thereby the proposed model achieved good performance. The anthracnose, healthy, insect attack, nutrient deficiency and wilt class achieved 44.3 mAP@0.5, 95.9 mAP@0.5, 79.8 mAP@0.5, 35.5 mAP@0.5 and 99.5 mAP@0.5, respectively, whereas all categories attained 71.0 mAP@0.5 (
The f score, sometimes known as the f1 score, measures a model’s accuracy on a dataset, as shown by the f1 curve in
The results showed that the healthy and wilt class expressed excellent precision, recall and mAP@0.5 (at > 93%); however, anthracnose disease had < 46% precision, recall and mAP@0.5, which was the lowest among all classes. On the other hand, the insect attack and nutrient deficiency class achieved more than 64% precision, recall and mAP@0.5. We suggest the following recommendations to improve the performance of the proposed guava multiple leaf diseases detection model:
The five classes in the GLDD used are significantly imbalanced. The infected class is highly underrepresented in the dataset. We can improve the performance of the proposed model by using the balanced dataset classes [ Increasing the dataset can improve the performance of the dataset. The deep learning methods require massive data samples for training; otherwise, overfitting occurs. Therefore, we can resolve the problem of overfitting by enhancing the dataset [ The complex background is another reason for misclassification. In the dataset, we used complex backgrounds, such as background colour, resembling leaf diseases, as shown in Another way to enhance the efficiency of the proposed method is by eliminating the curly or folded leaves from the training. The curly or folded leaf can mislead the insect eaten class, as shown in The small object sizes and annotation scheme are another reason for misclassification. In YOLOv5, only a rectangle is used to annotate the object. In guava leaf diseases, leaves are not in the proper shapes from the edges. And when we use a rectangle to annotate the object or disease area, the unnecessary information or area is also included, as shown in
To our knowledge, no work exists on multiple guava leaf disease detection from a single leaf. Therefore, a direct comparison with the existing method is not possible. However, we compare the performance of the proposed technique with other YOLO variants on the GLDD dataset. The values of precision, recall, and average detection time were compared with YOLOv3 [
Model | Precision (%) | Recall (%) | mAP@0.5 | Time (ms) |
---|---|---|---|---|
YOLOv3 | 66.2 | 65.0 | 65.1 | 48.205 |
YOLOv4 | 70.1 | 69.5 | 69.8 | 45.301 |
YOLOv4 tiny | - | - | 69.7 | 50.402 |
GMLDD model | 73.3 | 73.1 | 71.0 | 46.360 |
Real-time guava multiple leaf disease detection on a single leaf, guava infected patches and guava leaf segmentation were complex and challenging tasks. This research proposed a hybrid deep learning model to detect the multiple leaf diseases from a guava leaf and consisted of three models. The first model is GIP-MU-Net which segmented the infected region from the guava field. Then, the second GLSM model segmented the infected and healthy guava leaves. The last model is Guava Multiple Leaf Disease Detection (GMLDD), based on the YOLOv5 deep learning model. It detected and localized the multiple diseased regions from the infected area of the guava leaf. The performance of the deep learning models was evaluated on the self-acquired datasets: Guava Patches Dataset (GPD) and Guava Leaf Disease Dataset (GLDD). Both the datasets were developed by capturing the videos from the guava cultivation fields. Based on significant performance, we believed that the proposed multi-level deep learning model efficiently detected multiple diseases from a single instance, which helped the farmers increase the guava yield by detecting the guava disease early. It also helped to prevent the harsh effects of infectious diseases destroying the guava field and causing economic deprivation to the farmers. The proposed model was fully automated compared to the state-of-the-art existing leaf disease detection techniques. The plant leaves were manually captured and annotated to train the models. Moreover, the proposed model could be employed in real-time to detect plant leaf diseases. Moreover, the proposed model could be employed in real-time to detect plant leaf diseases. However, the proposed deep learning model is a prominent model for real-time detection of plant leaf diseases. Still, extensive development is required as research limitations of the environmental factors in various cultivation areas. The plant species represented different lesions in case of the same disease. The dataset was captured from a specific region and species of guava plants. These factors need to be addressed in the future to eliminate the effect of environmental changes, guava species, and disease appearance in different guava species.
The researchers would like to thank the Deanship of Scientific Research, Qassim University for funding the publication of this Project.