Revolutionizing Cucumber Disease Detection with Flooding-Based MobileNet V3

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Sep 7 2023

In an article in the press with the journal Computers and Electronics in Agriculture, researchers proposed a new flooding-based MobileNet V3 approach to identify cucumber diseases from cucumber leaf images in natural scenes.

Background

The cucumber is a creeping vine cultivated extensively for its cylindrical fruits that are commonly utilized as a vegetable. However, the cucumber yield across the world has been falling steadily owing to different pathologic diseases.

Traditional cucumber disease classification methods depend on manual diseased leaf inspection based on visual clues. However, the complex, labor-intensive, and time-consuming approach is susceptible to human errors, leading to unreliable outcomes and low classification accuracy/efficiency. Additionally, the shortage of agricultural experts who can accurately identify the diseases and the lack of expertise required for manual inspection among farmers are also adversely affecting the overall harvests. Thus, the development of technology and tools that can be used by farmers for early classification and detection of cucumber diseases has become crucial to overcome the challenges of manual inspection.

Mobile-based disease recognition technology can be an effective approach for cucumber disease classification. Mobile devices, such as smartphones, can be utilized to capture the images of cucumber leaves and then the captured images can be used to analyze and process the cucumber pathological results. Thus, an efficient and lightweight cucumber disease classification network can be developed that can be utilized on mobile devices to decrease the need for manual inspection.

Currently, few methods, such as volatile organic compound analysis, spectral analysis, and molecular analysis, are used for cucumber pathological analysis. However, these methods are expensive and cannot be implemented easily on a commercial scale.

Limitations of Computer Vision Processes

Computer vision holds significant potential for cucumber disease classification as the crop disease symptoms that appear as distinct features on plant leaves can be classified using proper strategies and image-based technologies. Specifically, the shape, texture, and color of diseased leaves in the images can be analyzed to detect and classify crop diseases.

However, the current computer vision methods cannot easily identify the fruit leaf diseases specific to China as these methods were trained predominantly using the PlantVillage dataset, which contains images obtained from farms in Switzerland and the United States (US).

Additionally, the datasets typically used to train models primarily consist of images captured by photographers and professional experts, while in real-world situations, most images are captured by farmers, who are not expected to capture images with perfect quality for analysis. The farmer-captured images also exhibit diverse sizes, colors, and backgrounds. Thus, datasets containing non-specialized leaf images must be used to train the disease classification models.

A Flooding-based MobileNet V3 Approach

In this paper, researchers proposed a flooding-based MobileNet V3 approach to identify crop leaves in ambiguous scenes. The proposed approach used a lightweight framework for effective recognition processing on mobile terminal devices to enable accurate and quick assessment of crop pathological conditions based on farmer-captured images.

Specifically, the algorithm can enable farmers to upload cucumber leaf images and classify them accurately. Seven kinds of cucumber leaf diseases can be classified precisely and quickly using the improved MobileNet V3-based cucumber disease classification network architecture/model.

Additionally, the network model can detect cucumber diseases from leaf images in natural scenes, leading to the realization of a tailored question-answering system for farmers. The accuracy of the network model can be improved by selecting appropriate batch capacities, optimizers, and parameters using the single-variable method.

Initially, batches of cucumber leaf image data were obtained from Chinese agricultural websites and then preprocessed through image size normalization. All images were captured randomly by farmers across China using mobile terminal devices.

The collected cucumber leaf images were annotated, and a Chinese cucumber leaf dataset (CCD) with complex image backgrounds was created, including seven types of cucumber leaf pathologies. Subsequently, researchers performed experiments to analyze and compare the network model, batch size, and optimizer using a single-variable approach to identify the optimal network model. Moreover, a novel training strategy, designated as the flooding method, was applied to the network model in place of the traditional loss evaluation strategy to mitigate the issue of overfitting.

Significance of the Study

Using the flooding method by replacing the loss evaluation strategy resulted in an additional 0.5% increase in the test set accuracy. The highest accuracy achieved using the proposed strategy on the CCD dataset used in this study was 83.3%. Additionally, the accuracy rates achieved for two public datasets, including Apple Disease and PlantVillage, selected for migration experiments were 98.1% and 99%, respectively, demonstrating the universality of the proposed flooding-based MobileNet V3 approach.

However, more research is required to investigate the use of more diverse datasets and advanced network architectures and the integration of large models such as Generative Pre-Trained Transformers to improve the performance of the disease detection algorithm.