Deep Learning for Disease Detection in Cauliflower Plants

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Oct 31 2023

In an article published in the journal Nature, the authors used Convolutional Neural Networks (CNN) to identify and detect various diseases in cauliflower crops. This research aimed to protect agricultural yields and human health by preventing the consumption of contaminated produce.

Background

Cauliflower, a vital crop in agriculture, is susceptible to various diseases that not only affect crop yields but also pose health risks to consumers. Detection and classification of these cauliflower diseases have garnered significant attention due to their economic and health implications. Several studies have delved into using advanced technologies to address this issue.

Existing research has explored the use of techniques like K-means clustering, CNNs, and traditional machine learning methods to identify and classify cauliflower diseases. These studies leveraged datasets, including VegNet, to develop algorithms for disease detection. Notably, some studies achieved high accuracy rates, with models like InceptionV3 demonstrating remarkable performance in disease classification.

Many studies used relatively small datasets, leading to concerns about class imbalances and the need for further fine-tuning to enhance accuracy. Additionally, the computational demands of certain models and the lack of diverse datasets hinder their practical implementation, and there are concerns about the limited generalization of these models.

The authors of the present study aimed to address these gaps by leveraging advanced deep-learning models for cauliflower disease detection. By combining these models, the study endeavored to create a more robust and widely applicable cauliflower disease detection system. The study also recognized the importance of reducing data dependency and sought to develop a model that can operate in real time and overcome the computational limitations of previous approaches. Overall, this research strived to enhance cauliflower production efficiency and safeguard both crop yields and human health.

Materials and Methods

Dataset: For this research, two sets of cauliflower images were used: an original image set and an augmented image set encompassing three disease classes and disease-free cauliflower. These images were collected from the Manikganj region in Bangladesh. In total, the dataset comprised 66 original images and 7,360 augmented images, categorized into various classes.

Data Pre-processing: The images, initially of size 224x224, underwent pre-processing. Using OpenCV, they were resized while preserving the aspect ratio. Grayscale conversion was applied to simplify the image structure. Morphological operations such as dilation and erosion were performed to manipulate the boundary pixels.

Exploratory Data Analysis: Visual representation of image pixel intensity was carried out. Histograms of red, green, and blue colors were presented to quantitatively show pixel intensities. Histogram Equalization (HE) expanded the intensity range, enhancing image contrast.

Feature Extraction: Several parameters were extracted from the images. These included area, aspect ratio, perimeter, equivalent diameter, max and min values, locations, and mean color intensity. Extreme points and contours were determined, and cropping was performed based on the largest contour. Adaptive thresholding was executed for image segmentation, where different threshold values were calculated for various image regions.

Data Division: The dataset was divided into training and validation sets in a 3:1 ratio. The training set contained 1,384, 1,440, 1,648, and 1,416 images from the bacterial spot rot, black rot, downy mildew, and no disease classes, respectively. The validation set consisted of 346, 360, 412, and 354 images from the same classes.

Applied Models: Several models were applied to detect and classify cauliflower diseases, including Xception, EfficientNet B0, B1, B2, B3, B4, MobileNetV2, DenseNet201, ResNet152V2, and InceptionResNetV2. Various evaluation metrics assessed their performance, including accuracy, loss, root mean square error (RMSE), precision, recall, and F1 score.

Study Findings

The evaluation results showed that EfficientNetB1 achieved the highest accuracy (99.37%) during training, while Xception had the best loss and RMSE values. On the other hand, ResNet152V2 had the lowest accuracy (52.42%) during training.
Additionally, for specific parameters and classes, models exhibited variations in performance. For instance, Xception showed high precision, recall, and F1 score for most classes. EfficientNetB1 performed well in several classes regarding accuracy, precision, recall, and F1 score. DenseNet201 also demonstrated strong performance, especially for recall, precision, and F1 score, while MobileNetV2 had relatively lower performance, particularly for black rot and no disease classes.

The confusion matrix and corresponding metrics were computed for different disease classes (bacterial spot rot, black rot, downy mildew, and no disease), indicating how well the models performed for each class. Overall, EfficientNetB1 and Xception consistently showed strong performance in various classes. The computational time for model training and evaluation varied, with ResNet152V2 taking the longest (1 hour and 36 minutes) and EfficientNetB0 the shortest (24 minutes and 49 seconds).
Comparing the accuracy of the proposed system with existing techniques, the research achieved higher accuracy (99.90% with EfficientNetB1) compared to previous methods. EfficientNetB1 outperformed all other models in the study.

Conclusion

In conclusion, this research illustrates how technology can enhance agriculture by developing an efficient and accurate cauliflower disease detection system. It utilized 7,360 images from four classes, applying image preprocessing and contour feature extraction techniques. Ten models were trained, with EfficientNetB1 achieving the highest validation accuracy of 99.90%.

While some limitations were noted, including potential misclassification, future work will focus on improving contour feature extraction and expanding the system to detect diseases in various vegetables. Overall, this research shows how deep learning models can boost agriculture, increase yields, and ensure global food security. The ongoing refinements promise a lasting positive impact on farming practices.