Relevant diseases and treatment in cucumber farming

In order to identify the most relevant and common cucumber diseases, research was carried out and existing data sets were analyzed. This resulted in the following diseases: Anthracnose, Bacterial Wilt, Belly Rot, Downy Mildew, Pythium Fruit Rot and Gummy Stem Blight [6].

Anthracnose commonly affects cucumbers as a fungal ailment [40]. Throughout the growing phase, this condition can lead to the rotting of the fruit, canker on stems, spotting on leaves and blight affecting all parts of the cucumber plant that are exposed to the air [6]. During warm and moist summer seasons, severe outbreaks of this disease can result in premature leaf shedding, diminished crop yields and a decline in the quality of the fruits [6,41]. The most commonly observed signs of this disease in agricultural fields are spots on the leaves [42,43]. The fungicidal active ingredient Fludioxonil is effective in the treatment of Anthracnose [44]. According to Li et al. [44], the addition of certain adjuvants to fludioxonil improves the effectiveness and efficacy of the fungicide in the treatment of diseased cucumbers. Parada et al. [45] have shown that fresh mushroom substrate made from the mushroom Hatakeshimeji and its water extract is also highly effective against Anthracnose.

Bacterial Wilt is a common disease affecting cucumbers [46]. The primary symptom of this disease is wilting, which eventually results in the plant’s death [47]. As the infection advances, leaves may develop yellowing (chlorosis) and dead (necrotic) areas [6]. Typically, symptoms swiftly spread along individual vines, moving towards the base of the plant (crown) and ultimately leading to the death of the whole plant [6,47]. The management of Bacterial Wilt primarily focuses on preventive strategies, including the utilization of healthy seeds, the eradication of infected plant matter, the sterilization of tools and the selection of more resistant varieties of plants [48]. Although chemical pesticides are applied, their efficacy may be diminished due to the pathogen developing resistance [49].

Belly Rot in fruits is caused by the fungus Rhizoctonia solani [6,50]. This disease manifests on cucumbers as water-soaked lesions that are tan to brown in color, typically appearing at the blossom end and underside of the fruit [6,30,50]. As belly rot progresses, these lesions become swollen and crater-like, adopting an irregular shape and eventually drying up [6,50]. The treatment of Belly Rot in cucumbers can be challenging [51]. Traditional methods such as fungicides often do not provide sufficient control [51]. Research is investigating biological control methods with antagonistic bacteria and fungi, but these may show inconsistent results under production conditions [52]. One promising method is the use of genetic resistance through resistant cucumber varieties, which can provide an economical solution alone or in combination with other control methods [51].

Pythium Fruit Rot is caused by various fungus-like organisms belonging to the genus Pythium [6]. Initial symptoms include water-soaked, brownish lesions that rapidly expand, becoming large, watery, soft and rotten [6]. This type of rot typically affects the parts of the fruit that are in direct contact with the soil [6,53]. Often, a brown to dark green blister can be seen on the cucumber fruit before it turns watery and begins to decompose and finally die [6,54]. To control Pythium, chemical fungicides, temperature and moisture control, compost use and biological control strategies are emphasized [55]. Chemical fungicides such as Metalaxyl and Azoxystrobin are common, but their effectiveness varies by crop and Pythium species [55]. Biological approaches, such as certain bacterial strains, show promise in suppressing Pythium colonization and promoting plant growth [55,56]. The use of microorganisms that suppress the main causes of Pythium has already been used successfully on pepper plants [56]. A combination of these methods could provide a more effective strategy for controlling Pythium-related diseases [55].

Gummy Stem Blight, a fungal infection, impacts cucumber plants by producing brown or grayish-brown lesions on the leaves that may appear gummy or sticky [6]. This disease can significantly affect the plant’s growth and yield by causing leaves to wilt and eventually drop off [6,57,58]. In severe cases, the fungus can cause the stem of the cucumber plant to rot and break down, leading to the complete death of the plant [6]. The treatment of Gummy Stem Blight in cucumbers can be achieved through the development and use of resistant cucumber varieties [59]. This reduces the use of fungicides, which also has a positive effect on environmental and food safety risks [59]. Another option is the use of plant growth-promoting fungi [60]. This variant not only effectively treats the disease and boosts plant growth but also represents an efficient and environmentally friendly strategy [60].

Cucumber plants can fall prey to Downy Mildew, a widespread fungal condition, particularly in wet or humid climates [6,61]. Affected cucumbers display yellow or light green spots or blotches on the top side of their leaves, accompanied by a downy growth on the leaf’s underside, which can appear white or grayish-purple [6,61]. The afflicted leaf areas might form lesions that expand and merge, causing the leaves to turn brown or die off [61]. Leaves may also curl or become misshapen, ultimately drying up and detaching from the plant [6]. The treatment of Downy Mildew involves the use of various fungicides, including Oxathiapiprolin, which is highly effective against the pathogen Pseudoperonospora Cubensis [62,63]. Fungicides are selected based on their ability to reduce symptoms and protect yields [63]. Resistance to some fungicides has been observed, emphasizing the need to alternate fungicides with different modes of action to slow the development of resistance [62,63].

Deep learning in agriculture

Advancements in machine learning and computer vision have led to the development of new approaches for automating plant disease recognition through image processing techniques. Significant progress has been made in identifying and diagnosing leaf diseases in cucumbers and other plants using traditional machine learning methods such as support vector machine, AdaBoost, k-nearest neighbor and probabilistic neural networks (PNNs) [64–67]. Recently, deep learning-based algorithms have become dominant in image classification and recognition. They offer promising results and superior potential for feature extraction and recognition by processing input data through multiple non-linear functions, outperforming traditional shallow methods. Deep learning models, in particular, have demonstrated a significant amount of potential for the automatic identification and categorization of plant diseases and have shown remarkable superiority over traditional methods [30,68–70]. Numerous papers, previously introduced, use transfer learning for image recognition tasks [7,14,71].

Transfer learning is a well-established method in machine learning. Pre-trained models, developed using extensive datasets, are employed to train additional models that have limited training data available [72]. Traditionally, in transfer learning, the lower layers of a pre-trained network, such as VGG-19, are frozen to leverage the generalized features that have been trained on broad and diverse datasets like ImageNet [73]. Only the last layers, which are specifically added for the new task, undergo fine-tuning [74].

During the fine-tuning process, the previously frozen layers are gradually unfrozen. These layers are then trained at a reduced learning rate to allow fine adjustments to the features without disrupting the already learned and useful representations [75]. This method maximizes the utilization of the pre-trained network architecture and prevents excessive specialization on the training datasets, potentially enhancing the model’s generalization capabilities [73].

Related work on cucumber disease detection

As a result, numerous studies have been dedicated to identifying cucumber diseases using machine learning techniques. Table 1 presents a summary of these studies and provides an overview of the related work.

Some studies focus on determining the severity of a disease on a cucumber leaf.

Lin et al. [31] utilized a semantic segmentation model based on a convolutional neural network (CNN), inspired by the U-Net architecture, to classify the cucumber leaf disease Powdery Mildew on cucumber leaves, using a dataset consisting of 50 images of affected cucumber leaves taken under laboratory conditions. Their model segments Powdery Mildew on cucumber leaf images at the pixel level, enabling precise quantification of the infected areas. Lin et al. achieved an accuracy of 96.08%.

Ozguren [32] aimed to contribute to the detection of Mildew disease and determination of the disease severity using cucumber leaf images. A dataset with a total of 175 images was used to classify the extent of the disease into five classes. The author utilizes image processing techniques to enhance feature extraction within the images. A Faster R-CNN model was used to achieve an accuracy of 94.84%.

Bansal et al. [30] investigated the cucumber leaf disease "cucumber leaf spots" by the severity categorized into 5 classes. The used dataset contains 50,000 cucumber leaf images. The deep learning model ResNeXt50 was applied to this dataset, resulting in an accuracy of 97.81%. ResNeXt50 is a deep CNN that builds upon ResNet but introduces grouped convolutions, improving model efficiency and accuracy.

Other studies focus on the recognition of different cucumber leaf diseases.

Zhang et al. [33] proposed an advanced CNN for the recognition of three different cucumber leaf diseases. The used dataset contained images of cucumber leaves taken in the field. They employed a two-stage segmentation method combining GrabCut and Support Vector Machine algorithms to extract disease spots from cucumber leaves. This approach utilized color, texture, and border features to accurately segment lesion areas from images captured in field conditions. Zhang et al. [33] achieved an accuracy of 96.11%.

Hussain et al. [28] used a fusion of two deep learning models, VGG and InceptionV3, to detect five different cucumber leaf diseases. With the self-collected dataset, an accuracy of 96.50% was achieved. The authors focused on feature extraction and automatic feature selection in their work. First, they extracted features like color, texture, and shape descriptors. Then they identified the most relevant features, which they subsequently used for classification.

Khan et al. [34] compared four deep learning models: VGG16, ResNet50, ResNet101 and DenseNet201. The used dataset aimed to distinguish six different cucumber leaf diseases. By using DenseNet201, an accuracy of 98.40% was achieved. The authors also used automatic feature selection with a technique called Entropy-ELM. That technique allows to retain the most informative features while eliminating redundant or irrelevant ones.

Ma et al. [7] compared traditional methods, random forest and support vector machine, with a deep CNN. With a large dataset of 14,208 images, four cucumber leaf diseases were detected. The deep CNN achieved an accuracy of 93.40%.

Omer et al. [35] built a CNN model from scratch to detect five cucumber leaf diseases with an accuracy of 98.19%. In their work, the self-designed CNN is compared with three other, pretrained models (AlexNet, InceptionV3, ResNet50), using a dataset consisting of 4,868 images. The proposed CNN model consists of five convolutional layers, partly followed by batch normalization and ReLU activation functions. Max-pooling and dropout layers are interspersed. The network concludes with two fully connected layers, culminating in a SoftMax output layer.

Nguyen & Nguyen [36] utilized a recurrent residual U-Net deep learning model for feature extraction in their approach. Then a support vector machine classifier is used with the extracted features for classification. For the recognition of four different diseases, the authors achieved an accuracy of 96.33%, using a dataset of 154 images.

Tani et al. [37] focused on the recognition of multiple infections on cucumber leaves. In their work, 25 classes with 11 different diseases are considered. 13 of the 25 classes show more than one infection of the cucumber leaves. With a CNN-based VGG-Net model, the authors achieved an accuracy of 95.50% for the entire dataset. The diseases of the cucumber leaves with multiple infections were detected with an accuracy of 85.90%.

Jeny et al. [38] focused exclusively on the recognition of diseases on the cucumbers and not on cucumber leaves. A random forest model was applied to a composite dataset from field images and open-source images. With an accuracy of 85.84%, five different cucumber diseases were classified.

Mia et al. [14] compared traditional machine learning methods with CNN-based transfer learning models for the recognition of diseases on cucumbers and cucumber leaves. The used dataset contains both photos from the field and photos under laboratory conditions. The six diseases were most accurately recognized with a MobileNetV2 with 93.23%. The most accurate traditional method was random forest with 89.93%.

Junior & Majiid [39] already addressed the same dataset for the detection of six different cucumber diseases, healthy cucumber leaves and healthy cucumbers. The models EfficientNetB0 and ResNet50 were used to achieve an accuracy of 94.20% using an averaging technique. However, the highest accuracy of 94.98% was reached using only EfficientNetB0.

Overall, the studies reviewed in this section highlight the effectiveness of ConvNets and other machine learning methods in detecting cucumber diseases. The related studies show very good results exceeding 90%. But Table 1 also shows, that studies that do not focus solely on the cucumber leaves but also on the cucumber itself produce poorer and inadequate results. Considering that, there’s a need for more comprehensive approaches that simultaneously address diseases affecting both cucumber leaves and the fruit itself, ensuring a holistic assessment of plant health.

Given the still promising results, we address this and develop an accurate model for the detection of cucumber diseases and cucumber leaf diseases exceeding the existing benchmark. As well, this study addresses several diseases within the same dataset, offering a more comprehensive diagnostic tool. This advancement not only allows for disease identification but also for assessing the overall plant health.

Model architecture

The VGG19 model, originally introduced by Simonyan and Zisserman in 2014 [76], is a deep CNN comprising 19 layers. Trained on the ImageNet dataset, it excels in tasks such as image classification or object recognition [76]. For its input, the model necessitates an image size of 224x224 RGB pixels.

The model consists of five convolution blocks. These layers are responsible for extracting features from the input images. Equation 1 shows the mathematical equation of a 2-D convolution. During convolution, the kernel (w) is moved over the input image (x). A point (i,j) corresponds to a specific point in the output. The summation formula describes a loop over all positions (m,n) [77] .

(1)

Subsequently, to each convolutional block, a max-pooling layer is applied to downsample feature maps and reduce their spatial dimensions. The fully linked layers use these features to execute the final classification or recognition. Notably, the VGG19 model employs very small 3×3 convolutional filters and can contain up to 512 filters in each layer. The Rectified Linear Unit (ReLU) function is implemented to activate the hidden layers [76], facilitating the model in learning complex features from input images. Furthermore, the deep network architecture allows the model to capture hierarchical representations of the input images.

The selection of the VGG19 model for image recognition is justified by its architecture, which uses small convolution filters to deepen the network without greatly increasing parameters [76]. This design enhances its ability to differentiate features and improve generalization through implicit regularization. Demonstrating effectiveness in various large-scale image recognition studies [78,79], VGG19 also serves as a robust base for transfer learning, adapting well to new datasets by adjusting pretrained weights to meet specific application needs [80].

Our model modifies the VGG19 base up to the last max-pooling layer. We added a global average pooling layer to reduce overfitting and convert feature maps to a one-dimensional vector for processing by the dense layers [81]. The network then includes a dense layer with 480 neurons and ReLU activation, followed by a 20% dropout layer to further prevent overfitting. The final dense layer has eight neurons with softmax activation, configuring the model for specific recognition tasks. The architecture of the model is shown in Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Our model architecture.

https://doi.org/10.1371/journal.pone.0320764.g002

Process of training

We trained the model with different hyperparameters which we identified by using the Hyperband tuner, a powerful and efficient hyperparameter optimization strategy provided by Keras. Hyperband offers several advantages, including its efficient resource allocation, which results in faster convergence to optimal configurations [82]. It dynamically allocates resources to more promising configurations through its unique bracketing and halving approach, leading to significant time savings and improved performance in machine learning model training [82]. During the process, we tuned the hyperparameters batch size, dropout rate, learning rate and the number of units for the first dense layer. The batch sizes of 16 and 32 were evaluated, the learning rates of 0.001, 0.0001 and 0.00001 were considered, the dropout rate ranged from 20% to 50% and the number of units for the dense layers varied between 192 and 512.

Batch size, the count of samples processed by the model in a single iteration, was set to 16. The number of epochs, meaning the amount of times the entire training data set was passed through the model during the training process, was 53 after a callback function was triggered. This callback was activated after 20 epochs without any improvement in the validation loss. The learning rate determines the size of the steps that a model takes during training to adjust its weights and was 0.00001. We employed the Adam optimizer [83] algorithm as a stochastic optimization method. It is used for first-order gradient-based optimization of stochastic objective functions, leveraging adaptive estimates of lower-order moments [83].

Our approach involved the utilization of a VGG19 model that had previously been pretrained on the ImageNet dataset [76]. We initiated the model customization by first removing its top layer. We subsequently froze the weights of the first eleven layers to ensure their preservation during training, leaving the last eight layers adjustable. Considering the depth of the VGG19 model, we wanted to retain the knowledge encoded in the earlier layers while allowing the later layers to be adapted to our dataset. This approach strikes a balance between leveraging the general feature extraction capabilities of VGG19 and tailoring the model’s learned representations to our specific eight-class cucumber disease recognition task.

To adapt the model for our specific task of cucumber disease recognition, we introduced new layers into the architecture. These additional layers encompassed a global average pooling layer introduced by Lin et al. [84], designed to reshape the extracted features, a dense layer comprising 480 units and employing the ReLU activation function to capture high-level features, a dropout layer randomly dropping 20% of the neurons that are deactivated during training and an output layer featuring eight units, each corresponding to one of the eight classes within our dataset. These newly introduced layers were seamlessly integrated into the existing model, effectively replacing the previously excluded top layer.

For the traditional transfer learning approach, the hyperparameter tuning was nearly equivalent. Batch sizes of 16 and 32 were tested, learning rates of 0.1, 0.01, 0.001, 0.0001 and 0.00001 were selectable, dropout rates ranged from 0.2 to 0.5 and the number of units for the dense layers spanned from 96 to 512. The final model was trained with a batch size of 16, a learning rate of 0.001, 288 units for the dense layer and a dropout rate of 30%. The model underwent 87 epochs of training, with fine-tuning beginning after 50 epochs. Additionally, the learning was successively reduced by a factor of 10 until it reached 1e-9 at the end. For both the initial training process and fine-tuning, the same callback function as in the novel approach was employed but only triggered during the fine-tuning process.

Novel transfer learning approach

In comparison to ordinary fine-tuning approaches, our approach significantly modifies this traditional methodology by making selective layers, in our case the last eight, of the VGG-19 model adjustable from the beginning of the training process. This method enables direct and intensive use and adaptation of the deeper layers of the network for the specific task of disease detection in cucumbers [85]. Immediate adjustment of these layers allows the model to swiftly respond to the complex and specific features of the target images, which is particularly advantageous when dealing with limited datasets [85].

This strategy offers potential benefits in terms of adaptation speed and the model’s responsiveness to new data [86]. By making critical layers trainable from the beginning, the flexibility of the model to react to subtle differences in the training data and effectively utilize them is enhanced [86]. This approach could prove particularly advantageous in scenarios where specific features, not adequately represented in broad training datasets, are crucial. While this less conventional method may heighten the risk of overfitting, proper management and calibration of the model could ensure that the benefits of a quicker and more focused adaptation significantly outweigh the potential risks. This paper compares both the traditional fine-tuning approach and our novel transfer learning method. We analyze the performance, adaptability and efficiency of each approach in the context of disease detection in cucumbers.

Evaluation method

During the model training the performance metrics loss and accuracy were calculated using the validation dataset.

After the model training we evaluated it by measuring the performance indicators (2)-(7). The equations for accuracy, recall, precision and F1 score are derived from Sokolova and Lapalme [87]. Balanced accuracy as a metric is based on Brodersen et al. [88]. This measure is especially important when dealing with imbalanced datasets where it provides a more realistic evaluation by considering both sensitivity and specificity [88]. It’s essential for ensuring that the performance improvements are not skewed by any class imbalance within the dataset.

In a multi-class scenario involving eight categories, True Positives (TP) refer to correctly identified instances belonging to a specific class, while True Negatives (TN) represent accurately recognized instances not belonging to that class [87]. False Positives (FP) occur when instances are incorrectly labeled as belonging to a certain class and False Negatives (FN) arise when instances are incorrectly labeled as not belonging to that class [87]. These metrics are crucial for assessing and calculating the performance of classification models consisting of the accuracy of the test dataset, precision, recall and F1 score [87].

The test accuracy provides an overall view of model correctness, while the precision focuses on the accuracy of positive predictions, recall assesses the model’s ability to find all positive instances and the F1 score offers a balanced metric that considers both precision and recall. We also present a confusion matrix to further evaluate every class within the dataset.

(2)

(3)

(4)

(5)

(6)

(7)

Optical model evaluation

To enhance the explainability of our model after prediction, we looked at the "output" images of our model. Therefore we utilized Local Interpretable Model-agnostic Explanations (LIME), which is a method that generates local, easy-to-understand explanations for individual predictions of models such as deep neural networks.

LIME explains predictions from complex “black-box” models by approximating them locally with simpler, interpretable models. To explain a single instance, LIME creates new, slightly perturbed samples around that instance and obtains predictions from the black-box model for each perturbed sample. Next, it trains a simple surrogate model—often a sparse linear classifier or small decision tree—using these perturbed samples and their corresponding predictions, while weighting points closer to the original instance more heavily. Because this surrogate model is easy to interpret, its coefficients or rules reveal which features most strongly influence the prediction in the instance’s local neighborhood. By focusing on neighborhoods rather than the entire data space, LIME delivers accurate local explanations without requiring knowledge of the black-box model’s internal architecture [89]. The output of LIME is therefore a heat map that shows the areas of the input image on which the applied model makes its decisions.

Fig 3 shows the complete methodical procedure that we applied in this work.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Methodical procedure.

https://doi.org/10.1371/journal.pone.0320764.g003

Data preprocessing

During data preparation 1,280 images were resized to 224 x 224 pixels, labeled according to their class name and scaled to a range of [0, 1] for normalization using the rescale parameter with a value of 1/255. To assess the model’s performance, a hold-out cross-validation methodology was employed. We randomly divided the dataset into training (70%), test (20%) and validation (10%) subsets to mitigate the risk of overfitting to specific image sequences. We utilized the training subset to train the model, while performance metrics were computed using the validation subset following the training process. Importantly, the testing subset remained unseen by the model during the training phase and was solely employed to evaluate performance metrics after model training.

Next, we applied different data expansion techniques for the training set using various augmentation layers from Keras. Data augmentation is a pivotal technique employed to enhance the diversity of the training dataset and improve the generalization capability of deep learning models. In the context of this study, a comprehensive set of random transformations and operations was applied. This process introduces variations in the images and reduces the risk of overfitting.

The operations encompassed random horizontal and vertical flips and rotations of 30% in each direction. Zooms ranged between 75 to 125% from the original image size. The width was up to 20% in expanding as well as 20% in narrowing. In addition, the brightness and contrast of the images were up to 25%. Changes in height and translations varied between -15 to 15%. Collectively, these augmentations served to introduce variations and diversity into the training data, thus mitigating overfitting risks and enhancing model robustness. The complete data preprocessing steps are illustrated in Fig 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Data preprocessing.

https://doi.org/10.1371/journal.pone.0320764.g004

Dataset

The dataset we used includes 1,289 RGB images and is provided by Sultana et al. [6]. Fig 5 shows the eight types of cucumber classes, namely Anthracnose, Bacterial Wilt, Belly Rot, Downy Mildew, Pythium Fruit Rot, Gummy Stem Blight, Fresh Leaf and Fresh Cucumber. Each class consists of 160 images and is organized in a separate folder. The folder with images of Pythium Fruit Rot contains 169 images, including nine duplicates that we removed during the data preprocessing. The dataset is freely available on the Mendeley website.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Exemplary images from the classes of the dataset.

https://doi.org/10.1371/journal.pone.0320764.g005

The data was collected from a field in Basundia in Bangladesh under natural weather with inconsistent lighting conditions [6]. The images were captured in the size of 512 × 512 pixels using a Nikon D5300 digital camera with the cooperation of an expert from an agricultural institution [6].

Limitations

While our proposed approach demonstrates promising results, it is important to acknowledge its limitations. Although we achieved high internal validity by employing a hold-out split, external validation of the model has not yet been conducted. To address this gap, additional testing of the algorithm on datasets comprising multi-class plant disease data is required. The generalizability of the model may also be constrained by the specific characteristics of the dataset used for training and evaluation. Variations in image quality, lighting conditions and disease severity across different datasets may affect the model’s performance. By visually evaluating the classification results, we have come to the conclusion that Gummy Stem Blight in particular is often classified incorrectly. The data set will be expanded to perhaps more comprehensively map this class in future research.

Future work

To address the limitations identified in this study, several approaches for future research can be explored. Firstly, investigating the transferability of the trained model to diverse datasets or real-world scenarios is essential to enhance its generalizability. Due to the lack of existing datasets including cucumbers and cucumber leaves, more data needs to be gathered [6]. This analysis would provide valuable insights into the model’s robustness and adaptability across different environments and conditions. In this context, ensemble learning could be used to combine the findings of several models into a common model. In future research, we will train several models individually with the few existing datasets and then merge them using ensemble learning to achieve a unified robust model [90].

Furthermore, the exploration of ensemble learning techniques presents an exciting opportunity to boost the model’s performance. By combining predictions from multiple models, ensemble methods leverage diverse representations of the data, potentially leading to improved classification accuracy and robustness. Another related approach could be to directly analyze the video streams from, for example, the drones using combined architectures from Recurrent Neural Networks and CNNs, which has already yielded promising results in other domains [91].