Computationally light deep learning framework to recognize cotton leaf diseases

Abstract

Cotton is an important commodity because of its use in various industries across the globe. It is grown in many countries and is imported/exported as a cash crop due to its large utility. However, cotton yield is adversely affected by the existence of pests, viruses and pathogenic bacteria, etc. For the last one decade or so, several image processing/deep learning-based automatic plant leaf disease recognition methods have been developed but, unfortunately, they rarely address the cotton leaf diseases. The proposed work presents a simple yet efficient deep learning-based framework to recognize cotton leaf diseases. The proposed model is capable of achieving the near ideal accuracy with early convergence to save computational cost of training. Further, due to the unavailability of publicly available datasets for this crop, a dataset is also collected comprising of three diseases namely curl virus, bacterial blight and fusarium wilt in addition to the healthy leaf Images. These images were collected from the Internet and fields of Southern Punjab region in Pakistan where the cotton crop is grown on thousands of acres every year and is exported to the Europe and the US either as a raw material or in the form of knitted industrial/domestic products. Experimental results have shown that almost all variants of our proposed deep learning framework have shown remarkably good recognition accuracy and precision. However, proposed EfficientNet-B0 model achieves 99.95% accuracy in only 152 seconds with best generalization and fast inference.

Keywords

Cotton leaf disease efficientnet mobilenet deep leaning agriculture

1 Introduction

Pakistan is the 4^th largest cotton producing country in the world. The crop is pronounced as "White Gold" because it contributes around 55% share in the country’s export. According to the official statistics, the estimated cotton crop production in the year 2020 - 21 remained the lowest in the last 20 years. This is due to the lack of technology in seed production and disease management [1]. Consequently, the number and type of species that harm cotton plant are continuously increasing. The viruses regularly mutate themselves and may vary from region to region [2]. Common cotton diseases [3] include cotton leaf mosaic, fungal root rot, angular leaf spot, alternaria leaf spot, fusarium & virticillium wilt and the most blatant one, the leaf curl virus. Generally the symptoms of the foliar diseases appear on the front and the back sides of the leaves. They may vary at different stages of plant growth and can be diagnosed inaccurately even by expert pathologists [4]. Therefore, incorporating the use of technology in disease diagnosis will not only help the farmer but will also increase the crop’s yield.

By understanding the cause and origin of a disease, its control management strategies can be devised accordingly. For example, cotton curl leaf disease [5], caused by the geminiviridae family of viral pathogen, is transmitted by whitefly. Further, the begomovirus specie is considered the most devastating pathogen across Pakistan and India [6] for the cotton plant. The spread and control of the disease is a challenge for researchers since its first outbreak in 1912 [7]. The symptoms start from thickening of the veins and then culminate with upward and downward curling of the leaves as shown in Fig. 1. Since the virus has several strains and mutations, the spread can only be managed by the implications of several long-term strategies [8].

Fig. 1

Snapshots of cotton leaf curl diseases [7].

Another biotic pathogen considered as a major threat to crop production all over the world and specifically in Pakistan is bacterial blight caused by Xanthomonas [9]. This soil-borne pathogen was first noticed in Pakistan in the early 70’s. Organism attacks the susceptible host leaf from the underside of the leaf. In favorable conditions, the symptoms develop from small water-soaked lesions to black streaks as the disease worsens [10] as shown in Fig. 2.

Fig. 2

Snapshots of cotton bacterial blight disease [10].

Fusarium wilt, a soil-borne fungal pathogen [11] is also a major disease for cotton crop. The fungal pathogen Oxysporum causes devastating loss to cotton yield in favorable conditions specially in countries like Australia, India, China, and Pakistan [11]. The fungi may start appearing at the seedling stage and can spread in the direction of flow of water. The onset of disease looks like vein clearing on the younger leaves followed by stunting, yellowing of the lower leaves, defoliation, necrosis and finally plant death [12] as shown in Fig. 3. The pathogen existence can be confirmed by splitting the host plant stem and then the presence of vascular brown lines confirm the attack of Oxysporum.

Fig. 3

Snapshots of cotton fusarium wilt disease [13].

Considering the adverse effects of these diseases on crop production, necessary and effective measures must be implemented using precision agriculture [14] tools of image processing. Tremendous advances in machine learning and deep learning techniques have inspired researchers for early and accurate identification of these biotic and abiotic stresses [15, 16]. A comprehensive survey of automatic plant leaf disease recognition techniques to identify cotton leaf diseases using conventional image processing techniques is carried in Section 2. It is interesting to note that cotton is the least addressed crop in automatic recognition of plant/crop leaf disease recognition. Further, the most famous and commonly used PlantVillage dataset in this field also does not include cotton leaf images. The authors in [16], presented a survey of more than 45 recent research contributions employing deep learning-based techniques for plant leaf stress identification; but none of them discusses cotton leaf diseases.

Training a deep learning network is computationally intensive and requires the use of specialized hardware commonly known as Graphical Processing Unit (GPU). Keeping in view the costly GPU hardware, the availability of cloud-based facilities has increased enormously for the last five years or so. Recently, the concept of transfer learning has also gained popularity where a major part of a pretrained model, previously trained on a large image dataset, can be used. The final fully connected layer in the original network is replaced with another layer with output neurons equal to the number of classes in the targeted application. By virtue of the transfer learning, pretrained versions of the famous deep learning models can be reused on other datasets using hardware with limited parallel processing capabilities.

With the rapid increase in the use of smartphones the need of develop light-weight plant disease recognition applications has increased too. MobileNets are frequently being used in many applications. A smartphone based mobile application will be more handy for a farmer to use in disease recognition as compared to any other system. But the computational limitations and hardware constraints of mobile phones make the use of traditional Convolutional Neural Networks (CNNs) almost impossible. Light-weight CNNs were proposed [17] to overcome the computational constraints. In this connection, MobileNet was proposed by Howard et al. [18] in 2017 for embedded applications. The computational is greatly reduced with almost same precision as it uses depth-wise convolution discussed in Section 3.

In the jargon of superb famous deep learning models, EfficientNet [19] is a recently proposed Light-weight deep learning model that has gained enormous popularity in a very short period. The network has eight sub-models similar to each other in their architecture. As mentioned earlier, the use of deep learning models for cotton crop disease recognition is very limited, the use of EfficientNet and MobileNet is a logical choice for this task.

EfficientNet was proposed by Tan et al. in 2019 [19]. It offers state-of-the-art accuracy with fewer parameters in several image classification algorithms [19] using a novel clever scaling method. It can be thought of as a logical extension to the MobileNet. Which is based on depth-wise separable convolution and has gained popularity for real-time applications running on low memory smartphones.

To the best of our knowledge, deep learning techniques have rarely been used for cotton leaf disease recognition [20]. The absence of publicly available cotton leaf disease dataset is another constraint. Although Digipathos [21], a publicly available dataset does contain cotton leaf disease images but they are very small in number and are not sufficient to evaluate the performance of a deep learning model on them. Increasing trend on the use of deep learning models [22] for crop leaf diseases led us to an interesting gap that the existing literature rarely addresses. Further, the need for an accurate and computationally light-weight model [17] with fewer parameters is inevitable for this particular crop. Owing to this, we have proposed a light-weight framework using eight versions of recently proposed models namely EfficientNet [19] and MobileNet [18] to recognize commonly found cotton leaf diseases. The performance of models are compared with each other and with other deep learning and machine learning models in terms of training time, parameters, accuracy, and network convergence on our self collected cotton leaf dataset. The major contributions of this research are as follows:

Addressing cotton crop leaf diseases a least addressed field, is investigated and near 99% accuracy is achieved using light-weight deep learning framework based on different versions of EfficientNet and MobileNet models.

Unavailability of cotton crop disease data is addressed by introducing a dataset comprising of 1, 711 images divided into 4 classes. Moreover, the images in the dataset also include background information to model the real-world scenarios. In contrast, in many widely used datasets in this field, e.g., PlantVillage and Digipathos etc., images are captured on plain background

The remainder of this paper is organized as follows: Section 2 presents a survey of cotton plant leaf disease recognition techniques based on conventional machine learning and deep learning approaches. Section 3 presents the details of our proposed deep learning light-weight framework to recognize plant leaf diseases for the cotton crop with the help of transfer learning. Section 3 presents details of the dataset, the experimental setup and the results of the experiments conducted on self-collected cotton disease dataset. It discusses the analysis of experimental results too. Finally, Section 4 gives the conclusion of this research and suggests possible future improvements.

2 Literature survey

For the last one decade or so, researchers are continuously trying image processing techniques for suitable evaluation and diagnosis of leaf disease in agricultural fields [23]. As the plant pathogens cause visible symptoms in the host, computer vision and image processing based techniques are the best to recognize the symptoms automatically for the purpose of diagnosis and possible treatment [23].

Several machine learning-based techniques have been proposed with marvellous results in an effort to recognize the cotton leaf diseases [14, 24]. Authors in [3] presented cotton disease classification system using support vector machine (SVM) classifier and also proposed an android application for soil monitoring and disease management system. Several machine learning and pre-processing techniques have been compared in [25]. This paper also summarizes several background removal and feature extraction techniques used by researchers for cotton leaf diseases using machine learning methods. Jenifa et al. [26] proposed a multi-support vector machine based classification system for four common cotton leaf diseases and compared the results with CNN. Cotton leaf spot diseases were identified and classified using SVM classifier in [27] and authors investigated the results for different feature extraction techniques too. Three common cotton leaf diseases were classified using back propagation neural network [28] with a classification accuracy of 85%.

More recently, deep learning based methods have got their way in the field of automatic plant leaf disease recognition but unfortunately their use to identify cotton leaf diseases is very limited. We could find only one significant contribution [20] where the authors used a CNN model to classify four diseases of cotton plant leaves and obtained an overall accuracy of 96% on their self collected dataset.

However, several recently proposed deep learning models have extensively been used to classify plant leaf diseases for other crops in the last few years. In a recently published work, authors of this manuscript have presented a review of 45 state-of-the-art and recently proposed deep learning based techniques proposed for different plants [16].

An improved CNN with considerably lesser number of parameters was presented by Zhang et al. for the maize crop [29] and obtained 98.8% accuracy. Chen et al. [30] presented a computationally light-weight version of the famous MobileNet model and experimented with the publicly available PlantVillage dataset in addition to their self collected data.

Deeper and complex networks with millions of parameters require training cost and time but are presented with many redundant parameters that have little or no effect on the classification accuracy [31]. Such parameters can be quantized, compressed or unnecessary connections can be skipped [32]. Owing to the need for such low-cost models, Barman et al. [33] compared the performance of MobileNet with their self-structured convolutional neural networks for citrus leaf disease. Several recently proposed deep learning based models like EfficientNet [19] and MixNet [34] etc. have obtained amazingly large accuracy in various fields including plant leaf disease recognition. The authors in [35] used B4 and B5 versions of the EfficientNet model to classify diseases of various plants found in the famous PlantVillage dataset. An algorithm namely “Few-shot learning” was presented by [36] for disease identification using smaller datasets. Applying this algorithm to deep learning networks enhances the recognition accuracy significantly. Further, thin and depth-wise separable models were proposed by Kamal et al. [37] for real-time crop disease diagnosis especially for resource constraint mobile devices, and achieved 98.34% classification accuracy on PlantVillage dataset using Reduced MobileNet. Low memory usage, high recognition accuracy and speed were attained by Liu et al. in tomato leaf disease recognition using MobileNetV2-YOLOV3 model [38].

3 Proposed computationally Light-weight deep learning model

It is well known that every iteration in a deep network contains several convolutional operations that require a lot of computational resources. It is evident from the current advancements in this field that the attainment of the best results can be achieved at a price [39]. The quest for bigger training data and deeper/denser networks have further escalated the use of specialized hardware equipment like graphical processing unit(GPU) and tensor processing units (TPU). Several regularization and optimization techniques [40] have relieved the complexity burden of deep networks, but it still needs to be further addressed. To evade this hardware complexity and computational cost, MobileNets [18] were proposed in 2017. The basis of these deep networks are depth-wise separable convolution layers that are a factorised version of the depth-wise and point-wise convolution layers. The output feature map G_k,l,n for standard convolution is for input (D_F, D_F, M)

$G_{k, l, n} = \sum_{i, j, m} K_{i, j, m, n} . F_{k + i - 1, l + j - 1, m}$ (1) here i, j are related to image indices and m, n deals with kernel indices. D_F is the spatial width and height of the input feature map and D_K is the spatial dimensions of the kernel. F is the input feature map and K is the kernel size with indices i, j showing stride of 1. Computational cost for standard convolution is calculated as:

$D_{K} . D_{K} . M . N . D_{F} . D_{F}$ (2) where M is the number of input and N is the number of output channels. The kernel size for spatial convolution is D_K × D_K with M channels and feature map size D_F × D_F. Input feature map F and the output feature map are shown in Equation 1. The convolutional operation filters features and combine them on the basis of kernels. This process of filtering and combining is split into factorized convolution known as separable convolution which can be written as [18] ${\hat{G}}_{k, l, m} = \sum_{i, j} {\hat{K}}_{i, j, m} . F_{k + i - 1, l + j - 1, m}$ (3)

In the above Equation $\hat{K}$ is the convolutional kernel of size D_K × D_K × M and m^th filter in $\hat{K}$ is applied to the m^th channel in F to produce a feature map of $\hat{G}$ .

The computational cost of depth-wise convolution is reduced and is given as $D_{K} . D_{K} . M . D_{F} . D_{F}$ (4)

Depth-wise convolution combined with point-wise convolution is called depth-wise separable convolution [37] as shown in Equation 3. The computational cost of which is shown as: $D_{K} . D_{K} . M . D_{F} . D_{F} + M . N . D_{F} . D_{F}$ (5)

MobileNets use depth-wise separable convolution and take 9 times lesser time compared to the standard convolution. MobileNetV2 is considered as one of the most optimal network with fewer parameters with good accuracy [41]. However, the EfficientNets [19] use the same type of convolutional operations and became very popular in the field of deep learning right after their inception in 2019.

Ever increasing use of handheld devices in our daily life led to the visualization of efficient mobile based deep networks such as MobileNets [18] and SqueezeNet [42]. In an effort to enhance the previous work, Tan et al. [19] proposed an efficient CNN. The compound scaling method proposed in EfficientNet can be adjusted with the size of input image. The authors presented a superlative equation for optimally scaling the network depth-wise, width-wise and resolution-wise.

The compound scaling is expressed as:

$depth, d = α^{φ}$ (6)

$width, ω = β^{φ}$ (7)

$resolution, r = γ^{φ}$ (8) where φ is a user specified coefficient to specify the amount of resources available and variables α, β and γ in Eqns. 6, 7 and 8 elaborate to specify the user-defined resources in scaling the width, resolution and depth. For EfficientNet-B0, the baseline variant, we assume φ = 1 which allows to scale the depth, width and resolution by 20%, 10% and 15% respectively. The same formulae derived after an extensive grid-search technique presents EfficientNet variants from EfficientNet-B0 to EfficientNet-B7. The EfficientNet model was found 6× faster than other deep networks due to the reduced parameters and has achieved larger accuracy on many existing datasets [19]. This breakthrough on the use of CNNs and with ever-increasing need to implement the crop-disease detection system on handheld devices motivated us to use EfficientNet for this work.

EfficientNet targeted the resource requirements in a very effective way by exploiting the advantages of both MobileNets and ResNets. A conceptual block diagram of the baseline EfficientNet model is shown in Fig. 4. It uses a special MBConv operation which is based on the inverted bottle neck depth-wise convolution [42] used previously in MobileNet [18]. The MBConv operation is optimized using a technique called Squeeze and Excitation operation [42]. Each block in the network starts with a depth-wise separable convolution and ends with a linear activation function. The varying kernel size k specifies the height of the 2D convolutional window. The variants from EfficientNet-B1 to EfficientNet-B7 have similar structure but the architecture becomes more and more complex and the number of parameters keep on increasing.

Fig. 4

Architecture of baseline EfficientNet model.

In this work, we present a deep learning framework with less parameters, better accuracy along-with early convergence. EfficientNet and MobileNet based cotton crop disease recognition systems were examined on our self collected dataset.

A block diagram of the proposed deep learning system based on the EfficientNet model and MobileNet model is shown in Fig. 5. The self-collected dataset is divided in to train and test splits whose further details are given in Section 3. The top layer of the EfficientNet’s model in use (EfficientNet-B0 to EfficientNet-B7) is replaced with flatten layer, which converts the output of convolution layer to 1-D array. Dropout is added to prevent the model from getting overfit. After adding the dense layer, a fully connected layer is used that has four output neurons corresponding to 4 classes in our dataset as shown in Fig. 5(a). The proposed model of MobileNet and MobileNetV2 are shown in Fig 5(b). Using the transfer learning approach we transferred the weights of imagenet to the next dense layer followed by batch normalization layer which reduces the training epochs and make the network converge more quickly. A dropout of 25% was added before another dense layer to effectively train all features. Several parameters are adjusted for each EfficientNet and MobileNet model to achieve an optimized deep learning workbench. Details of the parameters adjustment is given as under:

Since the EfficientNet and the MobileNet models are already trained, we used the same internal weights while employing them in our proposed deep learning framework.

All versions of the EfficientNet model (i.e., EfficientNet-B0 to EfficientNet-B7), the MobileNet and the MobileNetV2 were employed separately in our proposed framework.

In pursuance of larger recognition performance, batch sizes of 16 was used throughout the experimental work. Moreover, architecture of our proposed network is simple to keep the number of trainable parameters small.

For all experiments, Adam optimizer is used for EfficientNet models and Stochastic Gradient Descent (SGD) for MobileNets respectively because these settings were found to provide the best performance for the respective deep learning models in our proposed framework. The structural block diagram of the proposed deep learning framework employed is shown in the Fig. 6.

Fig. 5

Block diagram of the proposed deep learning framework.

Fig. 6

Functional diagram of proposed work.

Finally, the trained model with four output neurons is tested on test data for predictions. The performance of the trained model is evaluated using various evaluation metrics mentioned in Section 3.

4 Experiments and results

4.1 Dataset

An appropriate dataset is required for training of the proposed recognition model. Since there is no publicly available dataset for cotton leaf diseases, we collected more than 800 images from the internet and field which were further divided into four classes; 3 classes corresponding to three diseases and one healthy class. Details of the dataset used for experimentation is shown in Table 1. The in-field images were collected by Samsung A21s 48 megapixels mobile camera under different lightning conditions. To enhance data diversity plenty of annotated images were also downloaded from search engines.

Table 1
Details of the cotton leaf dataset used in the proposed work

Class No. of Images

Curl 418

Bacterial Blight 448

Fusarium Wilt 419

Healthy 426

Total 1,711

Class	No. of Images
Curl	418
Bacterial Blight	448
Fusarium Wilt	419
Healthy	426
Total	1,711

Since a reasonably large dataset is required to train a convolutional neural network, additional version of all images are created through augmentation techniques. This is achieved by applying rotation of 90°, 180°, 270° and 360°. Further, to reduce overfitting, contrast and saturation of collected images were also enhanced and multiple versions were created to replicate the infield lighting conditions. A programming subroutine was written to automatically expand the collected dataset. The resulting dataset finally contains 1, 711 images with complex background and varying lighting conditions. Some sample images from various classes are shown in Fig. 7. Images were manually annotated after careful examination with keyword search and were later labelled by the corresponding disease acronym.

Fig. 7

Snapshots of images corresponding to four classes in our dataset. From left to right, Healthy, Cotton Curl, Cotton Bacterial Blight and Cotton Fusarium Wilt.

4.2 Experimental setup

We have used the pretrained models of EfficientNet and MobileNet and MobileNetV2 using transfer learning approach [43]. In our proposed network, we removed the top layer while the other layers already trained on the imagenet [44] were retained to save the time and effort. To keep the model simple, the final fully connected layer of 1, 000 classes was replaced with a new layer having 4 output neurons representing four classes of our data as shown in Fig. 5.

Pre-experimentation revealed that RMSProp and SGD optimizers perform poorly for the EfficientNet model and hence were not used for the formal experimentation work of this research. We experimented with Adam optimizer only whose learning rate was set to 0.001 for all experiments. To further improve our training process, we invoked early stopping [45] with a patience value of 5. It helped to reduce the chances of overtraining or overfitting the model by halting the training epochs as soon as the validation accuracy stops improving. EfficientNet-B0 model achieved 99.95% accuracy on our dataset at the 8^th epoch. The variants EfficientNet-B6 and EfficientNet-B7 achieved similar accuracy values at 16^th and 18^th epochs respectively.

As far as fine-tuning of hyper-parameters is concerned, it was concluded that keeping the batch size small improves the convergence time. Therefore, it was kept equal to 16 to obtained the optimized performance. All the images were normalized by dividing them by 255. Image data generator in Python accepts the images in batches and randomly transforms and replaces the images by applying rotation, shear and resize operation etc. Hence the network is trained on transformed images and consequently, a better generalization is achieved. By using the image data generator class from the Keras library, we have performed the augmentation operations. The details and range of techniques applied are summarized in Table 3. We have proposed two other models using the streamlined architecture of MobileNet [18] that uses the same depth-wise separable convolutions making it another light-weight architecture for embedded applications. We introduced an additional dense layer as shown in Fig. 5(b) to the pre-trained versions of MobileNet and MobileNetV2. Batch normalization is used to reduce the internal co-variate shift problem and for better initialization of the parameters [46]. After the network parameters were fine tuned, a 25% dropout was also introduced to avoid overfitting. The classification accuracy of our dataset is analyzed by using another computationally low cost model namely MobileNet and MobileNetV2 that proved less efficient and took more epochs to converge. Pre-experimentation also revealed that SGD optimizer with a learning rate of 0.001 and decay of 4 × 10^-5 is the best choice for the pretrained MobileNet model. Early stopping was also used in the training process but it took 27 and 29 epochs for the MobileNet and MobileNetV2 models to reach a classification accuracy of 98.25% and 99.12% respectively.

Table 2
Parameters of the proposed CNN based on the EfficientNet and the MobileNet models

Parameters EfficientNet MobileNet

Batch size 16 16

Dropout 0.25 0.30

Learning rate 0.001 0.001

Decay 5 × 10^-5 4 × 10^-5

Batch Norm. No yes

Optimizer Adam SGD

Parameters	EfficientNet	MobileNet
Batch size	16	16
Dropout	0.25	0.30
Learning rate	0.001	0.001
Decay	5 × 10^-5	4 × 10^-5
Batch Norm.	No	yes
Optimizer	Adam	SGD

Table 3

Augmentation techniques applied on our dataset

Augmentation Technique	Values
Rescale	1/255
Rotate	25
Width Shift	0.1
Height Shift	0.1
Shear	0.2
Zoom	0.2
Horizontal Flip	True
Fill Mode	Nearest

For the experimentation of this research, Keras and Tensorflow libraries of Python were used on Kaggle, a free online platform for deep learning experiments. Kaggle is equipped with a Tesla P100-PCIE 16GB GPU to facilitate training of deep learning models. Further, the dataset used for this research is available for interested researchers 1 .

4.3 Evaluation metrics

To analyze the performance of the framework for our particular cotton crop disease classification problem, we have used a number of important performance metrics such as validation accuracy, loss, precision, recall and F1_score. Further, True positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN) for each category. Mathematical expressions to compute accuracy, precision, recall are given in Eqns. 9 through 11.

$Accuracy = \frac{TP + TN}{TP + TN + FP + FN}$ (9) $Precision = \frac{TP}{TP + FP}$ (10) $Recall = \frac{TP}{TP + FN}$ (11)

F1_score is the harmonic mean of precision and recall and is given in Equation 12.

$F 1_score = 2 \times \frac{Precision \times Recall}{Precision + Recall}$ (12)

To further investigate the prediction of individual classes, Recall and F1_score for each version of the various deep learning models are discussed in detail in Section 4.4.

4.4 Results

This section presents the results of the experiments performed on our proposed deep learning framework that consists of eight modified versions of the EfficientNet MobileNet and MobileNetV2 models.

4.4.1 Accuracy and Recall

Table 4 presents a comparison of the total number of parameters, time per epoch during training, validation accuracy and loss for the eight versions of the EfficientNet and two versions of the MobileNet. The comparison is also extended to other state-of-the-art machine learning and deep learning techniques. Deeper networks like VGG16 and ResNet50 presents competitive results in terms of validation accuracy/loss but at the cost of number of parameters and training time. Figs. 8 and 9 explain the accuracy and loss versus number of epochs for different EfficientNet and MobileNet models. It can be witnessed that almost all the models achieve ideal accuracy/loss performance. However, EfficientNet-B0 and MobileNetV2 give better generalization with fastest training time for the same data. But at the same time MobileNetV2 takes 28 epochs to converge and Efficient-B0 achieves the same in just 8 epochs.

Table 4
Accuracy and Loss computed on both deep learning and machine learning based methods

Model Trainable Parameters (Millions) Time per Epoch (sec) Valid. Acc. Valid. Loss

EfficientNet-B0 4.01 19 0.9995 0.0025

EfficientNet-B1 6.52 20 0.9532 0.0665

EfficientNet-B2 7.70 21 0.9854 0.0229

EfficientNet-B3 10.70 22 0.9854 0.0203

EfficientNet-B4 17.56 24 0.9971 0.0052

EfficientNet-B5 28.35 33 0.9708 0.0651

EfficientNet-B6 40.74 40 0.9993 0.0011

EfficientNet-B7 63.79 53 0.9991 0.0010

MobileNet 4.26 17 0.9825 0.0207

MobileNetV2 3.54 17 0.9912 0.0133

ResNet50 24.7 59 0.998 0.067

VGGNet16 15.6 50 0.994 0.025

SVM – – 0.93 –

Model	Trainable Parameters (Millions)	Time per Epoch (sec)	Valid. Acc.	Valid. Loss
EfficientNet-B0	4.01	19	0.9995	0.0025
EfficientNet-B1	6.52	20	0.9532	0.0665
EfficientNet-B2	7.70	21	0.9854	0.0229
EfficientNet-B3	10.70	22	0.9854	0.0203
EfficientNet-B4	17.56	24	0.9971	0.0052
EfficientNet-B5	28.35	33	0.9708	0.0651
EfficientNet-B6	40.74	40	0.9993	0.0011
EfficientNet-B7	63.79	53	0.9991	0.0010
MobileNet	4.26	17	0.9825	0.0207
MobileNetV2	3.54	17	0.9912	0.0133
ResNet50	24.7	59	0.998	0.067
VGGNet16	15.6	50	0.994	0.025
SVM	–	–	0.93	–

4.4.2 Per Class Performance

In order to further investigate the performance of the deep learning framework, four metrics namely precision, recall, F1_score and support are computed for each class. Table 5 gives these results in detail. The baseline version of EfficientNet-B0 achieved the maximum precision, recall and F1_score measure for all four classes. MobileNetV2 also obtained a decent set of evaluation measures for all four classes but performed the best for images infected with bacterial blight and fusarium wilt diseases. Table 5 also shows that comparable performance can also be obtained using more complex versions of the EfficientNet model but of course at the cost of more computational burden.

Table 5
Evaluation metrics on per class basis for various deep learning models

Network Class Precision Recall F1-score Support

EfficientNet-B0 Bacterial blight 1.00 1.00 1.00 86

Curl 1.00 1.00 1.00 80

Fusarium wilt 1.00 1.00 1.00 91

Healthy 1.00 1.00 1.00 85

EfficientNet-B1 Bacterial blight 0.97 1.00 0.98 86

Curl 0.87 1.00 0.93 80

Fusarium wilt 1.00 0.95 0.97 91

Healthy 1.00 0.87 0.93 85

EfficientNet-B2 Bacterial Blight 1.00 1.00 1.00 86

Curl 0.94 1.00 0.97 80

Fusarium Wilt 1.00 0.95 0.97 91

Healthy 1.00 1.00 1.00 85

EfficientNet-B3 Bacterial Blight 1.00 0.97 0.98 86

Curl 0.96 0.99 0.98 80

Fusarium Wilt 0.98 0.99 0.98 91

Healthy 1.00 1.00 1.00 85

EfficientNet-B4 Bacterial Blight 1.00 1.00 1.00 86

Curl 1.00 0.99 0.99 80

Fusarium Wilt 0.99 1.00 0.99 91

Healthy 1.00 1.00 1.00 85

EfficientNet-B5 Bacterial Blight 0.91 1.00 0.95 86

Curl 1.00 0.88 0.93 80

Fusarium Wilt 1.00 1.00 1.00 91

Healthy 0.99 1.00 0.99 85

EfficientNet-B6 Bacterial Blight 1.00 1.00 1.00 86

Curl 1.00 1.00 1.00 80

Fusarium Wilt 1.00 1.00 1.00 91

Healthy 0.99 1.00 1.00 85

EfficientNet-B7 Bacterial Blight 1.00 1.00 1.00 86

Curl 1.00 1.00 1.00 80

Fusarium Wilt 1.00 1.00 1.00 91

Healthy 0.99 1.00 1.00 85

MobileNet Bacterial Blight 0.99 0.99 0.99 86

Curl 0.99 0.97 0.98 80

Fusarium Wilt 1.00 0.96 0.98 91

Healthy 0.97 0.97 0.99 85

MobileNetV2 Bacterial Blight 1.00 0.98 0.99 86

Curl 0.96 1.00 0.98 80

Fusarium Wilt 1.00 0.96 0.98 91

Healthy 0.99 1.00 0.99 85

Network	Class	Precision	Recall	F1-score	Support
EfficientNet-B0	Bacterial blight	1.00	1.00	1.00	86
	Curl	1.00	1.00	1.00	80
	Fusarium wilt	1.00	1.00	1.00	91
	Healthy	1.00	1.00	1.00	85
EfficientNet-B1	Bacterial blight	0.97	1.00	0.98	86
	Curl	0.87	1.00	0.93	80
	Fusarium wilt	1.00	0.95	0.97	91
	Healthy	1.00	0.87	0.93	85
EfficientNet-B2	Bacterial Blight	1.00	1.00	1.00	86
	Curl	0.94	1.00	0.97	80
	Fusarium Wilt	1.00	0.95	0.97	91
	Healthy	1.00	1.00	1.00	85
EfficientNet-B3	Bacterial Blight	1.00	0.97	0.98	86
	Curl	0.96	0.99	0.98	80
	Fusarium Wilt	0.98	0.99	0.98	91
	Healthy	1.00	1.00	1.00	85
EfficientNet-B4	Bacterial Blight	1.00	1.00	1.00	86
	Curl	1.00	0.99	0.99	80
	Fusarium Wilt	0.99	1.00	0.99	91
	Healthy	1.00	1.00	1.00	85
EfficientNet-B5	Bacterial Blight	0.91	1.00	0.95	86
	Curl	1.00	0.88	0.93	80
	Fusarium Wilt	1.00	1.00	1.00	91
	Healthy	0.99	1.00	0.99	85
EfficientNet-B6	Bacterial Blight	1.00	1.00	1.00	86
	Curl	1.00	1.00	1.00	80
	Fusarium Wilt	1.00	1.00	1.00	91
	Healthy	0.99	1.00	1.00	85
EfficientNet-B7	Bacterial Blight	1.00	1.00	1.00	86
	Curl	1.00	1.00	1.00	80
	Fusarium Wilt	1.00	1.00	1.00	91
	Healthy	0.99	1.00	1.00	85
MobileNet	Bacterial Blight	0.99	0.99	0.99	86
	Curl	0.99	0.97	0.98	80
	Fusarium Wilt	1.00	0.96	0.98	91
	Healthy	0.97	0.97	0.99	85
MobileNetV2	Bacterial Blight	1.00	0.98	0.99	86
	Curl	0.96	1.00	0.98	80
	Fusarium Wilt	1.00	0.96	0.98	91
	Healthy	0.99	1.00	0.99	85

4.5 Network convergence

Computational efficiency and convergence time are important and can be called as direct metrics for light-weight CNNs [17]. Convergence time refers to the time taken by the number of training iterations to reach closer to the validation accuracy. The problem with deep CNN is that they tend to converge slowly because of vanishing gradient problem [46], several techniques are used by researchers to address the issue. ResNets speed up the network convergence by addressing the vanishing gradient problem but it is achieved at the cost of increased computational complexity. Experiments show that the EfficientNet converges in smaller number of epochs compared to other networks. Swish activation function [47] used in the baseline network architecture problems makes it a better network in terms of accuracy and convergence time. We have verified this by fine-tuning the learning rate, weight decay, use of the proper optimizer [48] and the batch size whose details are given in Table 2. The convergence times of different architectures are shown in Fig. 11.

4.5.1 Overall performance

A summary of the overall performance of the eight versions of the EfficientNet and two versions of the MobileNet is shown in Fig. 10. As we move from EfficientNet-B0 to EfficientNet-B7, trainable parameters keep on increasing and hence making the network more and more complex (please refer to) Fig. 10(a). However, many performance metrics e.g., validation accuracy, precision and recall do not significantly increase. Therefore, using the baseline version of the EfficientNet-B0 is the most obvious choice for this application. However, trainable parameters of the two MobileNet versions are comparable to the basic EfficientNet versions (EfficientNet-B0 and B1) but MobileNet took larger time to converge (Fig. 10(b)). Fortunately, as shown in Fig. 10(c), the accuracy of all competing models is comparable.

4.5.2 Discussion of Results

As mentioned earlier, pretrained models of the EfficientNet and MobileNet are used for all experimental purposes. For all versions of the EfficientNet and MobileNet models, 80/20 train test split is used according to the strategy depicted graphically in Fig. 5. The experimental results have shown that almost all versions of the EfficientNet model achieved excellent validation accuracy, precision and F1_score as shown in Tables 4 and 5. A comprehensive comparison of the validation accuracy for EfficientNet-B0 to EfficientNet-B7 versions, MobileNet and MobileNetV2 is given in Fig. 8 and 9. The accuracy of all other methods is close to 99%.

Fig. 8

Training vs validation accuracy and loss plots for different EfficientNet models.

Fig. 9

Training vs validation accuracy and loss plots for different MobileNet models.

Fig. 10

Trainable parameters, epochs to converge and validation accuracy for various deep learning models.

However, the objective is to achieve higher accuracy using a computationally light weight model [17]. EfficientNet-B0 achieved 99.95% in 152 seconds (8 epochs × 19 seconds per epoch). As the compound scaling increases for the higher order variants of the EfficientNet model, better accuracy can be achieved at the cost of large number of trainable parameters, which consequently means longer training times. The situation is depicted in Fig. 11.

Fig. 11

A comprehensive comparison of accuracy and convergence time for different deep learning models.

Further, trainable parameters of MobileNet are lesser than the EfficientNet-B0 model but the former takes longer to converge. One possible reason for this behavior is EfficientNet’s intelligent structure where varying kernel size is applied repeatedly. Although the MobileNetV2 pretrained model was trained the fastest by virtue of the least number of trainable parameters, the accuracy of the EfficientNet-B0 model remained the largest. The time of training for MobileNetV2 is 493 seconds which is much larger than EfficientNet-B0.

5 Conclusion and future work

Keeping in view that deep learning models have rarely been applied for cotton crop diseases, we have presented a low cost, precise and efficient solution in this paper. The proposed approach includes collecting a cotton plant leaf disease dataset consisting of 4 classes and a deep learning framework based on eight versions of the EfficientNet and two versions of the MobileNet models. Core versions of both these models were developed with the perspective of being computationally light-weight so that the trained model can run on handheld devices too. After extensive experimentation, it was concluded that our deep learning model based on modified EfficientNet-B0 converges the earliest and at the same time by being the most accurate on our augmented cotton leaf dataset. The encouraging results can be improvised henceforward to propose a light-weight deep learning model for massive plant leaf disease dataset.

In future, we want to expand our dataset for more real world images of cotton and other crops. Exploring disease severity and multiple disease on a single leaf for the cotton crop in particular is another interesting future direction.

Moreover, for the purpose of in-field evaluation, we also aim to implement our proposed algorithm on a GPU based cheap AI (Artificial Intelligence) computer such as Nvidia’s Jetson Nano Platform [49].

Footnotes

Acknowledgments

The authors would like to acknowledge the support received by the office of research, innovation and commercialization (ORIC), The Islamia University of Bahawalpur, Pakistan vide Research Grant No. 3873/ORIC/IUB/2021.

www.kaggle.com/dataset/546ad5032f88920df36e5f929f33be6bf 013f6e178efd34a163d08252e6077a6

References

Rehman

, Jingdong

, Chandio

A.A.

, Hussain

, Wagan

S.A.

and Memon

Q.U.A.

, Economic perspectives of cotton crop in Pakistan: A time series analysis (1970–2015) (Part 1), Journal of the Saudi Society of Agricultural Sciences 18(1) (2019), 49–54.

Saleem

, Nahid

, Shakir

, Ijaz

, Murtaza

, Khan

A.A.

, Mubin

and Nawaz-ul-Rehman

M.S.

, Diversity, mutation andrecombination analysis of cotton leaf curl geminiviruses, PLoSOne 11(3) (2016), 0151161.

Sarangdhar

A.A.

and Pawar

, Machine learning regression technique for cotton leaf disease detection and controlling using IoT, in: 2017 International conference of Electronics, Communication and Aerospace Technology (ICECA) 2 (2017), 449–454.

Abu-Naser

S.S.

, Kashkash

and Fayyad

, Developing an expert system for plant disease diagnosis, Journal of Artificial Intelligence 3(4) (2010), 269–276.

Rahman

M.-U.

, Khan

A.Q.

, Rahmat

, Iqbal

M.A.

and ZafarGenetics

, and genomics of cotton leaf curl disease, its viral causalagents and whitefly vector: a way forward to sustain cotton fibersecurity, Frontiers in Plant Science 8 (2017), 1157.

Briddon

R.W.

and Markham

, Cotton leaf curl virus disease, Virus Research 71(1–2) (2000), 151–159.

Farooq

, Farooq

, Mahmood

, Shakeel

, Rehman

K.A.

, Batool

, Riaz

, Shahid

M.T.H.

, Mehboob

, et al., An overview of cotton leaf curl virus disease (CLCuD) a serious threat to cotton productivity, Australian Journal of Crop Science 5(13) (2011), 1823.

Iqbal

, Sattar

M.N.

and Shafiq

, CRISPR/Cas9: a tool tocircumscribe cotton leaf curl disease, Frontiers in PlantScience 7 (2016), 475.

Rajendran

, Saravanakumar

, Raguchander

and Samiyappan

, Endophytic bacterial induction of defence enzymes against bacterial blight of cotton, Phytopathologia Mediterranea 45(3) (2006), 203–214.

10.

Kemerait

, Allen

, Lu

, Rothrock

, Faske

, Woodward

, Wheeler

, Isakeit

, Bart

, Phillips

, et al., Identification and Management of Bacterial Blight of Cotton, Cotton Incorporated Bulletin, Cotton Incorporated, Cary, NC (2017).

11.

Chen

, Bauske

, Musson

, Rodriguez-Kabana

and Kloepper

, Biological control of Fusarium wilt on cotton by use of endophytic bacteria, Biological Control 5(1) (1995), 83–91.

12.

Zhu

, Lujan

, Wedegaertner

, Nichols

, Abdelraheem

, Zhang

and Sanogo

, First report of Fusarium oxysporum f. sp. vasinfectum race 4 causing Fusarium wilt of cotton in New Mexico, USA, Plant Disease 104(2) (2020), 588–588.

13.

Ulloa

, Hutmacher

R.B.

, Davis

R.M.

, Wright

S.D.

, Percy

and Marsh

, Breeding for Fusarium wilt race 4 resistance in cotton under field and green house conditions, Journal of Cotton Science 10(2) (2006), 129.

14.

Kamilaris

and Prenafeta-Boldú

F.X.

, Deep learning in agriculture: A survey, Computers and Electronics in Agriculture 147 (2018), 70–90.

15.

Barbedo

J.G.A.

, A review on the main challenges in automatic plant disease identification based on visible range images, Biosystems Engineering 144 (2016), 52–60.

16.

Noon

S.K.

, Amjad

, Qureshi

M.A.

and Mannan

, Use of DeepLearning Techniques for Identification of Plant Leaf Stresses: AReview, Sustainable Computing: Informatics and Systems 28 (2020), 100443.

17.

Xiang

, Liang

, Sun

, Zhang

and Wang

, L-CSMS: novel lightweight network for plant disease severity recognition, Journal of Plant Diseases and Protection (2021), 1–13. doi:10.1007/s41348-020-00423-w.

18.

Howard

A.G.

, Zhu

, Chen

, Kalenichenko

, Wang

, Weyand

, Andreetto

and Adam

, Mobilenets: Efficient convolutional neural networks for mobile vision applications, arXiv preprint arXiv:1704.04861 (2017).

19.

Tan

and Le

Q.V.

, Efficientnet: Rethinking model scaling for convolutional neural networks, arXiv preprint arXiv:1905.11946 (2019).

20.

Jenifa

, Ramalakshmi

and Ramachandran

, Cotton Leaf Disease Classification using Deep Convolution Neural Network for Sustainable Cotton Production, in: 2019 IEEE International Conference on Clean Energy and Energy Efficient Electronics Circuit for Sustainable Development (INCCES), IEEE, 2019, pp. 1–3.

21.

Barbedo

J.G.A.

, Koenigkan

L.V.

, Halfeld-Vieira

B.A.

, Costa

R.V.

, Nechet

K.L.

, Godoy

C.V.

, Junior

M.L.

, Patricio

F.R.A.

, Talamini

, Chitarra

L.G.

, et al., Annotated plant pathology databases for image-based detection and recognition of diseases, IEEE Latin America Transactions 16(6) (2018), 1749–1757.

22.

Too

E.C.

, Li

, Kwao

, Njuki

, Mosomi

M.E.

and Kibet

, Deep pruned nets for efficient image-based plants disease classification, Journal of Intelligent & Fuzzy Systems 37(3) (2019), 4003–4019.

23.

Latif

, Alghazo

, Maheswar

, Vijayakumar

and Butt

, Deep learning based intelligence cognitive vision drone for automatic plant diseases identification and spraying, Journal of Intelligent & Fuzzy Systems 39(6) (2021), 8103–8114.

24.

Patrick

M.K.

, Weyori

B.A.

and Mighty

A.A.

, Capsule network with K-Means routing for plant disease recognition, Journal of Intelligent & Fuzzy Systems 40(1) (2021), 1025–1036.

25.

Prajapati

B.S.

, Dabhi

V.K.

and Prajapati

H.B.

, A survey on detection and classification of cotton leaf diseases, in: 2016 International Conference on Electrical, Electronics, and Optimization Techniques (ICEEOT), IEEE, 2016, pp. 2499–2506.

26.

Jenifa

, Ramalakshmi

and Ramachandran

, Classification of Cotton Leaf Disease Using Multi-Support Vector Machine in: 2019 IEEE International Conference on Intelligent Techniques in Control, Optimization and Signal Processing (INCOS), IEEE, 2019, pp. 1–4.

27.

Patil

S.P.

and Zambre

R.S.

, Classification of cotton leaf spot disease using support vector machine, International Journal of Engineering Research 3(4) (2014), 1511–1514.

28.

Rothe

P.R.

and Rothe

J.P.

, Intelligent pattern recognition system with application to cotton leaf disease identification, in: Innovations in Computer Science and Engineering, Springer, 2019, pp. 19–27.

29.

Zhang

, Qiao

, Meng

, Fan

and Zhang

, Identification of maize leaf diseases using improved deep convolutional neural networks, IEEE Access 6 (2018), 30370–30377.

30.

Chen

, Zhang

and Nanehkaran

, Identifying plant diseases using deep transfer learning and enhanced lightweight network, Multimedia Tools and Applications 79(41) (2020), 31497–31515.

31.

Cheng

, Yu

F.X.

, Feris

R.S.

, Kumar

, Choudhary

and Chang

S.-F.

, An exploration of parameter redundancy in deep networks with circulant projections, in: Proceedings of the IEEE International Conference on Computer Vision, 2015, pp. 2857–2865.

32.

Gong

, Liu

, Yang

and Bourdev

, Compressing deep convolutional networks using vector quantization, arXiv preprint arXiv:1412.6115 (2014).

33.

Barman

, Choudhury

R.D.

, Sahu

and Barman

G.G.

, Comparison of convolution neural networks for smartphone image based real time classification of citrus leaf disease, Computers andElectronics in Agriculture 177 (2020), 105661.

34.

Duong

L.T.

, Nguyen

P.T.

, Di Sipio

and Di Ruscio

, Automatedfruit recognition using EfficientNet and MixNet, Computers andElectronics in Agriculture 171 (2020), 105326.

35.

Atila

Ü.

, Uçar

, Akyol

and Uçar

, Plant leafdisease classification using EfficientNet deep learning model, Ecological Informatics 61 (2021), 101182.

36.

Argüeso

, Picon

, Irusta

, Medela

, San-Emeterio

M.G.

, Bereciartua

and Alvarez-Gila

, Few-Shot learning approach forplant disease classification using images taken in the field, Computers and Electronics in Agriculture 175 (2020), 105542.

37.

Kamal

, Yin

, Wu

and Wu

, Depthwise separable convolutionarchitectures for plant disease classification, Computers andElectronics in Agriculture 165 (2019), 104948.

38.

Liu

and Wang

, Early recognition of tomato gray leaf spot disease based on MobileNetv2-YOLOv3 model, Plant Methods 16(1) (2020), 1–16.

39.

Wiedemann

, Müller

K.-R.

and Samek

, Compact and computationally efficient representation of deep neural networks, IEEE Transactions on Neural Networks and Learning Systems 31(3) (2019), 772–785.

40.

Waheed

, Goyal

, Gupta

, Khanna

, Hassanien

A.E.

and Pandey

H.M.

, An optimized dense convolutional neural network model fordisease recognition and classification in corn leaf, Computersand Electronics in Agriculture 175 (2020), 105456.

41.

Chen

, Zhang

, Zeb

and Nanehkaran

Y.A.

, Identification ofrice plant diseases using lightweight attention networks, Expert Systems with Applications 169 (2021), 114514.

42.

Iandola

F.N.

, Han

, Moskewicz

M.W.

, Ashraf

, Dally

W.J.

and Keutzer

, SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and < 0.5 MB model size, arXiv preprint arXiv:1602.07360 (2016).

43.

Weiss

, Khoshgoftaar

T.M.

and Wang

, A survey of transferlearning, Journal of Big data 3(1) (2016), 9.

44.

Deng

, Dong

, Socher

, Li

L.-J.

, Li

and Fei-Fei

, Imagenet: A large-scale hierarchical image database, in: 2009 IEEE conference on computer vision and pattern recognition, IEEE, 2009, pp. 248–255.

45.

Zhang

, Bengio

, Hardt

, Recht

and Vinyals

, Understanding deep learning requires rethinking generalization, arXiv preprint arXiv:1611.03530 (2016).

46.

Khan

, Sohail

, Zahoora

and Qureshi

A.S.

, A survey of the recent architectures of deep convolutional neural networks, Artificial Intelligence Review 53(8) (2020), 5455–5516.

47.

Ramachandran

, Zoph

and Le

Q.V.

, Swish: a self-gated activation function, arXiv preprint arXiv:1710.05941 (2017), 1.

48.

Noon

S.K.

, Amjad

, Qureshi

M.A.

and Mannan

, Overfitting Mitigation Analysis in Deep Learning Models for Plant Leaf Disease Recognition, in: 2020 IEEE 23rd International Multitopic Conference (INMIC), IEEE, 2020, pp. 1–5.

49.

Cass

, Nvidia makes it easy to embed AI: The Jetson nano packs a lot of machine-learning power into DIY projects-[Hands on], IEEE Spectrum 57(7) (2020), 14–16.

Computationally light deep learning framework to recognize cotton leaf diseases

Abstract

Keywords

1 Introduction

3 Proposed computationally Light-weight deep learning model

4.1 Dataset

Table 1 Details of the cotton leaf dataset used in the proposed work Class No. of Images Curl 418 Bacterial Blight 448 Fusarium Wilt 419 Healthy 426 Total 1,711

Table 2 Parameters of the proposed CNN based on the EfficientNet and the MobileNet models Parameters EfficientNet MobileNet Batch size 16 16 Dropout 0.25 0.30 Learning rate 0.001 0.001 Decay 5 × 10-5 4 × 10-5 Batch Norm. No yes Optimizer Adam SGD

4.4.1 Accuracy and Recall

4.5.1 Overall performance

4.5.2 Discussion of Results

Footnotes

Acknowledgments

References

Table 1
Details of the cotton leaf dataset used in the proposed work

Class No. of Images

Curl 418

Bacterial Blight 448

Fusarium Wilt 419

Healthy 426

Total 1,711

Table 2
Parameters of the proposed CNN based on the EfficientNet and the MobileNet models

Parameters EfficientNet MobileNet

Batch size 16 16

Dropout 0.25 0.30

Learning rate 0.001 0.001

Decay 5 × 10^-5 4 × 10^-5

Batch Norm. No yes

Optimizer Adam SGD