Evaluation of CNN Models in Identifying Plant Diseases on a Mobile Device

Plant diseases are one of the utmost threats to crop production and expansion. Plant diseases can hamper the target of agricultural production every year where about 85% are caused by fungi or fungus-like organisms. The repeated failure to meet production targets can cause food insecurity for a country and even the world.

Using pesticides and insecticides could give a short solution but it is not helpful for long-term. Therefore, early detection and identification for plant diseases is one of best options in minimizing the damage level of plants. This option can be used because symptoms of plant diseases can mostly be identified on the leaves, such as chlorosis (the yellowing or lightening of the green parts of plants) and curl (rolling, cupping, or puckering of leaves).

For the farmers, who lack knowledge of plant diseases, will call the experts to help them identify plant diseases. In most cases, the experts need to use a microscope for carrying out in-depth analysis or need to convert the sample into electromagnetic spectrum. However, this method tends to be time-consuming and expensive, even while the experts are far away from the farmers or laboratory assistance required to diagnose the diseases.

The new technology appeared over the past few years like computer vision can be an alternative way to solve the limitations of traditional methods. The recent applications of computer vision through deep learning have made enormous advances in various studies, such as object segmentation, object recognition, image segmentation, and image classification. Furthermore, deep learning is not limited to being implemented on Personal Computer (PC), the others can be synchronized with mobile devices. Mobile devices have significantly increased usage during covid-19 pandemic, where people unwillingly work and study from home, or need to scan entries ticket via QR code if want to enter public places even apartments. These uncommon activities during the pandemic contribute to changing the perspective of mobile devices usage as an intermediary medium. This opportunity shown from mobile devices can be one of the potentially good mediums to reach farmers in rural areas.

However, some rural and plantation areas in Indonesia still not have enough internet connections support. Therefore, building a plant diseases diagnosis system with limited internet connectivity through a mobile application can be very helpful for farmers.

The system is expected to be able to identify diseases from leaf images and to provide accurate and quick labeling of a plant disease. The mechanism of identification refers to the similarity between the diseases on pictures of plants with actual plant disease by using deep learning algorithms.

Deep learning algorithms work using several layers of neural network algorithms to process data and create patterns in decision making automatically. The subset of deep learning algorithms is Convolution Neural Network (CNN) which is currently one of machine learning popular methods for solving image classification problems. There have been a lot of studies that use CNN in classifying plant diseases, as have been done by Jadhav et al. [1] that achieve accuracy rate of 96.25% by implementing AlexNet with some modifications in hyperparameters using 650 disease and 550 healthy soybean images. Upadhyay and Kumar [2] reached identification accuracy of 99.7%, using simple layers of CNN including batch normalization layer, one fully connected, and softmax layer with Rice Leaf Image dataset. Then, Sembiring et al. [3] classify ten tomato leaf diseases from PlantVillage Dataset and achieve the accuracy rate of 98.28%. Two versions of PlantVillage dataset (Augmented and Original) were evaluated by Atila et al. [4] by using several different CNN models, the highest accuracy of each version dataset, EfficientNet-B4 at 99.97% and EfficientNet-B5 at 99.91% respectively. Then, Liu et al. [5] achieved identification accuracy of 97.22% from DICNN model using 107,366 images of augmented 7,669 grape leaf images which are generated via image enhancement technique.

Based on previous research, there are only two studies that focus on the application of deep learning techniques to be used on mobile devices, specifically the research of plant disease classification. In 2019, Syamsuri and Kusuma [6] evaluate several pre-trained deep learning Convolutional Neural Network (CNN) on mobile devices, MobileNet reached accuracy of 92.83%, followed by MNasNet at 94.87%, and Inception V3 at 95.79%. Then, Bimorogo and Kusuma [7] expands the previous research by adding two classes of coffee plants and four classes of rice plants by using MobileNet V2 that achieve accuracy of 96.55%, then NasNet Mobile at 97,31%, DenseNet 121 layer at 96.21%, and Inception V3 at 98,45%. Both studies can achieve high accuracy and can be said to be almost perfect. However, the use of the PlantVillage dataset as the main dataset on both studies has drawbacks where [8] shows a decrease in accuracy by ten to twenty percent when tested with images of different lighting conditions. From the background above, there is a motivation to build a system that has the ability to recognize many variants of diseases with good accuracy performance and resource requirements in helping farmers in rural areas who are still difficult to reach by a fast internet network.

In this research, the authors use CNN model that has fewer parameters and has been demonstrated to be highly accurate on a certain disease dataset, namely Mason model [9], which compared to another less parameters CNN models that have been proven to be highly accurate on ImageNet dataset such as MobileNetV3 [10], ShuffleNetV2 [11], and EfficientNetB0 [12]. The motivation of using less parameters models is to reduce computation cost and require less resources.

Furthermore, this research extends more plant disease classes which consisted of 104,282 images from 21 kind of plants, in order to enrich the variation of dataset and build robust identification system with more variation of plant diseases.

The evaluation design in this research is carried out in two parts, namely on the workstation which is to find the best model performance based on its loss and then it is tested on a mobile device to evaluate the required resource requirements, namely file size, memory, CPU, detection speed, and power consumption. The evaluation results show that the Mason model shows better performance results compared to other models even though still needs some minor improvement in the future.

This paper is organized as follows. Section 2 presents the related work to the detection of plant diseases. Section 3 shows the experimental design method by mentioning merged datasets detail and illustrating workflow experimental on workstation and a mobile device. Section 4 discusses the experimental results on workstation and mobile devices which are displayed in table form and has an explanation. Section 5 shows the conclusion of the experimental results that have been carried out.

3.1 Dataset

Several sources of datasets were used in this research, these are The PlantVillage dataset that collected by Hughes and Salathé [13], with extended cassava leaf disease that used in competition held by Makerere University AI Lab [14], Plant Pathology dataset that was the dataset of Plant Pathology 2020 Challenge [15], a robusta coffee leaf images called RoCole dataset [16], then an apple scab dataset collected by Institute of Hulticulture [17], bean disease dataset from public domain, PlantDoc dataset [18], corn leaf disease dataset [19], cotton leaf disease dataset [20], DiaMOS plant dataset which contains four pear leaf disease [21], some disease species from Digipathos dataset [22] that have more than 20 images, rice leaf diseases dataset that gathered from a farming community [23], citrus leaf diseases dataset [24], cassava leaf diseases dataset [25], bean diseases dataset [26], and some manual searching for two classes, rice gray leaf spot and rice healthy. Total images were 104,282 images consisting of 21 plant species in 79 different classes with 19 healthy plants and 60 combinations of diseased plants. Next, the authors removed some background images from the datasets which have like the other and renamed properly all the images file after the dataset merged. The detail of the dataset is mentioned in Table 1.

Then, the dataset will be separated into three parts datasets, training, validation, and testing, using stratified random sampling with composition 60% (62,568 images), 20% (20,857 images), and 20% (20,857 images) respectively. Due to the manual image picking process in mobile, the testing data would be 158 images for testing on mobile devices and uses all of testing data for evaluating TFlite models’ accuracy on PC.

Table 1. Merged Dataset

3.2 CNN models

This research implemented transfer learning method to the CNN models. Transfer learning works by gained the weight of pre-trained CNN models before on a large-scale and general dataset, namely ImageNet. So, the CNN models can learn basic knowledge from existing models instead of starting from scratch each time. However, the authors cannot find two saved weights of transfer learning on ImageNet for ShuffleNetV2 and Mason model. The alternative, both models will be trained first using merged dataset with 20 epochs without early stopping using learning rate SGD 0.001 and Adam 0.001 respectively.

This work also fine-tuned the last ten layers of each model. By fine tuning, the information from transfer learning is not lost at all when the model to be retrained. For the final layers, each model used two fully connected layers (FCs) with relu activation function that has nodes same as the number of classes including a softmax layer.

3.2.1 Mason model

Mason model is popularized by Masud et al. [9] which has less parameters around 200 thousand. This model is supported by four convolution layers, a Global Average Pooling (GAP), two Fully Connected (FC), and a softmax output layer. Each convolution block has 3 x 3 rectified linear units (ReLUs), a connected convolution block 1 x 1 with ReLu and a max pooling block.

Zero padding is implemented before reaching convolution block for keeping the size of each convolution block. Mini-batch normalization is also used for accelerating the training. Then, in the middle of convolution block and max pooling, there is convolution block connector.

This model was able to obtain prediction accuracy of 97.9% in pulmonary nodule detection using CT Scan images of Lung Image Database Consortium (LIDC) dataset on mobile devices. 

3.2.2 MobileNetV3

MobileNetV3 is a next generation model from its predecessor MobileNetV2 which was made to enhance computation efficiency, accuracy, and latency reduction. This model was made by combination of MobileNetV2 and MNasNet as building blocks.

In the new architecture proposed, there includes non-linearity, h-swish, and new version of swish non-linearity for enlarging friendly computation and quantization. In addition, the new design of MobileNetV3 switched the final layer in MobileNetV2 design to after final average pooling position, so the final layer will only count 1 x 1 spatial resolution.

Overall, MobileNetV3 has two versions that consists of MobileNetV3-Large and MobileNetV3-Small. This work used the large one which still has less parameters and good accuracy at [10].

3.2.3 ShuffleNetV2

ShuffleNet V2 is the second edition of ShuffleNet which proposed channel split concept and switching channel shuffle to bottom block [11]. This concept is designed to reduce the cost of pointwise group convolutions and bottleneck structure.

Overall, the performance of the model from its authors shown 58% faster from MobileNet V2 and 63% faster from ShuffleNet V1.

3.2.4 EfficientNet B0

EfficientNet B0 is a sub-model of EfficientNet group which was introduced by Tan and Le [12]. The baseline of this model was built from a combination of MobileNet and ResNet. The compound scaling approach was proposed to scale up calculation stability of dimension, width, and resolution depth. By this approach, the layers can be stretched up or down automatically in getting receptive field and fine-grained based on the input images. 

3.3 Evaluation

The research was divided into two versions of implementation, those consist of experiments on workstation and experiments on a mobile device. In the experiments of workstation, mainly, the authors implement pre-trained CNN models with two hyperparameter (Adam and SGD) and learning rate (0.01 and 0.001) to find the best accuracy on PC. Then, in the experiments on a mobile device, primarily, the authors chose the four best models to be tested on a mobile device.

3.3.1 Experiments on a workstation

The experiments process supported by TensorFlow V2, TensorFlow-lite deep learning model, Python 3.7, and Visual Studio Code. The workflow in the experiments on workstation is illustrated by a flowchart that is shown in Figure 1.

11.png

Figure 1. Experiments process chart on workstation

The CNN model that has been trained was chosen based on the highest validation accuracy rate. Then, each model was saved in disk using “model.save()” function from TensorFlow. The experiments were carried out on a workstation which has the specifications that are shown in Table 2.

Table 2. The workstation specification

3.3.2 Experiments on a mobile device

In the beginning, the best trained models that saved by “model.save()” function was recalled to be converted into TensorFlow-lite (TFlite) models. The standard of converted TFlite models have default float32 version. This version needed to be optimized for reducing the storage size, less memory usage, and smaller download size without decreasing the accuracy. This work used float16 quantization because it provides minimal impact on accuracy and up to 50% model size reduction [26].

Then, the optimized TFlite models were carried out to predict the testing data on a workstation. Furthermore, the mobile application was built as a medium to read the TFlite models on a mobile device using kotlin language and android studio.

The experiments process on a mobile device is illustrated by a flowchart which is depicted in Figure 2.

2.png

Figure 2. Experiments chart on a mobile device

The application was used to evaluate model CNN on a mobile device illustrated in Figure 3.

There were ten repetitions to be tested on a mobile device for evaluating the resource requirements which consist of CPU, memory, and power consumption. The mobile battery was fully charged, on airplane mode, and restarted on every repetition to get the baseline and to ensure no other application are running. The numeric output data from the AccuBattery application, application to count resource requirements, during the testing on a mobile device was noted into excel. The experiments were carried out on a mobile device which has the specifications that are shown in Table 3.

3.png

Figure 3. Mobile application prototype

Table 3. The mobile device specification