Analysis of Plant disease in Power Plant Areas Using Deep Learning Techniques
L.Subash Dr G.Arulselvi Dr K.Kavitha
Research scholar, Associate Professor Associate Professor Dept of computer and Dept of computer and Dept of computer and Information science Engineering Engineering Annamalai university Annamalai university Annamalai university l.subashmenon @gmail.com [email protected] [email protected] Abstract:
Nowadays the plant disease control is a fundamental practice in agriculture food production. Many plant disease- preventing techniques have assumed critical parts in decreasing the recurrence and force of sicknesses. The exhaust gas from power plant industries can change the sensitivity of plants prompts the incidence of new disease caused due to the rise of atmospheric temperature and CO2 fixation. The main goal of the task is to actualize new deep learning techniques by combining three different architecture VGG 16, Google Net, and GAN offers to foresee the groundnut plant disease incidence and effectively conveyed on agrifield nearer to power plant zone and non-power plant zone. Furthermore, we inspected an assortment of deep learning applications with plant disease imaging, preprocessing, segmentation, and classification division that are firmly interlaced. The best-prepared model accuracy of groundnut disease of 6 classes were 98% for early leaf spot (BLS), 96% for late leaf spot, 95% for rust disease ), 98% for stem root disease, 96% for Alternaria,93 % for anthracnose and 96 % for healthy leaf .the results shows that new deep learning approach for plant disease investigation of field nearer to power plant zone and non-power plant zone offers a quick, moderate, and effectively deployable procedure for computerized plant disease recognition.
Keywords: Groundnut Disease Detection, Deep Learning, VGG-16, GoogleNet, Generative Adversarial Network (GAN)
1.Introduction:
Groundnut is an important oilseed crop in India contributing significant jobs in exporting oil production. Healthy organic sources are required for human beings to avoid causing disease. The increasing incidence of plant leaf disease can severely affect the yield of groundnut production, and it made a major financial misfortune to the groundnut growers (Singh 2018). The exhaust parameter from power plant industries such as environmental temperature and CO2 naturally affects plant growth and is the critical driver for warming our planet at a quickened rate. The planet earth is encountering an environmental change and CO2 concentration is a significant GHG, which expanded by almost 30% and temperature by 0.3-0.6°C [4].
Fig 1.1 Groundnut Plant Classes
The rise in temperature, CO2, and water stress increases the incidence of diseases such as fungi, viral, rust and leaf spot disease (Woodward et al. 2006). This has led to a reduction in yield and financial loss to the farmer. The occurrence of diseases, like an early leaf, late leaf spot, pepper spot, rust sickness, stems root infection Alternaria, and anthracnose, affect the growth and yield of groundnut plants. In India 40–60% of yield reduction in groundnut was reported due to various diseases, but it may up to 93% on the same occasions (Bharate and Shirdhonkar 2017.) In this study, we are planning to classify the groundnut leaf disease for two distinctive farming fields nearer to the power plant zone and the non-power plant zone. the groundnut disease classification at the early stage of crops had a significant benefit to the farmers to minimize the yield loss (Ojala et al. 2002).
The developing technology, specifically the accessibility of multimodality information from different sensors including the Internet of things and sensor organizations, has grown quickly [11]. Thus, a novel groundnut disease classification proof model dependent on new deep learning technology is intended to tackle the above issues. The proposed work contains leaf image preprocessing, segmentation, and classification with integrated learning architecture of VGG 16, Googlenet, and GAN, all through the entire cycle. We proposed an incredible and powerful technique of DCNN to recognize and order groundnut plant illnesses.
The major contribution of the research is to predict the groundnut leaf disease analysis in the field located near power plant areas and non-power plant areas. The groundnut plant images were collected at Neyveli and Kullanchavadi (Cuddalore district in Tamilnadu) are utilized for samples from power plant area and non-power plant area respectively. To implement the preprocessing tool like resize, colour enhancement, geo transformation, median filter, used to enhance the digital image processing. To find the disease classification of groundnut leaf using three different deep learning architecture like VGG 16, GoogleNet, GAN, and combination of
the three architecture were used to evaluate the groundnut disease level of an agriculture field in power plant zone and non-power plant zone. the study was undertaken to find the best performance of deep learning architecture by comparing the accuracy of disease classification and using performance metrics.
2.Related works:
A quiet lot of the works recommended the technology of computer-based disease identification in agriculture plants. (Vaishnnaveet al, 2020) proposed an effective strategy for the deep convolutional neural organization (DCNN) with the six consolidated layer to extract the significant features with no human oversight. The correlation investigation system conveys 95.28% accuracy esteem. (GeetharamaniG et al, 2019). Presented a strategy that accomplishes a normal accuracy of 96.46% in the arrangement of the testing set plant leaf pictures, and somewhere in the range of 92% and 100% for the individual class. This model can successfully order the 38 unique classes of Healthy and Non-Healthy plants utilizing leaf pictures and the transformation methodology provide the learnable parameter from 49,598 to 55,636.
The machine learning algorithm K-closest neighbor (K-NN) was actualized by El Houby (2018) for the identification of plant disease analysis. The strategy is utilized to extricate the features, and it groups the information dependent on their measures. In any case, it doesn't anticipate the infections effectively. GodliverOwomugisha and Ernest Mwebaze(2016) present an application of machine learning to the horticulture and farming field, resolve a specific issue of analysis of harvest disease based on plant pictures taken with a cell phone. This works by the farmer transferring a picture of a plant in his nursery and getting a disease level from a on-field worker.
Apple Tree Leaf Diseases (ATLDs) can handle by utilizing the learning procedure of altered DCNN outline work and it accomplishes a general precision of 98.82% in distinguishing the ATLDs higher than VGG-16 and DenseNet-201.
Ramcharan et al (2017) stated that cassava leaf pictures were taken with a 20.2-megapixel computerized camera in exploratory for three AI techniques were utilized, and results showed that SVM has the highest accuracies for predicting the six diseases concerning to cassava disease.Yanget al. (2003) proposed an incredible and adaptable technique for AI instruments to make the blend of well appropriate framework information. Li et al. (2010) presented the amalgamation of back propagation with a Neural Network (NN)- based idea for the recognition of cotton stress control. In the early stage, the noticeable territory of the plant like leaves and stem sicknesses is distinguished by using Smartphone applications.
A huge plant image dataset is retrained in transfer learning, where a model that has been prepared for new classes, offers an easy way to training DCNN learning models due to bring down computational prerequisites. This would have a particular benefit for field settings. Here we explored the potential for new integration DCNN model of VGG 16, Google net, GAN to recognize a level of Groundnut plant disease utilizing an on-field dataset of 2500 images from
nearer to power plant zone and 2500 from non-power plant zone comprising 6 Groundnut diseases caused by influencing boundary exhausted from power plant sector.
3.Data acquisition:
In this project, we have taken the groundnut leaf dataset from two kinds of villages that comprise various groundnut diseases. But we just spotlight on the dataset of 5000 groundnut leaf images of two different datasets contains category like healthy leaf early leaf spot, late leaf spot, rust infection, stem root sickness, Alternaria, and Anthracnose.. The datasets contain 2500 leaf images taken from the power plant zone, Neyveli, and 2500 leaf images taken from Kullanchavadi which is a non-power plant zone. The non power-plant zone (Kullanchavadi) is located 35 km away from the power plant zone. The raw data were recorded by utilizing the Canon 200D camera. To differentiate the plant infection examination for groundnut leaves, we have gathered the examples in 1 hectare from both Neyveli and kullanchavadi. All this information is gathered with the assistance of an agriculture expert. The list of disease-affected dataset images is portrayed in Fig. 12. A sum of 3200 pictures is utilized for the training set and 1800 pictures are utilized for the testing purpose. In this paper, each class uses 300 examples for every classification. A portion of the framework of groundnut disease classification is shown in Fig 2.
Fig 2.The Proposed Framework of Groundnut Plant Disease Classification
3.1 Preprocessing and Segmentation
The significant goal of this paper is to give a groundnut disease classification algorithm to evaluate the two different plant town consists of six classes by implementing image processing.
The steps include for groundnut leaf disease are preprocessing, segmentation and classification.
Initially, the median filter is applied to decrease the issue identified with the leaf surface.. The median channel is used to the sliding a window over the leaf image . The colour enhancement is applied to improve the input data for improving influenced a piece of image dependent on the intensity of the pixel. To improve the image, contrast is upgraded utilizing histogram
equalization. The transformation is further applied to build the number of input images in the dataset and reduce the overfitting ratio by adding mutilated images to the DCNN trained model data. The process of flipping, Gamma adjustment, revolution, and Scaling changes are utilized to make the augmented for the training module.
3.2 Deep learning background:
CNN for the most part uses deep learning models in dealing with images identified with tasks like distinguishing proof, location, characterization, etc. These frameworks regularly consist of convolution layers, pooling layers, and fully connected layers. (Kernals), pooling, fully connected (FC). The final layer Softmax can organize a thing with probabilistic characteristics some region in the assortment digitally. Convolution is the main layer to remove features include data from an information input test image. Convolution saves the relationship between pixel esteems by learning picture include features using little squares of data.
The pooling layers fragment would decrease the number of limits when the image tests are unnecessarily enormous. Completely Connected layers in neural organizations are those layers where every one of the contributions from one layer is associated with each actuation unit of the following layer. In most well-known AI models, the last couple of layers are full associated layers that arrange the information removed by past layers to frame the last yield. It is the second most tedious layer second to the Convolution Layer. Various models have been made since (Geetharamani et al 2019) the CNN-based plan was proposed to mastermind contamination in the PlantVillage dataset, and it performed better contrasted with the prominent Deep learning models of AlexNet, VGG-16, Inception-v3, and ResNet.Every model had moving hyper cutoff points or had utilized new strategies like Drop out, Image development, regularization, normalization, Batch standardization. In this assessment, the plant disease portrayal model is using the combined DCNN of VGG 19, GoogleNet, and GAN.
A.VGG 19
The proposed CNN that carefully utilized 3×3 channels with the stride of 1, alongside 2×2 max- pooling layers with stride of 2 is called the VGG-19 model. Contrasted with AlexNet, the VGG- 19 (16 Conv., 3 completely associated) is a more DCNN with more layers of ImageNet.
Additionally, it has challenges from which one was ImageNet Large-Scale Visual Recognition Challenge (ILSVRC). Input size of (224 * 224) RGB picture was given as a contribution to this organization which implies that the grid was of shape (224*224*3).
The procedure preprocessing that was done is that they deducted the mean RGB esteem from every pixel, processed over the entire preparing set. Utilized portions of (3 * 3) size with a stride size of 1 pixel, this empowered them to cover the entire idea of the image.. The max-pooling was maintained over a 2 * 2 pixel with the stride of 2. This was sequenced by the ReLu layer to acquaint non-linearity with cause the model to order better and to improve computational time.
The two fully connected layers with a size of 4096 and 1000 channel softmax layer have been implemented with VGG 19 trained model is shown in Fig 3.
Fig 3.VGG 19 Architecture B.GoogleNet
In image portrayal contention, GoogLeNet is the victor of the ILSVRC 2014, which has huge improvement over LeNet and AlexNet learning models and has a respectably lower learning rate differentiated than the VGGNet. GoogLeNet is furthermore called as Inception v1 model. It contains 1×1 convolution at the focal point of the framework. The global pooling likewise uses after the framework rather than utilizing related FC layers. Another procedure inception module is utilized in the GoogleNet model. The inception module has different sizes/kinds of convolutions for comparable data and to stack every one of the output. In GoogLeNet, 1×1 convolution is used as an estimation decline module to decrease the computation. As of now, FC layers are utilized close to the furthest limit of the framework.
The GoogleNet comprises various inception modules related together to go further. There are some transient softmax branches at the center. On the off chance that a structure manages different huge layers, it may go facing overfitting. To deal with this issue, a new planned GoogleNet convolution model is proposed and shown in fig 4. With this thought, the structure ends up being more extensive rather than more critical. Since neural structures are unstimulated and over the top to set up, an extra (1 × 1) convolution is used before the (3 × 3) and (5 × 5) convolutions to diminish the parts of the system and perform a fast evaluation.
Fig 4.GoogleNet Architecture C.GAN Architecture:
In this part, we will introduce CapGAN-based plant village image and classification .as per fig 5.
the generator of our CapGAN consolidates four deconvolution layers and 3 ReLu layers. It takes a uniform 100-dimensional vector as the info, and afterward the vector is up-examined to a 4 × 4
× 1024 dimensional vector by using the (FC) layers of the generator. Beginning from the subsequent layer, a progressive arrangement of four deconvolutions (DCONV) are utilized to change over this general portrayal into the image size of 64 × 64.
Specifically, each convolutional layer in the generator utilizes the BN to standardize the output of the component layer. Such preparation is valuable to accelerate the organization grouping, improve the network dependability and speed of the learning rate [27]. Then, The ReLU actuation work is implemented taking all things together layers except for that the Tanh work is utilized in the last layer. The explanation behind utilizing the Tanh work is that the last layer needs to extract an output image, and the pixel estimation of the input data is in the scope of [0, 255].
Fig 5.CapGAN Architecture D.Performance metrics:
To evaluate the deep learnings models like VGG 19, GoogleNet, CapGAN, and combined DCNN (combined VGG19, GoogleNet, and CapGAN ) we have used metrics such as accuracy sensitivity, specificity, precision, and F1 score, the formulation of these measures are mentioned as follows :
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (𝐴 + 𝐷)
(𝐴 + 𝐵 + 𝐶 + 𝐷) (1)
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = (𝐴)
(𝐴 + 𝐵) (2)
𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = (𝐷)
(𝐵 + 𝐷) (3)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (𝐵)
(𝐵 + 𝐷) (4)
𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =2 ∗ (𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 ∗ 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)
((𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 + 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛) (5)
From the above equations, the numbers of positive and negative true functions are denoted by using the factor of A and D. Similarly, the function of Band C is the representation of false positive and negative, respectively. The classified disease images with their performance measures are tabulated in Table 1. Thus, the overall accuracy measurement value of DCNN is 99.88%.
4.Experimental Results and discussion
The proposed research work utilized the Deep Learning Models like VGG 19, GoogleNet, and GAN pre-trained model for village plant leaf sicknesses recognizable 7 classes like early leaf spot, late leaf spot, rust disease, stem root disease, Alternaria, Anthracnose, and good leaf in power plant and non-power plant zone plant village. We have used 2500 samples used to analyze plant disease classification for both areas (PPZ and NPPZ) by using Matlab 2020 b. This paper utilized random translation, cropping for information upgrades during the training process, to reduce the possibility of overfitting. The initial learning rate for leaf illness identification is 0.001.
Table 1.Parameter Settings of DCNN Models
Parameter
VGG 19 GoogleNet GAN Proposed
DCNN
Input size 224x224x3 224x224x3 128x128 224 x224
Batch size 128 32 64 256
Optimizer SGD SGD SGD SGD
Momentum 0.9 0.9 0.9 0.9
Weight decay 0.1 0.1 0.1 0.1
Leaning rate 0.001 0.001 0.01 0.001
Epoch 50 50 50 50
The proposed model improves with the assistance of SGD analyzer with the loss of cross- entropy. This examination work runs the investigation up to 50 epochs to accomplish the best outcome model. The arranged model gives in Table 1. The model trains with the planned hyper- parameter features such as input size, learning rate, optimizers, epochs, batch size, weight decay, and momentum.
In this paper, the usage of Deep Convolutional Neural Networks has detail engineering to identify sickness on groundnut leaf image. The proposed methodology takes a stab at 7 classes of two different plant zone such as good leaf early leaf spot, late leaf spot, rust infection, stem root sickness, Alternaria, and Anthracnose. According to fig 6, .most of plant disease occurred in the power plant zone from the dataset of 2500 samples on both zone and the same fig shows that late leaf spot disease mostly affected in Neyveli and kullanchavadi with the ratio of 60:40. The results of the confusion matrix table from fig 7 a, 7b, and 7c clears that the GoogleNet model performs better than VGG 19 and GANmodel. Its shows that overall accuracies for seven classes are 89.6 %, 88.1 %, and 53.4 % concerning to GoogleNet, VGG 19, and GAN separately.
Fig 6.groundnut plant disease affected in PPZ and NPPZ
The groundnut class rust disease and stem root disease were classified with high accuracies of 100 % for all DCNN models.
0 100 200 300 400
No of Samples
Plant leaf classes
PPZ (Neyveli ) NPPZ(Kullanchavadi)
Proficiently, the pooling and the convolutional layer are the primary layers in the convolutional neural organization. The proposed model of DCNN contains three deep learning modelsGoogleNet, VGG 19, and GAN, and the ideal number of joined models are found by utilizing a network search measure. Eventually, the whole layer execution of DCNN with the three networks delivers a higher accuracy rate of 92.5 % for seven classes. Subsequently, the proposed DCNN performance evaluation results for power plant zone and non-power plant zone is mentioned in Table 2. The performance metrics such as accuracy, specificity, sensitivity, precision and f1 score has been tabulated for proposed model in Table 3. The comparative results of DCNN models are clear that proposed methodology has the overall accuracy of 96.5%
tabulated in table 4.
Table 2.Performance Results of Proposed DCNN for both PPZ and NPPZ
Groundnut leaf class
PPZ (Neyveli) images NPPZ(kullanchavadi ) images
Classification Accuracy by
proposed method (%) Dataset Augmented
data set Dataset Augmented data set
Alternaria 67 260 34 76 76.9
Anthracnose 57 220 26 68 100
Early Leaf Spot 123 564 61 466 83.3
Good Leaf 400 1200 400 1250 100
Late Leaf Spot 128 587 54 467 100
Rust Disease 97 300 37 76 100
Stem Root Disease 22 100 9 31 100
Fig 7a.Confusion Matrix for GoogleNet Fig 7b.Confusion Matrix for VGG 19
Fig 7c. Confusion Matrix for CapGAN Fig 7d. Confusion Matrix for Proposed DCNN Table 3.Performance metrics results of Proposed DCNN
Groundnut leaf class
Proposed Combined DCNN ACC
(%) SPE (%) SEN (%) PRE (%) F1 (%)
Alternaria 84.5 88.6 84.5 88.5 95.6
Anthracnose 86.5 83.52 86.5 90.4 92.5
Early Leaf Spot 95.5 92.5 92.5 92.5 95.4
Good Leaf 100 99.5 95.7 95.6 95.6
Late Leaf Spot 95.5 99.6 95.3 90 95.7
Rust Disease 100 100 100 95.6 100
Stem Root Disease 100 100 100 100 100
Table 4.Comparative Results of DCNN models Groundnut leaf
class
No of Images
VGG 19 (%)
Googlenet (%)
CapGAN (%)
Combined DCNN (%)
Alternaria 150 90 88.5 60 96.5
Anthracnose 76 90.5 85 75 95.5
Early Leaf Spot 289 76.35 85.5 50 90.5
Good Leaf 1500 85.5 100 80.5 100
Late Leaf Spot 350 80 80 45.5 90
Rust Disease 240 90 100 60 100
Stem Root Disease 20 100 100 100 100
Fig 8 . Accuracy performance of DCNN models for seven groundnut classes
5.Conclusion
In this paper, we proposed an efficient deep convolutional neural network (DCNN) using VGG 19, GoogleNet, GAN for the detection and classification of groundnut diseases in powerplant zone and non-powerplant Zone. The most groundnut plant disease affected leaf is predicted in powerplant zone when compared to non-powerplant zone with the ratio of 60:40 The deep learning process is intensely used to detect the leaf disease and the DCNN classification is utilized to categorize the diseases of seven classes like good leaf early leaf spot, late leaf spot, rust infection, stem root sickness, Alternaria, and Anthracnose. When compared to DCNNmodels, the proposed DCNN (combined pre-trained model of VGG 19, GoogleNet, GAN) has shown an accuracy rate of 92.8%. Moreover, the consistency and reliability of the proposed model are measured in terms of the confusion matrix table, accuracy, sensitivity, specificity Precisions, and F1 Scores. In future we can extend this model for self diagnosis for other parts of the plants, such as flowers, fruits, and stems. Moreover, we plan to conduct a deeper investigation of the training process without labeled images.
6.Reference
Singh KK (2018) An artificial intelligence and cloud-based collaborative platform for plant disease identification, tracking, and forecasting for farmers. In: Emerging markets, pp 49–56 Bergot M, Cloppet E, Perarnaud V, Déqué M, Marcais B, et al. (2004) Simulation of potential range expansion of oak disease caused by Phytophthoracinnamomi under climate change. Global Change Biology 10(9): 1539-1552.
Woodward JE, Brenneman TB, Kemerait RC Jr, Culbreath AK, Clark JR (2006) First report of Sclerotinia blight caused by Sclerotiniasclerotiorum on peanut in Georgia. Plant Dis 90(1):111
0 50 100
Performance accuracy in (%)
Plant leaf classses
VGG 19 GoogleNet GAN Proposed
Ojala T, Maenpaa T, Pietikainen M, Viertola J, Kyllonen J, Huovinen S (2002) Outex-new framework for empirical evaluation of texture analysis algorithm. Object recognition supported by user interaction for service robots 1:701–706
Bharate AA, Shirdhonkar MS (2017) A review on plant disease detection using image processing. In: International conference on intelligent sustainable systems, pp 103–109.
Y. Guo, X. Hu, Y. Zou, et al., "Maximizing E-tailers’sales volume through the shipping-fee discount and product recommendation system,” Discrete Dynamics in Nature and Society, vol.
2020, pp. 1–14, 2020
M. P. Vaishnnave• K. Suganya Devi • P. Ganeshkumar (2020) Automatic method for classification of groundnut diseases using deep convolutional neural network In : springer Soft Comput 24, 16347–16360
GeetharamaniG.ArunPandianJ (2019 ).Identification of plant leaf diseases using a nine-layer deep convolutional neural networking: Elsevier Computers & Electrical Engineering Volume 76, June 2019, Pages 323-338
El Houby EM (2018) A survey on applying machine learning techniques for management of diseases. J Appl Biomed 16(3):165–174
GodliverOwomugisha and Ernest Mwebaze (2016) Machine Learning for Plant Disease Incidence and Severity Measurements from Leaf Images 2016
Liu, B.; Zhang, Y.; He, D.; Li, Y.; Liu, B.; Zhang, Y.; Li, Y(2017 ), Identification of Apple Leaf Diseases Based on Deep Convolutional Neural Networks(2017 ) symmetry mdpi 10,11
Ramcharan, A.; aranowski, K.; McCloskey (2017) P Deep Learning for Image-Based Cassava Disease Detection Plant Sci. 8:1852.(Frontiers in plant science).
