Information Processing in Agriculture.pdf

发布时间：2022-05-31 发布人：admin 分类：说明书资料大小：2.17M 资料格式：pdf 举报版权申诉

hxf19961219-12255102-4744302542848079001.pdf-第1页.png

第1页 / 共9页

hxf19961219-12255102-4744302542848079001.pdf-第2页.png

第2页 / 共9页

hxf19961219-12255102-4744302542848079001.pdf-第3页.png

第3页 / 共9页

hxf19961219-12255102-4744302542848079001.pdf-第4页.png

第4页 / 共9页

hxf19961219-12255102-4744302542848079001.pdf-第5页.png

第5页 / 共9页

hxf19961219-12255102-4744302542848079001.pdf-第6页.png

第6页 / 共9页

hxf19961219-12255102-4744302542848079001.pdf-第7页.png

第7页 / 共9页

hxf19961219-12255102-4744302542848079001.pdf-第8页.png

第8页 / 共9页

Automatic grape leaf diseases identification via UnitedModel based on multiple convolutional neural networks

1 Introduction

2 Related works

3 Materials and methods

3.1 Dataset and preprocessing

3.2 Identification model of grape leaf diseases

3.3 Architecture of UnitedModel

3.4 Experimental environment

3.5 Evaluation metrics

4 Results and discussion

5 Conclusions

Declaration of Competing Interest

Acknowledgements

References

A v a i l a b l e a t w w w . s c i e n c e d i r e c t . c o m INFORMATION PROCESSING IN AGRICULTURE xxx (xxxx) xxx j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / i n p a Automatic grape leaf diseases identiﬁcation via UnitedModel based on multiple convolutional neural networks Miaomiao Ji a, Lei Zhang b, Qiufeng Wu c,* a College of Engineering, Northeast Agricultural University, Harbin 150030, China b School of Medicine, University of Pittsburgh, Pittsburgh 15260, USA c College of Science, Northeast Agricultural University, Harbin 150030, China A R T I C L E I N F O A B S T R A C T Article history: Received 17 April 2019 Received in revised form 11 October 2019 Accepted 15 October 2019 Available online xxxx Keywords: Grape leaf diseases Identiﬁcation Multi-network integration method Convolutional neural network Deep learning Grape diseases are main factors causing serious grapes reduction. So it is urgent to develop an automatic identiﬁcation method for grape leaf diseases. Deep learning techniques have recently achieved impressive successes in various computer vision problems, which inspires us to apply them to grape diseases identiﬁcation task. In this paper, a united con- volutional neural networks (CNNs) architecture based on an integrated method is pro- posed. The proposed CNNs architecture, i.e., UnitedModel is designed to distinguish leaves with common grape diseases i.e., black rot, esca and isariopsis leaf spot from healthy leaves. The combination of multiple CNNs enables the proposed UnitedModel to extract complementary discriminative features. Thus the representative ability of United- Model has been enhanced. The UnitedModel has been evaluated on the hold-out PlantVil- lage dataset and has been compared with several state-of-the-art CNN models. The experimental results have shown that UnitedModel achieves the best performance on var- ious evaluation metrics. The UnitedModel achieves an average validation accuracy of 99.17% and a test accuracy of 98.57%, which can serve as a decision support tool to help farmers identify grape diseases. Ó 2019 China Agricultural University. Production and hosting by Elsevier B.V. on behalf of KeAi. This is an open access article under the CC BY-NC-ND license (http://creativecommons. org/licenses/by-nc-nd/4.0/). 1. Introduction Grapes, as one of the most commonly cultivated economical fruit crops throughout the world, are widely used in the pro- duction of wine, brandy or nonfermented drinks and are eaten fresh or dried as raisins [1]. However, grapes are vulner- able to various different types of diseases, such as black rot, esca, isariopsis leaf spot, etc. It is estimated that losses caused by grape diseases in Georgia, USA in 2015 were approximately $1.62 million. Around $0.5 million was spent on diseases control and the rest was the loss caused by the diseases [2]. Thus, early detection of grape diseases can potentially cut losses and control costs and consequently can improve the quality of products. * Corresponding author. E-mail address: qfwu@neau.edu.cn (Q. Wu). Peer review under responsibility of China Agricultural University. https://doi.org/10.1016/j.inpa.2019.10.003 2214-3173 Ó 2019 China Agricultural University. Production and hosting by Elsevier B.V. on behalf of KeAi. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). For decades, the diseases identiﬁcation is mostly per- formed by human. The process of recognition and diagnosis is subjective, error-prone, costly and time-consuming. In addition, new diseases can occur in places where they were Please cite this article as: M. Ji, L. Zhang and Q. Wu, Automatic grape leaf diseases identiﬁcation via UnitedModel based on multiple convolu- tional neural networks, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.10.003

2 I n f o r m a t i o n P r o c e s s i n g i n A g r i c u l t u r e x x x ( x x x x ) x x x previously unidentiﬁed and, inherently, where there is no local expertise to combat them [3]. As a result, an automatic identiﬁcation method for grape disease identiﬁcation is in urgent demand. The development of sophisticated instruments and fast computational techniques have paved the way for real-time scanning and automatic detection of anomalies in a crop [4]. Although traditional machine learning methods have gained some valuable experience in identiﬁcation and diag- nosis of crop diseases, they are limited to following the pipe- lined procedures of image segmentation (such as clustering method [5], threshold method [6], etc.), feature extraction (such as shape, texture, color features, etc. [7]), and pattern recognition (such as k-nearest neighbor method (KNN) [8], support vector machine (SVM) [1], back propagation neural network (BPNN) [9], etc.). It is difﬁcult to select and extract the optimal visible pathological features and thus highly skilled engineers and experienced experts are demanded, which is not only to a considerable extent subjective but also leads to a great waste of manpower and ﬁnancial resources. In contrast, deep learning techniques can automatically learn the hierarchical feature of pathologies and do not need to manually design the feature extraction and classiﬁer. Deep learning method has so excellent generalization ability and robustness that it excels in many areas: signal processing [10], pedestrian detection [11], face recognition [12], road crack detection [13], biomedical image analysis [14], etc. In addition, deep learning techniques have also achieved impressive results in the ﬁeld of agriculture and beneﬁt more smallholders and horticultural workers, including diagnosis of crop diseases [15], recognition of weeds [16], selection of ﬁne seeds [17], pest identiﬁcation [18], fruit counting [19], research on land cover [20] etc, which has contributed to deal- ing with image classiﬁcation of areas of interest. Further- more, some applications focused on predicting future parameters, such as crop yield [21], weather conditions [22] and soil moisture content in the ﬁeld [23]. Inspired by the great success of CNN-based methods in image classiﬁcation, we propose an integrated model, denoted UnitedModel, for automatic grape leaf disease identiﬁcation. The rest of this paper is organized as follows. Section 2 starts with an overview of related works. Section 3 introduces dataset as well as data preprocessing and covers details of the proposed UnitedModel. In Section 4, we conduct all the exper- iments, discuss the limitation of the proposed UnitedModel and prospect the future work. In Section 5, we conclude this paper. 2. Related works Machine learning techniques were applied widely in plant disease classiﬁcation in the early stage. Li et al. [24] proposed a method based on K-means clustering segmentation of the grape disease images and a SVM classiﬁer was designed based on thirty-one effective selected features to identify grape downy mildew disease and grape powdery mildew dis- ease with testing recognition rates of 90% and 93.33%, respec- tively. Signiﬁcant progress has been made in the use of image processing approaches to detect various diseases in crops. Athanikar and Badar [9] applied Neural Network to categorize the potato leaf image as either healthy or diseased. Their results showed that BPNN could effectively detect the disease spots and could classify the particular disease type with an accuracy of 92%. The deployment of deep CNNs has especially led to a breakthrough in plant disease classiﬁcation, which can ﬁnd very high variance of pathological symptoms in visual appearance, even high intra-class dissimilarity and low inter-class similarity that may be only noticed by the botanists [25]. Lee et al. [26] proposed a CNN approach to iden- tify leaf images and reported an average accuracy of 99. 7% on a dataset covering 44 species, but the scale of datasets was very small. Zhang et al. [27] used GoogLeNet to address the detection of cherry leaf powdery mildew disease and obtained an accuracy of 99.6%. Their results also demonstrated trans- fer learning can boost the performances of deep learning model [15] ﬁne-tuned deep learning models pre-trained on ImageNet to identify 14 crop species and 26 diseases. The models were evaluated on a publicly available dataset including 54,306 images of diseased and healthy plant leaves collected under controlled conditions. They achieved the best accuracy of 99.35% on a hold-out test dataset. in crop disease identiﬁcation. Mohanty et al. Although the basic CNN frameworks, such as AlexNet [28], VGGNet [29], GoogLeNet [30], DenseNet [31] and ResNet [32] have been demonstrated effective and widely used in crop diseases classiﬁcation, most previous works had troubling boosting up the classiﬁcation accuracy rate to some extent. As a matter of fact, a single model can’t meet the further requests in terms of precision. In large machine learning competitions, the best results were usually achieved by the integration of multiple models rather than by a single model. For instance, the well-known Inception-ResNet-v2 [33] was born out of the fusion of two excellent deep CNNs, as the name suggests. Inspired by network in network concept [34], we think that integration is the most straightforward and effective way when the basic models are signiﬁcantly dif- ferent. We propose the UnitedModel in this study. The United- Model is a integration of the powerful and popular deep learning architectures of GoogLeNet and ResNet. We also take advantage of transfer learning to boost accuracy as well as reduce the training time. 3. Materials and methods 3.1. Dataset and preprocessing Our dataset comes from an open access repository named PlantVillage which focuses on plant health [35]. The dataset used for evaluating the proposed method is composed of healthy (171 images) and symptom images including black rot (pathogen: Guignardia bidwellii, 476 images), esca (patho- gen: Phaeomoniella spp, 552 images) and isariopsis leaf spot (pathogen: Pseudocercospor a-vitis, 420 images) (see Table 1). The identiﬁcation of grape diseases is based on its leaf, not ﬂower, fruit or stem. For one thing, the ﬂower and fruit of grape only appear in a limited time while the leaf presents most of the year. For another thing, the stem of grape can hardly present the symptoms of the diseases timely, while Please cite this article as: M. Ji, L. Zhang and Q. Wu, Automatic grape leaf diseases identiﬁcation via UnitedModel based on multiple convolu- tional neural networks, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.10.003

I n f o r m a t i o n P r o c e s s i n g i n A g r i c u l t u r e x x x ( x x x x ) x x x 3 Table 1 – Introduction of grape leaf dataset. Label Category Number Data No data augmentation Leaf symptoms Illustration augmentation Training samples Validation samples Test samples 1 2 3 Black rot 476 Esca Isariopsis leaf spot 552 420 4 Health Total samples 171 1619 288 288 288 96 960 72 72 72 24 240 116 192 60 51 419 See Fig. 1 ﬁrst row Appear small, brown circular lesions Appear dark red or yellow stipes See Fig. 1 second row Appear many small rounded, polygonal or irregular brown spots – – See Fig. 1 fourth row – See Fig. 1 third row Fig. 1 – The grape leaf images after data augmentation. the leaf is generally sensitive to the state of plants, whose shape, texture and color usually contains richer information. The raw images are divided into training dataset and test dataset. 360 Symptom and 120 healthy images are selected for training and the rest images are used for testing. To pre- vent over-ﬁtting, the training dataset is further split into training (80%) and validation data (20%). Thus the training dataset is 960 samples in total, validation dataset is 240 sam- ples in total and test dataset is 419 samples in total [36]. The original images in the PlantVillage dataset are RGB images of arbitrary size. Data preprocessing is necessary and all the images are resized to the expected input size of the respective networks, i.e. 224 224 for VGGNet, DenseNet and ResNet, and 229 229 for GoogLeNet. Model optimization and predic- tion are both performed on these rescaled images. To avoid over-ﬁtting and to boost the generalizability of the CNNs, data augmentation techniques which include rotating, ﬂipping, shearing, zooming and colour changing are randomly per- formed in training phase as shown in Fig. 1. Considering that the number of healthy samples in our work is relatively small with 171 samples in total, the propor- tion between healthy samples and diseased samples is set to 1: 3 while the class_weight ratio of them is set to 3: 1. By this way, the loss function can be adjusted in the training process and healthy samples can obtain more attention. 3.2. Identiﬁcation model of grape leaf diseases In this paper, the grape leaf diseases identiﬁcation can be for- mulated as a multi-class classiﬁcation problem. The classiﬁ- cation process of this work can be described as follows (Fig. 2). Assuming a mapping function from grape leaf images ‘‘Label2”, to corresponding predicted labels (i.e. ‘‘Label3” and ‘‘Label4”) f : X ! Y: Then the grape leaf training Þ. The train- dataset is denoted as ðx1; y1 ing dataset is the input of the CNN model A. During training, the weights of A are updated continuously to minimize the average loss function which is deﬁned as: JðxÞ ¼ 1 n ‘‘Label”, Þ ;:::; ðxn; yn yilogðhxðxiÞÞ Þ; ðx2; y2 ð1Þ Xn i¼1 where x is the training sample, n is the number of input data, yi is the true label and hxðxiÞ is the predicted label of the CNN model A given current weights x. Then, the end-to-end opti- mization can be performed using the stochastic gradient des- cent (SGD) method. The gradient of the loss function can be computed with the following equation: xtþ1 ¼ lxt arJðxtÞ ð2Þ where l is the momentum weight for the current weights xt and a is the learning rate. Noted that the iterative operation Please cite this article as: M. Ji, L. Zhang and Q. Wu, Automatic grape leaf diseases identiﬁcation via UnitedModel based on multiple convolu- tional neural networks, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.10.003

4 I n f o r m a t i o n P r o c e s s i n g i n A g r i c u l t u r e x x x ( x x x x ) x x x InceptionV3 is selected as one component of the UnitedMo- del. ResNet won the ﬁrst place in ILSVRC 2015 and COCO 2015 classiﬁcation challenge with error rate of 3.57%, which introduced residual units for addressing the degradation problem of CNNs. ResNet is stacked by many residual units. Given the size of the dataset used in this work, ResNet50 becomes our ﬁrst choice to be another basic network of the UnitedModel, which also has high performance in classiﬁca- tion tasks. Given sufﬁcient data, the most straightforward way of improving the performance of CNNs is increasing their parameters. In this study, GoogLeNet increases its width: the number of units at each layer while ResNet increases the depth: the number of network layer. The integration is most effective when the basic models are signiﬁcantly different because their combination will enable the integrated network to capture more distinguishing feature information, thereby improving identiﬁcation accuracy. An illustration for the workﬂow of the proposed method is shown in Fig. 3. Given a testing image, the well-trained Uni- tedModel extracts two different feature sets by using GoogLe- Net and ResNet and then performs the prediction based on the combined feature sets. In Fig. 4, we show the detailed structure of the UnitedModel. Considering that our dataset is relatively small, direct training from scratch on the dataset of grape leaf diseases will result in over-ﬁtting. And thus transfer learning is adopted. The training phase is as follows: Firstly, the input image is resized to the expected input size of the respective networks and propagated forward to each basic single model, i.e. InceptionV3 and ResNet50. The feature maps are extracted by multi-type convolution ﬁlters arranged in InceptionV3 and ResNet50. And then through global aver- age pooling (GAP), the feature maps extracted from convolu- tion layers will be output in the form of two-dimensional matrix and then merged operation concatenation (Fig. 3). After that, a dropout layer, a fully connected layer, an another dropout layer, a classiﬁcation layer (Fig. 4) are added. Within Fig. 2 – The mathematical model of grape leaf disease identiﬁcation. of x is limited in added layers because all parameters in con- volutional layers are set to the pre-trained network weights and ﬁxed. After a series of epochs, an optimal function g : X ! Y is obtained, which is the ﬁnal hypothesis. Then we can input the test dataset into A and get the correspond- ing predicted labels. 3.3. Architecture of UnitedModel The UnitedModel proposed for grape leaf disease identiﬁca- tion is based on two popular deep learning architectures i.e., GoogLeNet and ResNet. GoogLeNet, the winner of ILSVRC 2014, was precisely aimed to reduce the computational cost and complexity. The whole architecture of GoogLeNet is stacked by inception modules that combine multi-scale fea- ture extraction and dimension reduction layers based on the Hebbian principle and embedding learning. In our work, Fig. 3 – An illustration for the workﬂow of the proposed method. Please cite this article as: M. Ji, L. Zhang and Q. Wu, Automatic grape leaf diseases identiﬁcation via UnitedModel based on multiple convolu- tional neural networks, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.10.003

I n f o r m a t i o n P r o c e s s i n g i n A g r i c u l t u r e x x x ( x x x x ) x x x 5 Fig. 4 – Structure of UnitedModel. the classiﬁcation step, each neuron in fully-connected layers provides a full connection to all learned feature maps. Accordingly, the united two-dimensional feature matrix com- bines freely with each other, generating different feature maps and thus improving the ﬂexibility and characterization ability of the model [37]. Since the abundant training param- eters and complex structures in the proposed UnitedModel easily cause over-ﬁtting, dropout layers are used with rate 0.5. The softmax classiﬁer is used as the ﬁnal classiﬁcation layer to obtain the classiﬁcation results. 3.4. Experimental environment The proposed model is built using the open source Keras framework on top of TensorFlow. All experiments are con- ducted on a system with the speciﬁcations in Table 2. The original architectures of VGGNet, GoogLeNet, Dense- Net and ResNet are trained as the control experiments of Uni- tedModel to explore whether the integrated approach can improve the performance in the grape diseases identiﬁcation task and where the improvement is signiﬁcant. To enable a fair comparison between the results of all the experimental conﬁgurations, all the hyper-parameters are standardized. The SGD algorithm is used to learn the best set of weights and biases of the neural networks that minimize the loss function. While learning, the SGD algorithm works by ran- domly choosing a small number of training inputs referred to as batch size which is set to 16. The learning rate is set to 0.001. It is the rate at which a function moves through the search space. A small learning rate leads to more precise results but it requires more training time. The momentum is an additional factor to determine how fast the SGD algorithm converges on the optimum point. It is set to 0.9. Early stopping mechanism is used to deal with the problem of over-ﬁtting and speed up the training process. The training ends when Table 2 – Experimental environment. Hardware Software CPU: Intel Core i7-6700HQ RAM: 8 GB DDR4 GPU: NVIDIA GEFORCE GTX 960M Keras, Tensorﬂow Windows10 CUDA9. 0+CUDNN7.0 the network does not signiﬁcantly improve its validation per- formance after 3 epochs. 3.5. Evaluation metrics The evaluation of the different experimental conﬁgurations is based on a train-validation-test scheme. The actual training is carried-out on the training dataset while the validation dataset is used to avoid over-ﬁtting. The performance of each model is assessed on the test dataset. Precision, recall and average F1-score are used as evaluation metrics. Experimen- tal results are reported across all our experimental conﬁgura- tions. Due to its increased sensitivity to imbalances among the classes, the overall accuracy is also computed [38]. The precision, recall and F1-score are expressed in the following: ð3Þ precisionc ¼ TPc TPc þ FPc recallc ¼ F1 score ¼ 2 4 c¼1 TPc TPc þ FNc X4 recallc precisionc recallc þ precisionc ð4Þ ð5Þ where TP is true positive, FP is false positive, FN is false neg- ative, c is the class of grape diseases and healthy type is included. For a more detailed comparison at different operat- ing points, confusion matrix analysis for VGGNet, GoogLeNet, DenseNet, ResNet and UnitedModel is also performed. Please cite this article as: M. Ji, L. Zhang and Q. Wu, Automatic grape leaf diseases identiﬁcation via UnitedModel based on multiple convolu- tional neural networks, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.10.003

6 I n f o r m a t i o n P r o c e s s i n g i n A g r i c u l t u r e x x x ( x x x x ) x x x Fig. 5 – Validation accuracy and validation loss of various architectures are compared and the number of epoch is varied in the case of early stopping mechanism. VGGNet, GoogLeNet, DenseNet, ResNet and UnitedModel stop training with the epoch 11, 18, 11, 9 and 12, respectively. Fig. 6 – Comparison of the average accuracy for the different architectures. 4. Results and discussion As illustrated in Fig. 5, UnitedModel has a tendency to consis- tently improve in accuracy with growing number of epochs, with no signs of performance deterioration. On the other hand, loss curve of UnitedModel converges rapidly and has a big gap with other single models in terms of error rate. Con- sistent with Fig. 6, the performance of GoogLeNet, VGGNet, DenseNet and ResNet to UnitedModel. similar and all is inferior As we can see in Table 3, the average precision (99.05%), recall (98.88%) and F1-score (98.96%) of UnitedModel are the highest among all these models. And scores of VGGNet are the lowest with precision (94.16%), recall (93.32%) and F1- score (94.40%), which once again proves the superiority of integrated CNNs. The confusion matrices of different models on test dataset are shown in Fig. 7. The threshold is set to 0.5 and the fraction of accurately predicted images for each class is displayed in detail. The performance of VGGNet is poor. VGGNet tends to distinguish diseased samples as healthy, which is extremely harmful in actual agriculture produce. All the models except VGGNet can distinguish label3 and label4 easily, but label1 and label2 are prone to be misclassiﬁed. The large green area of healthy leaves and the golden appearance of leaves infected with isariopsis leaf spot disease make them easier to be distinguished. Meanwhile, the confusion between label1 and label2 is due to their similar pathological features. Please cite this article as: M. Ji, L. Zhang and Q. Wu, Automatic grape leaf diseases identiﬁcation via UnitedModel based on multiple convolu- tional neural networks, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.10.003

I n f o r m a t i o n P r o c e s s i n g i n A g r i c u l t u r e x x x ( x x x x ) x x x 7 Table 3 – Precision, recall and F1-score for the corresponding experimental conﬁgurations (best overall performance is displayed in boldface). Architecture Criterion Label Support VGGNet GoogLeNet DenseNet ResNet UnitedModel Precision Recall F1-score Precision Recall F1-score Precision Recall F1-score Precision Recall F1-score Precision Recall F1-score Label1 116 0.8699 0.9224 0.8954 0.9076 0.9310 0.9191 0.9642 0.9655 0.9372 0.9737 0.9569 0.9652 0.9825 0.9655 0.9739 Label2 192 0.9514 0.9167 0.9337 0.9577 0.9427 0.9501 0.9784 0.9427 0.9602 0.9742 0.9844 0.9793 0.9794 0.9896 0.9845 Label3 60 0.9828 0.9500 0.9661 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 Label 4 51 Total/Average 419 0.9623 1.0000 0.9808 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9416 0.9332 0.9440 0.9663 0.9684 0.9673 0.9722 0.9771 0.9744 0.9870 0.9853 0.9861 0.9905 0.9888 0.9896 Fig. 7 – Confusion matrices for the grape diseases dataset are presented with an overall accuracy associated for individual classes. The confusion matrices are given in terms of percentages, not absolute numbers. However, UnitedModel improves the test accuracy of label1 (97%) and label2 (99%). The UnitedModel is a combination of InceptionV3 and ResNet50 and can improve the representa- tive capability by enriching features using multi-network integration method. In this study, we conjecture different CNN models to generate complementary features and we boost the performance by combining the complementary fea- tures. In addition, the use of transfer learning makes this pro- cess more concise and efﬁcient. The loss and accuracy curves of the UnitedModel during training phase are presented in Fig. 8. The green and black curves are the loss function values on the training dataset and validation dataset, respectively. The two curves drop rapidly in the ﬁrst two epochs and start to converge after around 6 epochs. And then, the training loss continues to des- cend slightly until roughly the ﬁnal epoch while validation loss is no longer signiﬁcantly reduced. The red and blue the training accuracy and validation curves represent Please cite this article as: M. Ji, L. Zhang and Q. Wu, Automatic grape leaf diseases identiﬁcation via UnitedModel based on multiple convolu- tional neural networks, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.10.003

8 I n f o r m a t i o n P r o c e s s i n g i n A g r i c u l t u r e x x x ( x x x x ) x x x support tool to help farmers identify the grape diseases. We also provide a practical study for dealing with the insufﬁ- ciency and unbalance of dataset. The methods of data aug- mentation techniques, early stopping mechanism and dropout are used to improve the generalization ability of model and reduce the risk of over-ﬁtting. In addition, we have proposed an effective multi-network integration method which can be used to integrate more state-of-the-art CNN models and can also be easily extended to other plant disease identiﬁcation tasks. Declaration of Competing Interest We declare that we have no ﬁnancial and personal rela- tionships with other people or organizations that can inap- propriately inﬂuence our work, there is no professional or other personal interest of any nature or kind in any product, service and company that could be construed as inﬂuencing the position presented in, or the review of, the manuscript entitled ”Automatic Grape Leaf Diseases Identiﬁcation via UnitedModel Based on Multiple Convolutional Neural Networks”. Acknowledgements This work was supported by the Public Welfare Industry (Agri- culture) Research Projects Level-2 under Grant 201503116-04- 06; Postdoctoral Foundation of Heilongjiang Province under Grant LBHZ15020; Harbin Applied Technology Research and Development Program under Grant 2017RAQXJ096 and National Key Application Research and Development Program in China under Grant 2018YFD0300105-2. R E F E R E N C E S [1] Kole DK, Ghosh A, Mitra S. Detection of downy mildew disease present in the grape leaves based on fuzzy set theory. In: Advanced computing, networking and informatics. Berlin (Germany): Springer; 2014. p. 377–84. [2] Martinez A. Georgia plant disease loss estimates. link: http:// www.caes.uga. edu/Publications/displayHTML.cfm?pk_id= 7762. 2015. [3] Srdjan S, Marko A, Andras A, Dubravko C, Darko S. Deep neural networks based recognition of plant diseases by leaf image classiﬁcation. Comput Intell Neurosci 2016;2016:1–11. [4] Wang G, Sun Y, Wang J. Automatic image-based plant disease severity estimation using deep learning. Comput Intell Neurosci 2017;2017:1–8. [5] Xu G, Yang M, Wu Q. Sparse subspace clustering with low- rank transformation. Neural Comput Appl 2017;2017:1–14. [6] Singh V, Misra A. Detection of plant leaf diseases using image segmentation and soft computing techniques. Inf Process Agric 2017;4(1):41–9. [7] Ma J, Du K, Zheng F, Zhang L, Gong Z, Sun Z. A recognition method for cucumber diseases using leaf symptom images based on deep convolutional neural network. Comput Electron Agric 2018;154:18–24. [8] Xu X, Liu Z, Wu Q. A novel K-nearest neighbor classiﬁcation algorithm based on maximum entropy. Int J Adv Comput Technol 2013;5(5):966–73. Fig. 8 – The trend of train accuracy, validation accuracy, train loss and validation loss of the UnitedModel. accuracy, respectively. In Fig. 8, we can observe that after a few epochs they overlap almost completely, showing that UnitedModel gets sufﬁciently trained. Experimental results have shown that the proposed Uni- tedModel integrating InceptionV3 and ResNet50 can achieve promising results for the grape leaf diseases identiﬁcation task. However, since the training dataset used in our work is composed of samples in a simple background scene, the trained model cannot be applied for real-time diagnosis of grape leaf diseases in the complex background. One direction of future work is to create datasets under un-controled sce- nario which can be used for training model to improve the generalizability. Thus the proposed model can be used for the development of mobile systems and devices. Another direction of future work is to compress the model while keep- ing the same performance. The proposed UnitedModel is a combination of two popular models and is of huge number the parameters. The training of the UnitedModel is great com- putational cost. It is important to ﬁnd out an effective pruning mechanism for model compression to reduce computational resources. 5. Conclusions In this work, we have developed an effective solution to auto- matic grape diseases identiﬁcation based on CNNs. We have proposed the UnitedModel which is a united CNNs architec- ture based on InceptionV3 and ResNet50 and can be used to classify grape images into 4 classes, including 3 different symptom images i.e., black rot, esca, isariopsis leaf spot, and healthy images. UnitedModel takes advantage of the combination of InceptionV3’s width and ResNet50’s depth and thus can learn more representative features. The repre- sentational ability of UnitedModel is strengthened by way of high-level feature fusion, which makes it achieve the best performance in the grape diseases identiﬁcation task. The experimental results demonstrate that our model can outper- form the state-of-the-art basic single CNNs including VGG16, InceptionV3, DenseNet121 and ResNet50. The proposed Uni- tedModel achieves an average validation accuracy of 99.17% and a test accuracy of 98.57% and thus can serve as a decision Please cite this article as: M. Ji, L. Zhang and Q. Wu, Automatic grape leaf diseases identiﬁcation via UnitedModel based on multiple convolu- tional neural networks, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.10.003

分享到：

赞收藏

资料库

Information Processing in Agriculture.pdf

相关推荐

人工智能

热门标签

最新资料