Deep Learning based Automated Disease Detection and Classi�cation Model for Precision Agriculture

Plant phenotyping and Precision agriculture are information-and technology-oriented fields with specific challenges and demands for the detection and diagnosis of plant disease. Precision agriculture can be referred as a crop management method related to the spatial and temporal variability in soil and crop factors within a field. Accurate and early diagnosis and detection of plant diseases were major factors in plant production and the reduction of quantitative and qualitative losses in crop yield. Advancement of automatic disease detection and classification system is significantly explored in precision agriculture. In recent times, research workers have investigated numerous cultures leveraging dissimilar parts of a plant. This article develops a new Deep Learning based Automated Plant Disease Detection and Classification (DL-APDDC) Model for Precision Agriculture. The presented DL-APDDC algorithm concentrates on the recognition and classification of plant diseases in leaf and fruit regions. In the initial stage, the leaf and fruit regions are extracted by the use of U2Net based background removal. Next, the Adam optimizer with SqueezeNet model is exploited as feature extractor and the hyperparameters are tuned by Adam optimizer. Finally, the extreme gradient boosting (XGBoost) classifier performs classification of plant diseases. The experimental validation of the DL-APDDC technique is tested on benchmark plant disease dataset. The simulation values indicated the enhanced outcomes of the DL-APDDC approach over other models.


Introduction
Precision agriculture is a concept of farm management based on measuring, responding, and monitoring crop variability [1].The study aims to determine the decision support system (DSS) for farming management by improving the returns on input when maintaining resources.
MPlant disease was a threat to food security around the world but has disastrous consequences for small-holder farmers whose livelihood depends on healthy crops [2].In the developing world, around 80% of agriculture production is produced by smallholder farmers (UNEP, 2013), and reports of crop loss of above 50% due to diseases and pests are prevalent.Moreover, a larger fraction of hungry people lives in smallholder agricultural households, which makes the smallholder farmer a group that's mainly susceptible to pathogen-derived disruption in food supplies [3].Due to crop diseases, numerous efforts are being made to prevent these losses.
Machine vision was extensively applied to assist precision agriculture by giving automatic solutions to tasks that are performed conventionally [4].Manual method tends to be error prone and tedious.Machine vision provides efficient and accurate solutions to assist agricultural activity.Furthermore, ML algorithm enabled the analysis of great amounts of information precisely and rapidly, providing a means for the application of machine vision in agriculture [5].ML algorithm enables to analyze huge amounts of information, irrespective of complexity, accurately and quickly.Its use is already common in several fields namely credit analysis, fraud detection, image recognition patterns, fault prediction models, product quality analysis, and intelligent spam filters [6].But considering the different alternatives, it is crucial to know the individual features of every technique and the better scenario for its usage.
The wide usage of precision farming worldwide is due to the existence of advanced machine learning (ML) and deep learning (DL) approaches [7], efficient computational devices, and high-speed internet access.DL architecture tackles crucial problems by using different feature extraction techniques [8].Because of the higher computation cost of training models from SVM, Google Net, scratch VGG, and so on, are used that previously have present weight.Pretrained network provides fast convergence and is utilized in advanced techniques to tasks such as activity recognition, object detection, and segmentation [9].DL network achieves advanced performance in other fields of study are not applicable for agricultural tasks of crop management namely fertilization, irrigation, picking, and pesticide spraying.This situation demonstrates the requirement to create proper crop dataset by using different devices for wider and deeper networks to produce best outcomes [10].

Related Works
Venkataramanan et al. [11] modelled a DL algorithm for classifying and detecting plant diseases by analyzing the leaf of a given plant.In this study, the classification can be accomplished in many stages for eradicating possibilities at every stage, thereby rendering higher accuracy levels during estimations.To extract a leaf from the input image a method named YOLOv3 object detector was employed.This extracted leaf will be examined by a sequence of ResNet18 techniques.Such ResNet18 techniques have been trained by utilizing TL.Hossain et al. [12] modelled a method uses the KNN algorithm for detecting and classifying plant leaf disease.The texture features can be derived from leaf disease images for classifying purposes.In this study, the diseases such as leaf spot, alternaria, canker, bacterial blight, alternata, and anthracnose of many plant species will be classified by KNN classifier.Sardogan et al. [13] devised a Learning Vector Quantization (LVQ) algorithm-oriented approach and CNN method for classifying and detecting tomato leaf diseases.The author has devised a CNN for automatic classification and feature extraction.In plant leaf disease research, color information is actively used.In this method, based on RGB elements the filters were implemented to 3 channels.The LVQ was given to the output feature vector of convolution part to train the network.In [14], the author formulated an effective rice plant disease detection approach related to CNN method.This study had a focus towards 3 renowned rice diseases they are bacterial leaf blight caused by bacteria, brown spots, and leaf smut caused by fungus.This presented method would apply Otsu's global thresholding approach for performing image binarization to eliminate background noise of images.
Ashwinkumar et al. [15] present an automated method for classifying and detecting plant leaf diseases with the use of an optimal mobile network-oriented CNNs (OMNCNNs).It encompasses Kapur's thresholding-oriented image segmentation and bilateral filtering (BF) oriented preprocessing for identifying the affected areas of the leaf image.Likewise, the MobileNet technique was enforced as a feature extracting method where the hyperparameters will be optimized through the emperor penguin optimizer (EPO) technique for improving the plant disease recognition rate.At last, ELM related method was employed for allocating suitable class labels to the applied plant leaf images.Guo et al. [16] presented a mathematical technique of plant disease recognition and detection relies upon DL, which will improve training efficiency, precision, and generality.Firstly, the region proposal networks (RPNs) are leveraged for localizing and recognizing the leaves in complex surroundings.Next, images segmented on the basis of the outcomes of RPN method have the feature of symptoms by utilizing Chan-Vese (CV) method.

The Proposed Model
In

U2Net based Background Removal
Primarily, the leaf and fruit regions are extracted by the use of U2Net based background removal.The U2-net is a two-level nested U-architecture [17].The outer layer is the largest Uarchitecture comprising 11 stages.Every phase is populated through a residual U-block (RSU) (inner layer).Ideally, the nested U-architecture enables the extraction of multiscale and multilevel features most effectively.It comprises three parts: (1) map fusion module, (2) encoder, and (3) decoder.(1) There are 6 phases in the encoder phase.Every phase is comprised of RSU.The feature map can be decreased for increasing the receptive field and to attain more largescale data.In the next two phases, dilated convolution is utilized for replacing the pooling function.This phase is needed for preventing context data loss.Note that the receptive field is improved while the feature map is not decreased.(2) The decoder stage has structure same as the encoder stage.Every decoder phase concatenates the up-sampled feature map from its preceding phase and those from symmetrical encoder phase as the input.(3) Feature map combination with the deep supervision approach is the final phase used to produce a probability map.The model generates six side outputs.Then, this output is up-sampled to the size of input images and merged with a concatenation function.To summarize, the U2-net design has a deeper structure with rich multiscale features, low memory, and computing costs.Furthermore, the U2-net structure is based on the RSU block and doesn't utilize pre-trained backbone, it is easy and flexible t

Feature Extraction Model
In this study, the SqueezeNet model is exploited as feature extractor.The most important benefit of CNN over traditional classification algorithms is that, in CNN, the classifier and the representation of the features are deployed in a similar network, which eliminates the requirement for them to be reliant on each other [18].Convolution layer is comprised a fixed set of learnable filters, are considered the significant layers in CNN.2D activation maps are made by sliding the filter over input visual dataset during the forward pass.The strength and location of the recognized visual feature of the input images are characterized by the activation maps.The significant feature of CNN is the pooling layer, which is mainly applied between consecutive convolutional layers to gradually decrease the spatial presentation size concurrently maintaining the relevant data.This assists in controlling the overfitting during the learning procedure.Regularization is a common methodology in CNN to prevent the over-fitting consequence by adding a significant amount of penalty to the loss function.The neuron of the FC layers is coupled with each activation of the previous layer in the final section of CNN architecture to minimalize feature dimension.CNN has a last pooling layer that smooth the convolution layer that is afterward transferred to the node of network that is connected wholly.In the next phase, the activation is calculated using matrix multiplication that is followed by the implementation of bias factor.
A CNN is a FFNN that mainly comprises pooling and convolutional layers.In comparison to the designed image feature artificially in conventional computer vision, the image feature extracted through the CNN could better reflect the real characteristics of an image and are very efficient when compared to conventional computer vision method in the fields of image recognition and classification.Thus, CNN is utilized rather than the conventional bag-of-words (BoW) to implement loopback recognition.SqueezeNet was intended by UC Berkeley and Stanford researchers, not to accomplish better CNN detection performance, however, to simplify the difficulty of the network and accomplish the detection performance of public networks.The major configuration of SqueezeNet is the Fire model.
SqueezeNet comprises 8 Fire models from fire2 to fire9, whereas the architecture of fire2, whereby H and W denote the height and width of feature maps, correspondingly, and e3 shows the channel count.SqueezeNet decreases the number of flops and parameters in the network through Fire module that comprises expand and squeeze models.The squeeze model makes use of e1 1 x 1 convolution kernel to decrease the dimensionality of feature maps, and the feature maps are H x W x e1 afterward ReLU activation.The feature map is H x W x e2 afterward the two taps of expanding (with 1 x 1 and 3 x 3 convolution kernels, correspondingly) are separately calculated, and later the feature map of the two taps is stitched together by the concat layer to produce H x W x e3.Only the number of channels varies meanwhile, the convolution kernel of the fire network is similar in size.The fire module can be labeled with the number of channels of the fire model output e3.Once the network input is 224 x 224 x 3 RGB images, the computation of network is 837 MFlops, hence the network architecture is very applicable for lightweight devices like intelligent mobile robots.
The hyperparameter tuning process is carried out by the use of Adam optimizer.The Adam optimizer was broadly applied in the DL fields by virtue of its relatively fast convergence speed and self-adaptive learning rate [19].Thus, the study adopts the Adam model to upgrade the SqueezeNet model parameter.The training step is defined in the following: Wherein,   signifies the average value of gradient index at  time, and   signifies the squared gradient at  time. −1 denotes average value of gradient index at preceding time, and  −1 denotes squared gradient at preceding duration. 1 = 0.9, and  2 = 0.999, This two values were hyperparameter that controls attenuation of moving average.From the expression: R indicates the maximal iteration count;  denotes the attenuation factor, within the value of 0.99; it is viewed that development of learning rate depending on the value of the rate of learning in the preceding phase and gradient value of present phase was utilized for adaptive adjustment.Afterward the time  is accelerated, which improves the possibility to prevent the oscillation zone.

Plant Disease Detection Model
At the final stage, the XGBoost classifier performs classification of plant diseases.XGBoost was an ensemble model depends on gradient boosted tree [20].The outcome of prediction was the sum of the score forecasted by  trees, as follows: In Eq. ( 6),   denotes -ℎ of the training samples,   (  ) represent the score for -ℎ tree, and  indicates space of function comprising each gradient boosted tree and it can be attained as follows: In Eq. ( 7), the previous ∑

Results and Discussion
The experimental validation of the DL-APDDC method is tested using two plant disease datasets: potato leaf disease dataset [21] and citrus fruit dataset [22].The potato dataset has 5702 samples with three classes and the citrus fruit dataset comprises 73 samples with five classes as defined in Table 1.In Table 2 and Fig. 7, an overall plant disease classification results of the DL-APDDC model on potato dataset are given.The outcomes exhibited the DL-APDDC method has attained effectual outcomes under both TR and TS datasets.For instance, on TR set, the DL-APDDC model has attained   of 96.44%,   of 96.55%,   of 96.44%, 1  of 96.45%, and   of 96.64%.Meanwhile, on TS set, the DL-APDDC technique has achieved   of 95.62%,   of 95.75%,   of 95.55%, 1  of 95.62%, and   of 99.71%.Simultaneously, the linear discriminant, Quadratic SVM, and Cubic SVM techniques have reported moderately closer   values of 74.09%, 77.15%, and 78.82% correspondingly.

Conclusion
In Deep Learning based Automated Plant Disease Detection and Classification (DL-APDDC) Model for Precision Agriculture.The presented DL-APDDC approach concentrates on the recognition and classification of plant diseases in leaf and fruit regions.In the initial stage, the leaf and fruit regions are extracted by the use of U2Net based background removal.Next, the Adam optimizer with SqueezeNet model is exploited as feature extractor and the hyperparameters are tuned by Adam optimizer.Finally, the extreme gradient boosting (XGBoost) classifier performs classification of plant diseases.The experimental validation of the DL-APDDC technique is tested on benchmark plant disease dataset.
this article, we have introduced an automated plant disease classification model, named DL-APDDC technique.It mainly focuses on the categorization of different plant diseases affected in the leaf and fruit regions.The presented DL-APDDC technique encompasses U2Net based background removal, SqueezeNet feature extraction, Adam optimizer, and XGBoost based classification.Fig. 1 represents the working process of DL-APDDC system.

2 𝑡𝜂
later obtain the final value of the updating variable:   =  −1 −  *  ̃ + √  ̃ , whereas the first value of update learning rate was 1 = 0.01, and in sampling training procedure, it is noted that, The convergence tendency is closer to power exponential function.Thus, this study adds a power exponent correction term to learning rate at  time   =

Fig. 6 .
Fig. 6.Potato Dataset a) Confusion Matrix Training Set b) Confusion Matrix Testing Set c) PR-Curve Training Set d) PR-Curve Training Set e) ROC Training Set f) ROC Testing Set

Fig. 9 .
Fig. 9. Citrus Dataset a) Confusion Matrix Training Set b) Confusion Matrix Testing Set c) PR-Curve Training Set d) PR-Curve Training Set e) ROC Training Set f) ROC Testing Set

Fig. 10 .
Fig. 10.Result analysis of DL-APDDC system under Citrus dataset Table 5 and Fig. 11 exhibits an overall   inspection of the DL-APDDC technique on citrus dataset.The results exhibited the Linear SVM technique has achieved least   of 74%.

Fig. 11 .
Fig. 11.Comparative analysis of DL-APDDC system under Citrus dataset Contrastingly, the Otsu method has managed to portray considerable   of 83.95%.But the DL-APDDC technique has shown maximal classification performance with   of 96.08%.By observing the results and discussion, it is confirmed that the DL-APDDC model has gained maximum plant disease classification performance.
this article, we have introduced an automated plant disease classification model, named DL-APDDC technique.It mainly focuses on the categorization of different plant diseases affected in the leaf and fruit regions.At the initial stage, the leaf and fruit regions are extracted by the use of U2Net based background removal.Next, the Adam optimizer with SqueezeNet model is exploited as feature extractor and the hyperparameters are tuned by Adam optimizer.Finally, the XGBoost classifier performs classification of plant diseases.The experimental validation of the DL-APDDC technique is tested on benchmark plant disease dataset.The simulation values indicated the enhanced outcomes of the DL-APDDC approach over other models.In future, the classification accuracy of the DL-APDDC algorithm can be boosted by the DL classification techniques.

Table 2
Result analysis of DL-APDDC system with distinct measures under Potato dataset

Table 3 and
Fig. 8 highlight an overall   examination of the DL-APDDC model on potato dataset.The outcomes signified that the CNN-RF method has attained least   of 79%.Simultaneously, the CNN-SVM, GoogleNet, and VGGNet techniques have reported moderately closer   values of 84%, 86%, and 86% respectively.Contrastingly, the CNN-ANN model has managed to portray considerable   of 92%.But the DL-APDDC model has shown maximum classification performance with   of 96.44%.

Table 3
Comparative analysis of DL-APDDC system with other approaches under potato

Table 4
Result analysis of DL-APDDC system with distinct measures under Citrus dataset

Table 4 and
Fig.10, an overall plant disease classification outcomes of the DL-APDDC approach on citrus dataset is given.The outcomes designated the DL-APDDC approach has acquired effectual outcome under both TR and TS datasets.For example, on TR set, the DL-APDDC approach has accomplished   of 96.08%,   of 98.40%,   of 94.92%, 1  of 96.45%, and   of 99.91%.In the meantime, on TS set, the DL-APDDC technique has reached   of 95.45%,   of 98.33%,   of 90%, 1  of 92.46%, and   of 100%.

Table 5
Comparative analysis of DL-APDDC system with other approaches under Citrus