Internet of Things (IoT) paves a new direction in the domain of smart farming and precision agriculture. Smart farming is an upgraded version of agriculture which is aimed at improving the cultivation practices and yield to a certain extent. In smart farming, IoT devices are linked among one another with new technologies to improve the agricultural practices. Smart farming makes use of IoT devices and contributes in effective decision making. Rice is the major food source in most of the countries. So, it becomes inevitable to detect rice plant diseases during early stages with the help of automated tools and IoT devices. The development and application of Deep Learning (DL) models in agriculture offers a way for early detection of rice diseases and increase the yield and profit. This study presents a new Convolutional Neural Network-based inception with ResNset v2 model and Optimal Weighted Extreme Learning Machine (CNNIR-OWELM)-based rice plant disease diagnosis and classification model in smart farming environment. The proposed CNNIR-OWELM method involves a set of IoT devices which capture the images of rice plants and transmit it to cloud server via internet. The CNNIR-OWELM method uses histogram segmentation technique to determine the affected regions in rice plant image. In addition, a DL-based inception with ResNet v2 model is engaged to extract the features. Besides, in OWELM, the Weighted Extreme Learning Machine (WELM), optimized by Flower Pollination Algorithm (FPA), is employed for classification purpose. The FPA is incorporated into WELM to determine the optimal parameters such as regularization coefficient C and kernel
Agriculture is the most important source of income for human beings across the globe [
Indian population is undergoing phenomenal growth due to which cultivation and agricultural practices should also be improved robustly. Being a staple crop, Rice is the most sought-after food crop in India across the nation [
In order to overcome the issues discussed above, various hybrid methods have been proposed that generated supreme results. Internet of Things (IoT) is defined as a consortium of sensing devices, communication components, and users. IoT has the capability to generate a superior opportunity through which energetic industrial networks as well as real-world domains can be developed through the deployment of wireless sensors. It becomes possible to combine and introduce the observation mechanism in data gathering with the help of IoT and embedded mechanism. The application of IoT is the prominent step in this network. This is because it possesses an inference module to predict plant diseases and categorize it under nutrient insufficiency by applying wireless communication system. This system is considered to be an optimal path in the mitigation of issues involved in prediction and grouping of plant diseases. Next, the dataset gathered from farming is leveraged to create agriculture hazard alerting technology and corresponding Decision Support System (DSS). The selection of best features is highly significant as IoT information is produced in a fast manner. The availability of massive and dissimilar data mitigates the generalization function. However, this problem can be overcome with the application of Machine Learning (ML) technique which is enhanced with classification accuracy by reducing the count of parameters.
Image processing technique has few steps to be followed in disease detection such as image acquisition, preprocessing, segmentation, feature extraction and classification [
The current research article introduces a novel Convolutional Neural Network-based inception with ResNset v2 model and Optimal Weighted Extreme Learning Machine (CNNIR-OWELM)-based rice plant disease diagnosis and classification model in smart farming environment. The CNNIR-OWELM method utilizes different IoT devices to capture the images of rice plant and send it to cloud server via internet. The CNNIR-OWELM method makes use of histogram segmentation technique to determine the affected regions in the rice plant image. Moreover, Deep Learning (DL)-based inception with ResNet v2 method is applied for feature extraction. Here, OWELM is deployed for classification purposes. Flower Pollination Algorithm (FPA) is also incorporated with WELM to determine the optimal parameters such as regularization coefficient C and kernel
Few research investigations have been conducted earlier regarding the recognition and classification of rice plant infections. Based on Deep Convolutional Neural Network (DCNN), a new rice plant disease prediction method was deployed by Lu at al. [
Different classification methods are in use for representation whereas Random Forest (RF) model resolves alternate models. In this framework, a dataset with defected and non-defected leaf images is employed. Under the application of IP modules, Nidhis et al. [
Here, Naïve Bayes (NB) classifier is a simple classification model used for classifying the disease into diverse classes. This classifer analyzed and classified the rice plant infection into three major classes with the help of a single feature. Hence, it is a robust mechanism that consumes the minimum computation time. Under the application of IP schemes, disease prediction is performed automatically in paddy leaves as intended [
Kaya et al. [
In the prediction of rice leaf disease, classical models like human vision-related models are applied. An expert suggestion is required here whereas it is time consuming, costly and possess dense limitations. The correctness of human vision model depend upon the eyesight of the expert. ML-related model is activated to find the classes of diseases, to make proper decisions, and to decide on better remedy. The benefit of using ML models is that it computes consistent tasks compared to experts. Hence, to overcome such limitations of previous technologies, a novel ML-based classifier has to be developed. However, there is a gap exists in both detection and classification due to less development in this ML-based plant leaf disease prediction domain.
In order to accomplish the best classification and segmentation of rice plant images, Contrast Limited Adaptive Histogram Equalization (CLAHE) method is deployed to enhance the density of contrast image. The modified CLAHE is useful in this process to discard noise amplification. Further, various histograms are computed using CLAHE in which the nodes depend upon the standard image region. The histogram gets disseminated to remove the additional amplification whereas the remapping intensity measures are determined through shared histogram. For rice plant images, the newly-developed CLAHE scheme is deployed to enhance the contrast nature of the image. In order to describe the efficacy of CLAHE, few procedures are employed namely entire input production, input pre-processing, processing of background region that intends to develop a mapping to grayscale, and attaining the concluded CLAHE image from random gray level mapping.
Segmentation is defined as the classification of image into a set of non-overlapping parts. Through the identification of optimal threshold values, one can achieve proficient segmentation than the alternate models depending upon the computation of histograms. An image histogram shows reliable frequency for different colors based on its existence such as grayscale level of the input image. Additionally, a digital image can be projected with a value of L colors, while the histogram is assumed as a discrete function as illustrated in
where the count of pixels is
In general, a global threshold is computed effectively when the peaks in image histogram are of interest and the background is categorized. Furthermore, it is apparent that the current method could not be applied for illuminated images. The valleys of histograms are longer and get expanded with diverse measures of heights. Followed by, a local adaptive threshold finds a threshold for each pixel depending upon the range of intensities, attained from local neighborhood. Next, thresholding is carried out for an image in conjunction with a histogram that lacks original peaks. Moreover, specific methods are employed to split the image into sub images for which the mean value is applied. If a specific feature is involved in an image, then prior to this procedure, the operation of picking a threshold becomes simple whereas the root cause for deciding a threshold is to assure the present task is accomplished. A new technology called P-tile applies data in the form of darker objects when compared to another background. It applies a definite fraction such as percentile (1/p) of complete image ration and for printed sheets. Therefore, a threshold is used for examining the intensity target fraction of image pixels than the previous value. Next, the intensity is found under the application of cumulative histogram:
The measure of threshold
The supervised learning model use CNNs with reduced attributes and effective training speed in comparison with deep Artificial Neural Network (ANN) especially with possible advantages like image segmentation, prediction, and classification. A feature map of initial layer is obtained by reducing the input volume after which conv. kernels are applied. Conv. kernel is comprised of
where
where
Conv. layer contains filters which have to be convolved across the width and height of input data. Besides, the simulation of a Conv. layer is attained with the application of dot product between filter weight content and exclusive location of input image. It is considered that 2D activation map offers filter responses in spatial location. The excess variables are number of filters, filter weight, size, stride, padding, and so on. These variables are used in examining the size of outcome. Typically, pooling layers concentrate on removing the overfitting problems whereas non-linear down-sampling is applied on activation maps that decrease both dimension as well as complexity. Consequently, the computation speed gets improved. The attributes applied are filter size and stride, where padding is not essential in pooling. Furthermore, pooling is applied in input channels effectively. Likewise, both output and input channels are similar. It is classified into two classes namely, max pooling and average pooling.
Max pooling: The working principle of this layer is similar to Conv. layer. However, it primarily varies on two aspects i.e., capturing a dot product from input and filter and maximum neighboring value from input image
Average pooling: The values are processed by exclusive position from input image
The Fully Connected (FC) layer is also termed as hidden layer and is used for unwanted NN. Next, the input array is changed as 1D vector with the help of a flattening layer. Here, a node in input is connected to all nodes present in the output.
Softmax layer is applied in the layer consequent to CNN, which implies a categorical distribution between labels and it offers the feasibilities of an input that belongs to a class.
The base model for inceptions is applied in the training of diverse portions as the monotonous block is categorized as subnetworks and is generally utilized in the activation of comprehensive process in memory. However, inception technique is referred to as a simple technique that shows the probability of changing the number of filters accomplished from remarkable layers without influencing the heartiness of complete trained network. To improve the training efficiency, a layer size should be changed into better value so that an appropriate trade-off can be attained from diverse sub-networks. Unlike earlier, tensor flow advanced inception methods are deployed without repeated partitioning. Moreover, inception-v4 is applied to eliminate irregular process, which has developed similar mechanism for inception blocks in each grid size.
Each inception block is used by a filter expansion layer that guides in enhancing the dimension of filter bank prior to input depth estimation. Hence, it is a significant task to replace the dimensional cutback that results in the inception block. Among different techniques of inception, Inception-ResNet V2 has gradual speed since it is presented with massive count of layers. The excess alignment between residual and non-residual inception is termed as Batch-Normalization (BN) and is applied in the classical layer rather than being used on residual estimation. Therefore, it is assumed as a logical concept at the time of anticipating a large-scale BN. A BN of TensorFlow applies massive amount of memory and to reduce the number of layers, BN is employed under diverse positions.
Here, it is marked that a filter score can be found. This phenomenon depicts an unreliable system which gets expired during primary phase of training implying that the target layer, prior to initializing a pooling layer, creates the zeros from distinct process. But it is not possible to eliminate the training score. Further, the measures are reduced prior to identification of better learning. Usually, scaling factors exist within the radius of 0.1 to 0.3 and are applied for scaling accumulated layer activations correspondingly.
ELM is utilized in the classification of balanced dataset whereas WELM is utilized in the classification of imbalanced datasets. So, this section briefly establishes WELM. During training, dataset has
where
where
Based on Karush-Kuhn-Tucker hypothesis, Lagrangian factor is established to alter the training of ELM into a dual problem. A resultant weight
where
where
It is from
So, KELM classification performance is decided by two parameters such as kernel function parameter
To maintain the original benefits of ELM, it is presented that WELM allocates weights to different instances so that imbalanced classification problems can be resolved. Its resultant function is computed as given herewith.
where W is a weight matrix. WELM consists of two weighting systems as given herewith.
where
Regularization coefficient C and kernel
Flower constancy is defined as the exact solution. For global pollination, a pollinator sends pollen from long distances to higher fitting. Alternatively, local pollination is processed in a tiny region of a flower in shading water. Global pollination is carried out under the feasibility that is termed as ‘switch probability.’ Once the above process is removed, local pollination is deployed. In FPA model, around four rules are present as given herewith.
Live pollination and cross-pollination are named as global pollination for which the carriers of pollen pollinator use levy fight algorithm
Abiotic and self-pollination are considered to be local pollination
Pollinators are nothing but the insects which can develop flower constancy. It is depicted as a production possibility for the flowers.
The communication between global and local pollination is balanced using switch possibility.
Therefore, 1st and 3rd rules are illustrated as follows.
where
where
where
The performance of the CNNIR-OWELM model was validated using a benchmark rice leaf disease dataset [
Classes | Number of images |
---|---|
Bacterial leaf blight | 40 |
Brown spot | 37 |
Leaf smut | 38 |
Different Classes | Bacterial leaf blight | Brown spot | Leaf smut |
---|---|---|---|
TP | 38 | 32 | 36 |
TN | 72 | 73 | 74 |
FP | 2 | 5 | 2 |
FN | 3 | 5 | 3 |
Total images | 115 | 115 | 115 |
Measures | Sensitivity | Specificity | Precision | Accuracy | F-score |
---|---|---|---|---|---|
Bacterial leaf blight | 0.927 | 0.973 | 0.950 | 0.957 | 0.938 |
Brown spot | 0.865 | 0.936 | 0.865 | 0.913 | 0.865 |
Leaf smut | 0.923 | 0.974 | 0.947 | 0.957 | 0.935 |
Average | 0.905 | 0.961 | 0.921 | 0.942 | 0.913 |
Methods | Sensitivity | Specificity | Precision | Accuracy | F-score |
---|---|---|---|---|---|
CNNIR-OWELM | 0.905 | 0.961 | 0.921 | 0.942 | 0.913 |
VGG-16 CNN (2020) | 0.898 | 0.932 | 0.902 | 0.929 | 0.899 |
DNN (2019) | 0.735 | 0.894 | 0.749 | 0.900 | 0.815 |
DAE (2019) | 0.680 | 0.872 | 0.676 | 0.860 | 0.770 |
ANN (2019) | 0.633 | 0.816 | 0.609 | 0.800 | 0.683 |
CNN (2019) | 0.940 | 0.940 | 0.940 | 0.938 | 0.940 |
KNN (2019) | 0.650 | 0.780 | 0.720 | 0.700 | 0.650 |
SIFT-SVM (2016) | 0.867 | 0.832 | 0.867 | 0.911 | 0.867 |
SIFT-KNN (2016) | 0.900 | 0.864 | 0.906 | 0.933 | 0.901 |
This research work presented a novel CNNIR-OWELM model for rice plant disease diagnosis in smart farming environment. The proposed method is executed at the server side and incorporates a set of processes such as image collection, preprocessing, segmentation, feature extraction, and finally classification. Primarily, the IoT devices capture the rice plant images from farming region and transmit it to the cloud servers for processing. The input images are preprocessed to improve the contrast level of the image. In addition, the histogram-based segmentation approach is introduced to detect the diseased portions in the image. Followed by, the feature vectors of the segmented image are extracted by Inception with ResNet v2 model. Subsequently, the extracted feature vectors are fed into FPA-WELM model for the classification of rice plant diseases. The experimental outcome of the presented model was validated against a benchmark image dataset and the results were compared with one another. The simulation results infer that the presented model effectively diagnosed the disease with a higher sensitivity of 0.905, specificity of 0.961, and accuracy of 0.942. In the future, the researchers can extend the presented model to diagnose diseases that commonly affect the fruit plants, apart from rice plants.