Enhanced Hybrid Neural Networks (CoAtNet) for Paddy Crops Disease Detection and Classification

1.1 Different types of rice plant diseases

Rice plants can be infected with many different types of diseases. An infection of rice plants was caused by a different kind of infection than that which affected other plants [4]. Brown Spot: The main field diseases in the paddy crop were shown in Figure 1. In addition to brown spots, sesame leaf spots and Cochliobolus miyabeanus are also considered fungal diseases. The majority of cases occur in West Bengal, Orissa, Andhra Pradesh, and Tamil Nadu [5] Symptoms: Affects nurseries and main fields alike, Seedlings are affected by blight, the patches correlate with each other, measuring between 0.5 and 2.0mm in breadth, Brown-colored infection also appears on the neck of the panicle, in severe cases, yields are reduced by 50%.

1.png

Figure 1. Samples of crops leaf diseases unhealthy

a. Cochliobolus miyabeanus,

b. Xanthomonas oryzae pv. Oryzae,

c. Magnaporthe oryzae,

d. Rice tungro virus

Rice Blast: Magnaporthe oryzae was responsible for RB disease. Leaf spots with dark green borders on the grayish-green circulars are symptoms of this disease. A spindle-shaped or ellipsoidal lesion was usually observed. Various countries around the world are experiencing some of the most devastating rice diseases [6]. Bacterial Leaf Blight (BLB): A bacterial pathogen called Xanthomonas oryzae is responsible for BLB. This disease was characterized by a lesion that begins at the tip of the leaf and spreads down several inches. Rice production is significantly affected by BLB, which is one of the most destructive diseases. Sheath Blight (SB): Rhizoctonia solani is responsible for causing Sheath Blast disease. Plants show symptoms from the lower sheath, progressing up to the upper sheath and then to the leaves. Several major rice-growing countries have faced a severe economic impact from this disease [7, 8].

Problem definition: A human vision-based approach was typically used for detecting leaf diseases [9]. It was time-consuming and very expensive to seek expert opinion in these situations. Methods that rely on human vision have many limitations. An expert's eyesight was crucial to the accuracy and precision of a human vision method. It was possible to classify the various different types of diseases, make the appropriate decisions, and choose the appropriate treatment using a deep learning method. Its consistency as compared to human experts has the advantage of using deep learning technology. Therefore, a neural network based classification approach was necessary to overcome the drawbacks of conventional methods. Using deep learning to detect leaf diseases on plants was a very rare development, and it was only a few paddy leaf disease detection and classification that recent developments have been made.

Insects attack the rice plant in more than 100 species, 20 of which can cause economic damage. There are various stages of attack in the rice crop. Generally, there are the Vegetative stage, Reproductive stage, Ripening stage, and Soil-inhabiting pests. A stem borer is considered to be one of the most harmful pests of rice in the world and attacks plants from seedlings to maturity. Scirpophaga incertulas and Chilo suppressalis cause steady annual crop losses of 5-10% in Asia, with localized outbreaks of up to 60% at times. Mostly found in the tropics, the scirpophaga incertulas insect is also found in temperate areas with annual rainfall exceeding 1,000 millimeters and tem temperatures above 10℃. The species is predominant in Bangladesh, India, and many other countries. During earlier years, they were monitored and controlled using light traps because of their strong phototaxis. As stem borers are the most serious pests in rice cultures worldwide, they attack young plants up to maturity. These insects are very danger for the rice plants, which it will affect the maximum productivity. In deep learning model the insects captured image were trained to predict which type of diseases attacked by rice plants.

1.2 Motivation

The design paradigm of the proposed design is directed towards finding the image in the right class, disease classification, and detection. I have taken the data set from the farmer land following location such as Villupuram district longitude and latitude 11.9401°N, 79.4861°E. The preprocessing process completed convolution auto encoder take a place. The results are divided into two different neuron networks such as feature extraction using ResNet 152 and feature extraction using proposed CoAtNet. The selective attention is provided to the outlier attribute extraction as a size, color, resolution, and shape. The different diseases are arranged into different classes to train the model. The main goal is designed to develop a neural network model to identify the diseased plant leaf in an earlier stage to increase productivity. By adopting a neuroscience perspective to classify distinct plant leaves based on their infection classes.

In this research article the complete layout is initial dataset collected from farmer land, then preprocessing the loaded images, select the CNN model for training purpose to achieve best results. Hybrid model neural networks achieve the maximum best results from after validation process.

3.1 Acquisition of images

Image acquisition was the procedure for gathering images that will be utilized in this study. The input images dataset of the paddy plant leaf were take snapshot on the farmer field using a high and low-resolution using digital camera and mobile phone. In this study, we have captured 3071 leaf images for the proposed system. Dataset link (figshare - My data) Rice plants suffer from diseases and pests in various parts [17]. Several factors may contribute to their occurrence like sunny, moisture, cloudy, rainfall, monsoon, nutrition, plant leaf, etc. A range of weather conditions was used for image collection - in winter, summer, and overcast conditions to ensure that a fully representative set of images was obtained. In these images, the total training purpose we have considered 2,771 and the test purpose we have used around 300 images. Four categories are limited to the most common disease categories: BLB, Leaf Blast, Brown Spot, Tungro / Leaf smut.

3.2 Pre-processing

In an input dataset images are resized and cropped in pre-processing to a dimension of 300 X 450 pixels to minimize the memory and compute requirements. In the current phase, one of the biggest concerns was to eradicate the image setting by fusing hue values. An RGB image was first transformed into an HSV image. With the HSV representation the value S was made the first consideration because it predominates over the whiteness. We select a threshold value by experimenting with several values. Image with background removed only shows the diseased portion of the leaf. Non-linear filtering approaches were applied to remove the unwanted noise from the images.

3.3 Convolutional Neural Network (CNN) models

CNN is an image classification defined as the work of converting various different images into one or more predefined classes. The convolutional networks model was a dedicated kind of neural network for dealing out suitable for 2D information that has a great grid-like topology [10]. It was designed for the functionality of human neuron behavior. The configuration of the CNN has own image filter influence on the performance of the model without human intervention. Convolution was a specialized kind of linear operation. CNN have 3 to 150 layers to communicate each another layers such as output acts as another layer input.

In Figure 2 list the evolution of the different deep learning model in every year-wise described. To increase the precision, complication, and real-world impact deep learning has every time improved in its ability to afford accurate recognition and prediction [18]. We have worked to full fill the farmer challenges and created opportunities to get better deep learning even further and to bring it to new frontiers [19].

2.png

Figure 2. Evolution of deep learning model year-wise proposed

There are two networks structures were used in our research work, they are ResNet-152, and CoAtNet, and are described as follows.

3.4 Architecture RESNET-152

In order to overcome the “vanishing gradient” problem, ResNet is a Convolutional Neural Network (CNN) architecture that can construct networks up to thousands of layers, which outperform shallower networks. ResNet-152 was deep CNN based on the constructs of pyramidal cells in the cerebral cortex [19]. One of the most powerful deep neural networks, ResNet (residual connections), achieved outstanding results in the ILSVRC 2015 classification challenge [20]. The network consists of different unique layers such as input layer, convolution, pooling layer, rectified linear unit layer, and classification layer. ResNet models typically shown in Figure 3 have two or three layer skips, containing rectified linear unit (ReLU) and batch normalization. In general, skip connections are added to avoid the vanishing gradients problem or to moderate the accuracy saturation difficulty. The feature space of a neural network without residual parts is larger. Consequently, it was more susceptible to perturbations that cause it to leave the manifold and requires more training data to recover [21]. The forward propagation is an output of the neurons in the layer defined as $a^l$, the activation function for the layer mentioned as g, $\mathrm{W}^{\mathrm{l}-2, \mathrm{l}}$  total weight matrix lies l-1 & l.

$a^l=g\left(W^{l-1,l} * a^{l-1}+b^l+W^{l-2,l} * a^{l-2}\right)$          (1)

$a^l=g\left(Z^l+W^{l-2,l} * a^{l-2}\right)$          (1.1)

The backward propagation is defined in equation number 2 as follows:

$\Delta \omega^{l-1, l}=-\eta \frac{\partial \mathrm{E}^{\mathrm{l}}}{\partial \omega^{l-2^2}}=-\eta \mathrm{a}^{\mathrm{l}-1} * \delta^{\mathrm{l}}$           (2)

η is a learning rate less than 0, $\delta^\iota$ an error signal of neurons at layer ι and $a^l$ activation neurons.

ResNet was a network model that uses residual blocks to enhance the depth of the CNN. Initially Resnet architecture has been implemented with minimum number of layers, which obtained good results up to thousand images. Hence number of convolutional and pooling layer is increased. Finally, multiple separate units are connected through fully connected layers. Then the given image is classified as normal leaf or diseases by using softmax classifier.

3.png

Figure 3. Block diagram of Resnet-152 architecture for classification

3.5 Proposed Hybrid CoAtNet Architecture (PHCNA)

A convolutional layer does not connect neurons to every pixel in the input image, but only to pixels in their receptive field. In the training process, learnable filters or kernels are convoluted over the image. A filter learns to recognize specific patterns, and low-level filters are related to more complex patterns. The Google AI team introduced a hybrid model, which combines self-attention and convolution to achieve top accuracy levels [22]. The CoAtNets pronounced "coat" nets, combining the strength from different architecture. There are two key insights provided to perform in a better way: i) Convolution and self-attention can be naturally integrated through simple relative attention. ii) Attention layers that consider their competence and computational requirements at each phase (resolution) effective at were improving simplification, capacity, and efficiency [23]. What is the best way to mix CNN with transformers? The basic idea was MBConv block, which combines depthwise convolution with inverted residual bottleneck, initially because this expansion-compression scheme is the same as the Transformer's FFN module. The depthwise convolution in addition to self-attention can both be expressed as a per-dimension weighted sum in a predefined receptive field.

$\mathcal{Y}_i=\sum_{\mathcal{J} \in \mathcal{L}(i)}^n \mathcal{W}_{i-j} \odot x_j$           (3)

The depth wise convolution was defined as $y_i$ is the output of provided above expression. The sum of value in a pre-defined local receptive field $\mathcal{J} \in \mathcal{L}(i)$, here convolution relies on a fixed kernel to gather information from a local receptive.

$\mathcal{Y}_i=\sum_{\mathcal{J} \in \mathcal{G}}^n\left(\frac{\exp \left(x_i^{\mathcal{T}} x_j\right)}{\sum_{\mathcal{K} \in \mathcal{G}} \exp \left(x_i^{\mathcal{T}} x_k\right)}\right) x_j$           (4)

Self-attention allows the receptive field to be global and instead of fixed kernels, it utilizes dynamic attention weights that vary based on the inputs. To mix convolution and self-attention, we can simply add a global static convolution kernel with an adaptive attention matrix $\exp \left(x_i^{\mathcal{T}} x_j\right)$ the softmax normalization. The attention weights are decided by a combination of a CNN kernel and an adaptive matrix [24, 25]. This Eq. (4) interestingly corresponds to an especial variant of attention called relative self-attention, in which relative position representations are used.

$\mathcal{Y}_i=\sum_{j \in \mathcal{G}}^n\left(\frac{\exp \left(\mathcal{X}_i^{\mathcal{T}} \mathcal{X}_j+\mathcal{W}_{i-j}\right)}{\sum_{\mathcal{K} \in \mathcal{G}} \exp \left(\mathcal{X}_i^{\mathcal{T}} \mathcal{X}_k+\mathcal{W}_{i-\mathrm{k}}\right)}\right) \mathcal{X}_j$          (5)

The global static convolution matrix was derived from Eq. (5) $\mathcal{W}_{i-\mathrm{k}}$ & $\mathcal{W}_{i-j}$.

Figure 4 shows the complete overall architecture diagram for CoAtNet, considering above block diagram with size of HxW as an input, we pertain convolutions in the initial stem phase (S0) and shrink the size to H/2 x W/2 [23]. In every stage, it continues to reduce the size. The Ln represents the total amount of layers. Premature stages (stage 1 and stage 2) use MBConv blocks consisting of depthwise convolution. At the end of two stages (stage 3 and stage 4) transformer blocks are mostly used with comparative self-attention.

4.png

Figure 4. Proposed Hybrid CoAtNet Architecture (PHCNA)

Instead of using the earlier transformer blocks in ViT, here we make use of a pooling method between different stages, similar to the channel transformer. Transformers may lack the generalization capability that CNNs possess, and rely on a substantial amount of data to compensate. As a final step, we implement a classification to generate class predictions [26-28]. The proposed workflows were shown in Figure. 5 to identify the healthy and diseased paddy leaves. In this research work, a new hybrid model was designed to detect the rice leaf images. There are four different categories of leaf diseases provided as input. After that, all the RGB images are resized as per the requirements. The Convolutional Auto Encoder (CAE) was used for filter purposes. The CAE has been dropping the dimensionality of the input rice leaf images. We have used two different network models such as ResNet-152, CoAtNet to train the model. We have divided it into 80:20 ratios for the dataset training and testing purpose. CNN helps to predict, rice leaf images have been classified into two different categories such as healthy or unhealthy. The model inference was detected as which category of the diseases as well calculate the various measures.

5.png

Figure 5. Proposed workflow classification for identifying healthy and diseased paddy leaf

Table 2. Images classified into four categories and respective samples count

In Table 2 training parameters were plotted and four different classes such as C_A, C_B, C_C, and C_D contain each diseases name along with a total count of samples.

Our research will be based on providing high-end computing services to farmers in order to achieve high levels of productivity. The population will grow in the future, and we will need to meet the basic need for food. Plants may be harmed by various insects as a result of dynamic climate change. Early disease detection using low-resolution images, high-resolution images, videos, and data captured by IoT devices (Drones). Disruptive technology will aid in the provision of solutions to farmers. Practically speaking, the majority of farmers are unaware of how to use this high-end application. We can carry out the suggested work after gaining government consent and for the bulk of the agricultural location. For the time being I have identified several diseases using deep learning. The major diseases are allowed by the farmer at the beginning stage, to prevent these diseases in detail manner discussed. There are four different types of diseases were considered. There are various different types of rice varieties available in the market. Due to various seasonal the rice cultivation process may vary place to place. In this article, we have obtained the accuracy level 96.5% compared to all other values it is high efficiency. We have used two different network model such as ResNet-152 and CoAtNet model as proposed one. We have used to perform various rigorous analyses of results. In comparison to existing work not achieved by above 95%.