1 Introduction

The Rice plant has grown to be an enormously important food crop for society, particularly for those living in Asia, where rice is a staple food for Sri Lankans. As per polls, approximately there are 3 billion people depend on rice on a daily basis. Paddy farming has a long-rooted history within eastern civilizations. As an Asian country, Sri Lanka has its own ancient rice varieties including Ma wee, Pachcha perumal, Suwadal, Kurulu thuda, Masuran, Kahawanu, Kalu heenati [1]. In Sri Lanka, Paddy Cultivating process is done without considering the locations as well as it continues rice production under various weather conditions from the wettest areas to the driest areas in a tropical country like Sri Lanka. Due to the current state of disease-related concerns, several diseases do exist that have a direct impact on the nation’s rice production. According to past studies fungi’s, viruses and bacteria related diseases has primarily caused to this reduction in annual rice production. Around 37% of annual rice production tend to become decreased due these rice plant diseases [1]. Still, there are a considerable amount of people involving rice cultivation as a life strategy same as in the past, thus the majority of those people have a proper understanding of paddy cultivation and may not have needed to be taught how to identify these diseases. However, the current scenario is a little different; individuals from younger generations are less aware of such disorders, thus the government is providing agricultural instructors with technical expertise about those problems. It can be a challenging chore at times since, because there are so many diseases, even the instructors cannot determine the particular disease without visiting the fields. When considering all these things you may understand why should we need to find some efficient method to control this issue.

Farmers may use advancements in technology to diagnose rice diseases, which would be a highly effective solution to these problems. Many studies are now being undertaken on the use of Deep Learning technologies to diagnose such disorders. Several studies employed deep learning concepts and approaches, while others built their own techniques to handle this issue. This research discusses how deep learning models may be employed as pre-plant disease detection techniques. The aim of conducting this study is to examine various methodologies by evaluating the merits and weaknesses of those studies and analyzing them. There have been numerous studies on detecting illnesses in other plants such as tomatoes and peach, but there have been few studies on disease diagnosis in rice plants. However, it can be shown that the technique employed to diagnose illnesses in those all-important plants is the same. According to the author’s best of knowledge, they have identified various rice leaf diseases, among those diseases some already widespread in some other countries. But some of those diseases are unique to Sri Lanka “Pecky rice”, “Grain spotting”, “Brown spot”, “Leaf scald”, “Root-knot”, “Narrow Brown Leaf spot”, “Bacterial blight”, “False smut”, “Sheath rot”, “Bacterial leaf streak”, “Rice blast”, “Rice sheath blight” [1] are some common types of rice plant diseases (see Table 1).

Authors have already secured a publication titled “Plant Disease Diagnosis and Controlling using Convolutional Neural Networks” [2] and in that study was mainly focused to review related studies in the literature which were conducted about common plant disease detection and control. Through the extended study, authors have observed that even though there were ample of studies that have been conducted regarding disease diagnosis of common plants such as tomato, pepper, potato, etc. there is very few studies has conducted regarding rice plant disease diagnosing and controlling. From this study, authors have reviewed a number of studies that were conducted regarding rice plant disease diagnosing and controlling to identify existing better-performing technologies and models in rice plant disease diagnosing with the intent of implementing a better model for detecting rice plant diseases as the further work. A summary of the findings have been included in Table 2 and they are further been described in the discussion section. Table 3 will reflex the merits and demerits of those reviewed papers in detailed.

When it comes to the structure of the manuscript, this section provides a brief introduction to the study conducted by the authors. The next section discussed the deep learning technologies that have been used in the studies conducted in the literature. The third and fourth sections briefly describe and compare the co-findings of the study that has been conducted. Under the discussion section, it further analyzes and discusses the findings. Under the final section, it is concluded with the best suited technologies and models which will be suitable to build a better and more accurate rice plant disease diagnosing model.

Table 1 Types of rice diseases. Source: [1]
Full size table

2 Deep learning for rice plant disease detection

Deep Learning was created as the assistance of neurons to act or think like a person. DL was created utilizing a multi-neural network design with numerous convolution layers to forecast the needed outcomes. DL includes several network architectures ranging from simplest to more complex structures. Deep Learning originated as a subclass of Machine Learning (ML) in the early 1940s, parallel with the invention of “threshold logic”. It has been used to create computer simulations which have been closely resembled human biological pathways. Figure 1 demonstrates the general flow of image processing techniques used in rice plant disease detection by state of the art researches.

Fig. 1
figure 1

Image processing pipeline. Source: [3]

Full size image

2.1 CNN for image processing

Convolutional neural networks are widely recognized as a unique and necessary core of deep learning due to their excellent feature extraction capabilities. Convolutional autoencoders and CNNs are two of the main Deep learning methodologies that have been employed in a variety of Computer Science applications due to their success in image data processing. Convolution operations are used by these two primary technologies to extract spatial and temporal information from picture data. As presented in Fig. 2, in CNNs, there are common Layers which are known as input layer, convolutional layer, pooling layer etc. The input layer can be considered as the input of the entire CNN in the image recognition process, and it commonly reflects the picture’s pixel matrix. In contrast, the Convolutional layer is utilized to extract picture characteristics. Normally, layer which is known as pooling is used to minimize the size of the feature maps. As a result, it decreases the number of parameters to learn as well as the amount of processing power done in the network. Whereas, all inputs from one phase are linked to every activation unit in the layer below by Fully linked layers. The activation function is included in the Softmax layer of NN models that predict a multinomial probability distribution. It may contain multiple pooling and activation layers inside hidden layers of Convolutional Neural Networks and there are numerous activation and pooling layers. Fully Connected Levels are the network’s ultimate layers.

Fig. 2
figure 2

Basic diagram for CNN. Source: [4]

Full size image

2.2 Feed forward neural network (FFNN)

Fig. 3
figure 3

Feed forward network. Source: [5]

Full size image

Feed forward neural network process information sequentially from input to output and there is no feedback between layers as demonstrated in Fig. 3. If the user of a Neural Network is primarily concerned with the input or the output layers, then these Feed Forward Neural Networks are extremely important, and they are frequently employed in non parametric data analysis. R, N, and S are the inputs, neurons in the buried layer and generates numbers accordingly. Networks input vector is denoted by x. The input and hidden values matrices are denoted by iw and hw, accordingly. ho is the hidden layer’s output vector and y is the network’s output vector. The bias vectors of the output neuron are denoted by hb and ob, accordingly.

2.3 Single layer perception (SLP)

Fig. 4
figure 4

Single layer perception. Source: [6]

Full size image

Single Layer Perceptron, as shown in Fig. 4, is kind of a model which transforms a linear into a nonlinear function with the help of activation function. According to the Single Layer Perception, vectors can be allocated to one of two classes.

2.4 Multi layer perception (MLP)

Fig. 5
figure 5

Multi layer perception. Source: [7]

Full size image

This is a kind of perception in a conventional network that comprises only one hidden layer, as seen in Fig. 5. Multi Layer Perception is a kind of perception with several layers that comprise of output, hidden and input layers. Input layer of the ANN is also called as the initial passive layer of an ANN, acts as a data intake conduit. The hidden layer in the second layer of this MLP increases the network’s capabilities and enables for the modeling of more challenging issues. The last layer is referred to as the output layer, and it is responsible for generating network output signals.

2.5 Honen’s self organizing map (SOM)

Fig. 6
figure 6

SOM. Source: [8]

Full size image

SOM, which is illustrated as in Fig. 6, can be defined as unsupervised learning network which has a design essentially based on the Feed Forward network, which, it has two levels, those are known as the Kohonen layer and the input layer. A Self-Organizing Map network’s neurons are organized as a grid, it can be either rectangularly or hexagonally. The input layer is connected to the Kohonen layer. Maps are usually created in the input space of the network.

2.6 Radial-basis function network (RBFN)

Fig. 7
figure 7

RBF network. Source: [8]

Full size image

The RBF network seen in Fig. 7 is composed of three layers which are known as input and output layers, hidden layer of RBF neurons. In data modeling, hidden layer of this network can be introduced as crucial. When observing the Fig. 7 x, y(x), ci, and M signify input, output, center, width, and number of basis functions centered at ci, respectively, while wi specifies weights.

2.7 Probabilistic neural network (PNN)

Fig. 8
figure 8

Probabilistic neural network diagram. Source: [9]

Full size image

A general feed forward network is a CNN that contains input, output and hidden layers is referred to as a Probabilistic Neural Network. The hidden layer also known as pattern layer. A sample network diagram can be visualized as in Fig. 8. A Bayesian classifier uses in this PNN. The PNN incorporates a non parametric estimation method to obtain multi-variate possibility and density estimation. PNN is currently the finest neural network for dealing with classification problems. According to the diagram, the output layer has M classes, and each class m includes N m hidden pattern neurons in the pattern layer and In the summation layer there is only one G m summation neuron. Training patterns are imported into the pattern layer and separated into M groups as it get one for each class.

3 Research findings

This section compares and contrasts the previous studies that were reviewed. A comprehensive analysis under selected parameters are evaluated in Table 2, while the merits and demerits of these analyzed studies and techniques have been discussed in Table 3.

When observing past researches on this technical arena, it was able to realize that the three primary procedures in plant disease detection which are known as feature extraction, classification and segmentation. Among these procedures k mean segmentation approach is commonly used for image segmentation. The GLCM method is utilized for extracting features, and a categorization strategy is used to forecast illness names, which are both classic machine learning methodologies. When it comes to rice plant disease detection there were some advanced machine learning approaches has found such as RasNet, DensNet, MobileNet, VGG16 etc. This section focuses on discussing about technologies that were used in the literature by contrasting each of them, following Table 2 will give brief idea about these advanced machine learning technologies used in each study.

Table 2 Comparing previous studies
Full size table
Table 3 Comparison of the drawbacks of the previous studies
Full size table

4 Comparison of datasets used

ML is heavily dependent on datasets. This is the most significant feature that allows algorithmic training and explains why ML has grown in popularity in recent years. According to many researches, size of the dataset will directly be affecting to the accuracy of the developing models. Although data can take various forms, machine learning models rely on basic data kinds; Categorical data, Numerical data, Text data and Time-series data. When you are searching datasets of rice plant diseases you may realize that there are lack of datasets available in the literature, therefore researches has spent more time to gather required datasets to build their models. Some of them has used existing datasets while some of them use their own datasets.

4.1 Newly collected datasets

Researches who have used their own datasets to train and test their models have used devices like mobile phones, cameras to collect those images. In studies [9, 10] they have used these kind of methods to collect their datasets (refer Table 2). In the study which has been conducted by Ruoling et al. [10], he has used a dataset which contains images taken by mobile phone with high resolution. According to the study using this mobile device he could able to collect about 33,026 images which are related to six types of rice diseases during two year period of time [10]. In the study [12] which was conducted by Anandhan et al. they have use a dataset of 1500 paddy leaf Images collected through a Sony camera and using a Mobile device called Vivo V9 [12]. Through collecting the images for the dataset by their own, researches can increase the accuracy of their implementation, but the problem is according to the above reviewed studies collecting data to the dataset and create their own dataset is not that much easy process it takes lot of time. But they don’t want to limit their development only for available dataset, they can expand the model if they have the ability to collect more verities or more classes of data.

4.2 Existing datasets

Some researchers have used existing datasets in order to train and test their models. In studies [12,13,14,15] they have used these kind of methods to collect their datasets (refer Table 2). In some studies they have used modified datasets which are created by merging two or more existing datasets. As an example, in study [14] which was conducted by Surya Pratap et al. they have used a combination of two different datasets which are having 2092 and 120 images of rice plant diseases [14]. Kaggle is a very important web resource which contains thousands of datasets which can be used to train deep learning models. Kaggle allows users to search for and upload data sets, analyze and build models in an internet-based data science environment, cooperate with other data scientists, including machine learning specialists, and compete to solve data science issues. It is a community network for attracting, nurturing, training, and challenging computer scientists from everywhere in the world to tackle data science, deep learning, and advanced analytic challenges. It has around 536,000 active users from 194 countries and receives over 150,000 contributions every month. Yibin et al. and Andrianto et al. has used Kaggle datasets which contains 5200 images [13] and 1600 images [15] respectively for the studies [12, 14] even though these datasets makes researches dataset finding process easy as using the same datasets in many studies sometimes on the same models may not produce innovate outcome to the research world some of the researches like Yibin et al. [11] used some other public datasets which are available on the internet for training and testing purposes of their studies in that study they have used about 2370 rice leaf samples in total, even though the data collecting process is easy. Accuracy of collected data may be low as they were taken from untrustworthy resources. So, the accuracy of collected data sets may directly affected to the accuracy of the model.

5 Discussion

Many ML techniques have been implemented with the intent of recognizing and classifying various diseases of plants. Moreover, as Machine Learning and Deep Learning technologies progress, this area of study seems to consist significant capacity for enhanced accuracy. To recognize and characterize the symptoms of Rice plant ailments, many original and customized deep learning structures, as well as many visualization methodologies, have been explored. The deep learning models used to diagnose rice plant illnesses are described in depth in this research. The research examines updated and well known deep learning architectures, along with visualization mapping methodologies for detecting plant diseases. It makes suggestions for future improvements in detection and visualization.

Observing past studies, several researchers employed hybrid models to meet their study goals. In the study of “Automatic Diagnosis of Rice Diseases Using Deep Learning” [10] which was conducted by R. Deng et al. has used a hybrid model which contains three sub models, they are ResNeSt 50, DenseNet 121 and SE ResNet-50. Here they created an Android application. It was created with the use of deep learning methods and for the application, they employed a huge dataset including around 33,026 photos of rice illnesses known as sheath. The Ensemble Model was tested using a multiple collection of photos, proving the model’s efficiency. The author employed a set of characteristics for the validation procedure, such as recall accuracy, learning rate, illness recognition accuracy, and precision accuracy. They attained 91 percent accuracy with this model. According to this implementation, ResNeSt 50 is very accurate model which has reached better accuracies, modified version of this model has used by Krishnamoorthy et al. In the study of “Rice leaf diseases prediction using deep neural networks with transfer learning ” [13], they have developed deep learning model InceptionResNetV2. This is a CNN model that uses a transfer learning strategy to detect illnesses in rice leaf pictures. The suggested model’s parameters have been tuned for the classification job, and it has an excellent accuracy of 95.67 percent. To diagnosis plant diseases, several research employ regional convolution NNs and quicker regional convolution neural networks. This sort of research was carried out by Assistant Professor of Galgotias University whose name was Anandhan. When analyzing the research [10], DL models were employed to identify the early stages of a rice blast illness such as blast disease, leaf streak, sheath blight, and brown spot. According to the findings of the experiments, R-CNN model is the most appropriate model for detecting or recognizing several rice blast illnesses such as Sheath blight-94.5%, Brown spot-95%, and Blast-96%, with the intent of identifying various leaf diseases, they developed a unique picture dataset collected from local village rice plants. The suggested deep learning models offer the best results when using mask R-CNN and R-CNN. This strategy will aid farmers in the healthy and safe prevention of rice plant illnesses.

According to the previous studies, which has been conducted by researchers, VGG16 and VGG19 models has performed very well with better accuracies in training, testing and validating the data. Simonyan and Zesserman suggested VGG19, is an architecture with two distinct layer variants, 16 layers one is known as VGG16 and 19 layers one is known as VGG19. VGG had taken first and second place in the localization and classification tracks, respectively. VGG19’s structure is completed by three fully linked layers and five blocks of convolutional layers. Surya et al. has conducted a research on detecting rice plant diseases using mobile application in the study of “Rice Plant Infection Recognition using Deep Neural Network Systems” [14], after the trial was successfully finished, it was discovered that the accuracy of the LeNet5, VGG-19, and MobileNet-V2 was 76.63 percent, 77.09 percent and 76.92 percent, respectively. According to these details developer has chosen VGG19 model for the development. VGG16 model has used by another Indonesian researcher whose name is Andrianto has conducted his study on implementing a “Smartphone Application for Deep Learning Based Rice Plant Disease Detection ” [15], according to this paper, they describe that they have already developed a deep learning-based rice disease detection system, which consists of a machine learning model on a distant server and a mobile application. The train accuracy of the rice leaf disease detection system utilizing the VGG16 model is 100%, while the test accuracy is 60%. The author of this study claims that the “Hispa” class is the most accurate when compared to other classes since it has the most data from the test results. That proves us using large number of data for the training purposes will also increase the accuracy of the model.

When referring to previous review papers, it can be understood that different types of technologies have been used by them which are varying from simplest to complex. Some researchers have utilized existing machine learning algorithms, whilst some have gone through their own methods and algorithms in order to recognize images. And, some of the algorithms they have used has been mentioned in the above Table 2. According to the paper “Rice leaf diseases prediction using deep neural networks with transfer learning.” [13], which has written by Krishnamoorthy et al., CNN is also an algorithm of deep learning techniques that has been successfully invoked for handling computer vision issues such as picture classification, object segmentation, and image analysis. According to the author, CNN is efficient in detecting visual representations. It is kind of a feed forward ANN composed of three separate layers as input, output and hidden. Transfer learning is a technique for repurposing a previously trained CNN for a new issue and in this paper author has discussed about the potent algorithm. According to the study it is a powerful deep learning algorithm that has been introduced into the world of agriculture to address many challenges such as weed and seed identification, plant disease categorization, root segmentation and fruit counting.

When observing the previous studies, some researchers have utilized existing deep learning models, whilst some have gone through developing their own deep learning models. “Attention based Depth wise Separable Neural Network with Bayesian Optimization” or in short term ADSNN-BO is that kind of innovative model which has introduced by Yibin through his study on “Rice Diseases Detection and Classification Using Attention Based Neural Network and Bayesian Optimization ” [11], where they have suggested to identify and diagnose rice illness using photographs of rice leaves. The author presented the ADSNN-BO model, which is based on the MobileNet structure and an enhanced augmented attention mechanism, to accomplish AI-assisted speedy and accurate illness diagnosis. As the name contains the term mobile, MobileNet is indeed a CNN architecture which specializes in categorizing Mobile Versions and pictures. In depth convolutions, which filter the inputs without creating new features, and point to point convolutions. The Bayesian optimization approach is used to optimize the model’s hyperparameters. On the basis of a public rice illness dataset with 4 groups, cross verified classification studies are carried out. The experimental findings show that this mobile compatible ADSNN-BO model surpasses all of the previous models examined, with a test accuracy of 94.65 percent. Author has stated that the findings of this study will encourage the use of artificial intelligence in the agricultural area for rapid plant disease detection and control. To improve the performance of all CNN designs, pretrained ImageNet weights are employed. As a further step, they intend to investigate their suggested ADSNN-OB model further using various optimization approaches and hyperparameter optimization to improve the efficiency and make the procedure more successful. Hence, In summary, this study examined standard and customized Deep Learning structures, as well as regular mapping approaches, that are employed in rice plants disease detection. There are very few number of studies has been conducted regarding rice plant disease detection. Through this study it encourage to build software which are capable of effectively identifying rice plant diseases. And it focuses on identifying best machine learning technologies used in the past studies for Rice plant disease detection. When discussing about the limitations of this study it mainly focuses on rice plant related diseases not about the all plant diseases. It could be further expand to discuss about more plant diseases in the future. When it comes to the future scope this study has conducted with the aim of identifying best Machine learning technologies used in the previous studies in order to develop an new ensemble model which is more accurate than the existing models in rice plant disease identification. It would be very beneficial for increase annual rice production by effectively identifying rice plant diseases.

6 Conclusion

CNN are widely regarded as the most successful technique for any prediction problem involving input image data. It is critical since it needs very little pre-processing. This work examines how DL aids in the diagnosis and categorization of rice plant diseases based on past research. Because image detection comprises numerous phases such as classification, recognition, segmentation and detection each with its own set of accuracies and efficiencies. The ultimate accuracy of picture identification must be determined by taking into account all of these stages. Researchers have used multiple efficient models to train their datasets according to their findings VGG16 and VGG19 models has performed very well with better accuracies in training and they have also use models like mobileNet, LeNet5 and ResNet. According to the past research papers conducted MobileNet-V2, LeNet5, and VGG-19 has given 76.92%,76.63% and 77.09% accuracies respectively. And when it comes to the hybrid model identified through the review process, it could able to find that it as it needs less parameters for training, it is also more accurate than the other identified models. When comparing with the all other previous studies regarding to rice plant disease detection, R-CNN model has been identified as the most appropriate model for identifying and detecting various rice plant diseases. as it shows 96% accuracy for blast diseases, 95% accuracy for Brown Spot disease, and 94.5% for Sheath blight disease. Hence, in summary, this review study examined standard and updated Deep Learning structures, as well as regular mapping approaches, employed for disease detection in rice plants.