Advancements in Artificial Intelligence (AI) for enhanced insights and automation in rice agriculture: A systematic review

With the rising global demand for rice, improving production efficiency through advanced technologies like artificial intelligence (AI) is crucial. This systematic review gathered recent literatures on learning algorithm models applied to automate rice agriculture tasks. The objectives were to analyze the performance accuracy of different machine learning algorithms for rice classification and determine the most effective models. The 116 studies from 2016-2023 were screened and 70 were included. The algorithms were evaluated by weighted mean accuracy percentage across studies while maintaining consideration to sample sizes. The results showed the DenseNet121 deep convolutional neural network achieved the overall highest accuracy of 99.98%, also topping rice disease detection. For variety classification, Deep Neural Networks reached 99.95% accuracy by learning complex visual differences. Adaptive Neuro-Fuzzy Inference System led in grading quality of 98.6% by discerning grain features. Larger datasets improved accuracy indicating that the more training data has, it enhances model accuracy. The review demonstrates AI’s significant potential to automate essential aspects of rice production. Further research expanding standardized algorithm evaluations is recommended to strengthen the evidence-base and support integration of AI for intelligent, sustainable rice agriculture.


Introduction
For billions of people rice is more than just a food especially in Asia, Africa and South America [1].Nearly half of the world's population depends on rice as their dietary requirements.Each year, throughout the world nearly 510 million metric tons of rice are produced according to the Statista [2,3].Currently, rise in demand for rice is happening especially to places that are not a rice growing region and need a vast rice production to meet day to day food supply.
Agricultural research and innovation in rice production is crucial.It needs to be developed and advanced to increase output and keep up with the rising demand for rice driven by population growth and surge in food consumption.The deployment of advanced technology for rice production makes procedures in production faster and accurate.
Over the years Artificial Intelligence (AI) offers solutions that will help in farming and produce more with fewer resources, increase crop quality and quick production [4].Technology in artificial intelligence is responsible for different objectives including vision, learning and decision making.AI improves as it is constantly learning, applying its components such as Machine Learning and Deep Learning.Under those components there are different algorithms or architectures that are used to understand and process information similarly as the human brain operates [5].With that, those advancements are more utilized in the field of agriculture to enhance the automation of the productions.

Data Analysis
To synthesize and analyze the data gathered to determine the most efficient artificial intelligence algorithm used and its ranking, the method used is the weighted mean.Weighted mean is computed based on the following formula [6]; ℎ  =  (   ℎ)   ℎ The weighted mean considers the different sample sizes of each article for each algorithm's performance evaluation.This is utilized by multiplying the algorithm's scores by the sample size used for that measurement, which is then divided by the total sample size to calculate the weighted average accuracy score for each algorithm [6].This allows us to rank the algorithms while maintaining consideration of the difference in sample sizes.
Using this method provides a quantitative way to account for the different sample sizes used for each accuracy percentage.The weighted mean allows us a more precise ranking system by giving more influence to measurements with larger sample sizes and simplifies comparing the algorithms.The weighted mean consolidates the performance ratings into a combined accuracy benchmark.

Literature Review
This section comprises the existing systematic reviews about the advancements in technology applied in the field of agriculture.Several systematic reviews were gathered to synthesize the reviews of every paper and highlight the importance of adapting advanced technology in agriculture.
Surge of demand in securing the proper methods of food production calls information technology as a solution for the development in agricultural production [7].Agricultural researchers should provide more synthesized learnings and systematic reviews about the adoption of technology in agriculture, specifically, artificial intelligence.By using artificial intelligence, it will help farmers to have an improved farming process and efficiencies, waste reduction in biofuel and food production which will be non-destructive to the environment [4].Identifying the applicability of computer vision and artificial intelligence in precision agriculture of five major grains in the world.
Using computer vision systems and artificial intelligence in agricultural and industrial food production for automation tasks in the field showed that there are gaps in the development of intelligent devices but there are alternatives for future improvement.

Extracted data of the included studies
The included studies for the review, shown in Table 2, contains the information that will be synthesized to come to a conclusion of determining the most efficient artificial intelligence algorithm applied in rice production.From all the gathered data, the researchers come up with a way to determine which learning algorithm model has the highest performance accuracy ranking them from lowest to highest accuracy percentage and combine the learnings from all the studies included.Thus, this systematic review of gathering several studies related to the application of artificial intelligence in automated classifications of rice properties, to improve and hasten rice production, was conducted.

Summary result of all the analyzed learning algorithm models
The collected data has been analyzed and evaluated in terms of getting its weighted average in accuracy percentage per learning algorithm of each study.All the collected learning algorithms were ranked from lowest to highest to conclude which has the best performance that can be applied in rice production which is shown below (Figure 1).Additionally, not all the extracted learning algorithms had more than one supporting studies, some still lack evidence and foundation for it to be applied practically and so the data were separated into two: algorithms applied in more than 1 study (Figure 2) and algorithms applied in only 1 study (Figure 3), which is ranked respectively.From the analyzed data it was observed that the studies related to rice production provide different purposes as rice production is a broad topic.The researchers decided that depending on the rice production category in which a study focuses, data will be segmented into categories: rice varieties, rice health/diseases, and rice quality or grading.The graphs provided below will present the rank of the accuracy for application of AI in classification of rice varieties (Figure 4), rice health and disease detection (Figure 5), and rice quality/grading (Figure 6).Segmentation of the data will provide more specific information regarding the efficiencies of application of AI in rice agriculture.The review would broaden knowledge and yield an exact ranking of the application of different learning algorithm models per rice production category that would explain which is best to use in a specific rice production.

Discussion
The result from the analyzed data of showing the overall average accuracy ranking of all collected learning algorithm models displayed (Figure 1) that the DenseNet-121 got the best accuracy percentage of all and in contrast Color Feature + SVM was the lowest.With a further learning after the overall ranking as the data were separated, considering if a learning algorithm has only one or more than one supporting study the results of ranking are shown below, respectively.

Accuracy ranking of applied algorithm in more than one study
In the analyzed data the lowest average accuracy percentage with more than one study was the SVM + HOG.Histogram of Oriented Gradients with Support Vector Machine (SVM + HOG) is a traditional image classification method which is proposed in 2005 by Dalal and Triggs [84].In the study of Sethy et.al., SVM + HOG was included to be examined for grading the rice panicle blast which achieved an accuracy of 76% [45].Moreover, this method was also used for rice quality classification obtaining an accuracy of 85.06%.Combining these studies which applied SVM + HOG to classification of rice properties, considering the number of data samples included, results an average accuracy percentage of 81.62%.This method was originally the basis of pedestrian detection in which HOG characterizes object appearance based on the information of local intensity gradients or edge directions and the combined feature vector will then be classified by SVM [84].However, it should be pointed out that this traditional method contains simpler operations which leads to slow computational speed and its application to real-time system is limited due to that.With some modifications to the SVM + HOG or applying other algorithms may enhance the effectiveness of it.
Deep Neural Network (DNN) gained the highest accuracy percentage in all the learning algorithms applied in more than one supporting studies.A recent study about a pixel-based rice seed classification which implements DNN found that in high temperature grown rice seed HSI DNN can achieve an accuracy of 94.83% and the average accuracy in five different treatments at each exposure times and six different temperature treatment was 91% using DNN [30].Another study conducted implementing DNN was for the classification of rice varieties.The study concluded that one of the classification successes was from the model of DNN with an accuracy of 99.95% which is one of the highest among the other models tested [43].Summing up the existing studies that have been gathered in this review the result showed that the average accuracy percentage of DNN was 99.95%.In different areas of artificial intelligence, DNN has been successfully used in many machine learning applications, taking note that the multi-layer structure was the foundation architecture of DNN [85].The more hidden layers and neurons, it will allow DNN's capacity to extract complex features leading to a better algorithm for accurate classification of DNN in some cases.The findings suggest that DNN would be a successful method for rice properties classification in rice production.

Accuracy ranking of applied algorithm in only one study
The traditional image classification method with the lowest accuracy in the analyzed data was Colour Feature + SVM, with an accuracy of 60.3% based on a single study [45].Colour Feature + SVM refers to using colour features extracted from images along with a Support Vector Machine (SVM) classifier for image classification.SVM is a machine learning algorithm commonly used for classification tasks [84].In the study by Sethy et al. [45], Colour Feature + SVM was one of the traditional image classification methods examined for grading rice panicle blast severity.It achieved an accuracy of only 60.3%, lower than more advanced methods like ShuffleNet [45].This indicates that colour features alone are not robust enough for accurately discriminating between panicle blast severity levels.Other methods using more sophisticated feature extraction and learning algorithms were able to capture more meaningful features that helped improve classification accuracy significantly.Relying only on colour features limits the representational power of the model.
In the analyzed studies, the DenseNet-121 model achieved the highest accuracy among all learning algorithms, in a single study.DenseNet-121 is based on deep convolutional neural networks (DCNNs), which are a type of artificial neural network that utilizes deep learning.DCNNs have emerged as one of the most promising neural network architectures and now dominate nearly all recognition and detection tasks [87].DCNNs are a specialized class of deep neural networks designed for processing grid-like topology data like images.Their architecture of convolutional, pooling and fully connected layers allows efficient feature extraction and learning from visual data [86].By using DCNN, DenseNet-121 is able to effectively analyze images and perform recognition tasks with high accuracy of 99.98% [16].
The exceptional performance of DenseNet-121 highlights the capabilities of DCNN models for image classification.

The best learning algorithm model applied in classification of rice varieties, rice health and disease detection and rice quality
Classification of rice varieties takes the difference of each varieties features such as texture, shape, and color.With the help of classification machines, it would be easier to determine its qualities.In this rice production method, from the analyzed data it resulted that Deep Neural Network (DNN) achieved the highest average accuracy of 99.95% (Figure 4) in classification of rice varieties.In an existing study, DNN obtained a high classification success (99.95%) considering 75000 data set in classifying rice varieties because in large data sets, wide range of learning could be performed in this method [43].This method is also effective in hyperspectral rice seed classification, more importantly, in high temperature grown rice seed HIS [30].Considering the synthesized learnings, it is proven that DNN can be applied successfully in classification of rice varieties.
Ensuring rice health in rice production is essential to achieve the best quality of rice.Applying artificial intelligence in this field would improve the detection of the rice disease and evaluate the rice health.According to the analyzed data (Figure 4) DenseNet121 (Dense Convolutional Network) approach achieved the highest average accuracy (99.98%) in rice health and disease detection.More specifically, the supporting study that proves the efficiency of this method concludes that with purpose of classifying the nutrient deficiency of rice, an accuracy of 99.98% was obtained.The reason behind for the high-performance accuracy of DenseNet121 is that additional new layers were inserted, early stopping, model check points and five-fold cross validation was conducted [16].
Many instances that rice quality is not secured and not accurately graded accordingly to its respective class quality was put on sale have occurred due to several reasons one of it was lack of advancements in production.Adaptive Neuro-Fuzzy Inference System (ANFIS) which obtained the highest average accuracy percentage of 98.6% (Figure 5) in the field or rice grading outperformed other machine learning algorithms.This method accurately classified broke and whole grain to determine the quality of the rice.Surge of using advanced rice quality classifier has become a new research trend which gradually applies to the development of AI in agriculture.

Conclusion
This systematic review aimed at gathering recent artificial intelligence advancements in rice production.The goal is to analyze their performance accuracy for determining the most efficient learning algorithm models for future real-world applications.The findings showed that DenseNet121 deep learning architecture has the highest performance accuracy of 99.98% in all the reviewed articles.Specifically, this DenseNet121 also is the top algorithm for rice health and disease detection.The deep convolutional neural network allowed for a robust feature extraction and learning capability which is effective for classifying rice diseases.For the rice variety classification, Deep Neural Networks (DNNs) are the most accurate with a 99.95% performance accuracy.The multi-layer processing of DNNs allows for complex differentiation between rice varieties based on texture, shape, and color.In terms of rice quality and grading, the Adaptive Neuro-Fuzzy Inference System (ANFIS) is the top-performing learning algorithm model with 98.6% accuracy by discerning whole and broken grains.
The analysis showed that with greater dataset sizes, comes with greater algorithm accuracy across most studies.By using more training data, the learning algorithm models can have more performance accuracy.Overall, this review demonstrates the significance of artificial intelligence in automating and enhancing the various aspects of rice production.Although DenseNet121 and DNNs are already leading in performance, there is room for more accuracy and efficient rice variety classification, disease detection, and quality grading.These will be beneficial for improving crop yields of farmers, ensuring rice quality for consumers, and optimizing production.Expanding the evidence-base by evaluating more algorithms on larger standardized datasets is recommended.Some specific focus areas may include early disease prediction and integrating spectral imaging for a holistic view of rice properties.This will lead to more intelligent artificial intelligence, more sustainable rice production, and trusted rice agriculture practices.

Figure 5 Figure 6
Figure 5 Rank of the accuracy for application of AI in rice health and disease detection

Table 1
Review of related literatures

Table 2
Application of Artificial Intelligence in Classification of Rice Properties