Machine Learning Techniques for Coffee Classification: A Comprehensive Review of Scientific Research

1. Introduction

Coffee stands as a cornerstone crop globally and its consumption has been increasing in recent years [1]. This surge underscores coffee’s pivotal role in the global market, given its status as a crucial economic food commodity [2].

Current methodologies for managing agricultural data in coffee farming are increasingly considered outdated—labor-intensive, time-consuming, and, most importantly, prone to inaccuracies and errors. In this context, the advent of Machine Learning (ML) and intelligent data analytics emerges as a transformative force, propelling sustainable agriculture by enhancing food production and addressing the pressing challenges in this sector [3].

Although a considerable volume of work addresses the use of Artificial Intelligence (AI) in coffee culture, it should be noted that research has followed different paths, depending on its various aspects. There is a considerable volume of work on the use of AI in agriculture, especially in the category of predictive analysis [4]. Indeed, there are numerous areas in coffee production where ML techniques can be applied. Intelligent systems can provide advances and improvements, from grain harvesting to commercialization. In a scenario where specialized systems surpass human diagnoses in terms of trust, speed, and accuracy, it will be of fundamental importance for coffee growers to make their use widely applicable and essential to increase productivity in agriculture.

Therefore, the present work aims to review and summarize the available data regarding how ML has been used in coffee production and marketing in classification activities, which encompass the classification of defects, roasting, maturation, and sensory characteristics of coffee beans and also leaf diseases that affect the coffee plantation. This research is important and justified by understanding the relationship between AI and coffee production, identifying trends in AI applications in coffee classification, and, mainly, identifying research gaps that can be eliminated to improve processes throughout the coffee production chain.

Considering that global coffee consumption is increasing, and its importance of participation in revenue in more than 70 developing countries economies [1], combined with the capacity of AI techniques for problem-solving and decision-making, this paper presents a detailed summary of research covering coffee and AI.

Table 1 compares our literature and related literature reviews. The first column corresponds to the reference of the literature review; the second column describes the paper; the third column lists the number of papers analyzed in the work; and the last column informs the time range of the analyzed works. The review’s authors [5] analyzed seven studies that classify the quality and type of coffee beans based on images and using ML. The reviews [6,7,8] analyze research works and list challenges of disease detection of various plants. The literature review [6] focuses on leaf diseases. The study of [9] analyzes technologies for detecting roast-level coffee beans, but not all studies use ML techniques. Many studies related to smart agriculture [10,11] address machine learning techniques for classifications of other types of crops, such as rice [12,13,14,15], corn [16,17] and soybean [18].

A Comprehensive Review was used to achieve the proposed objective of synthesizing and understanding how ML techniques for coffee classification are presented in scientific research. The review was carried out in the IEEE, Science Direct, and Springer databases, with terms compatible with the object of this research.

The main goal of this review is to survey current research in the coffee area that uses machine and deep learning techniques. This research lists the works related to the classification of coffee beans and leaves in a comprehensive way. In this context, our main contributions include:

• A more up-to-date and comprehensive literature review of current ML research in the classification and detection of coffee beans and coffee leaves, covering 72 papers and different ML algorithms, including comparisons and discussions among them;
• An analysis of the evaluation metrics used in ML for classification and detection of coffee beans and coffee leaves for understanding results and reproducing them, benefiting future research;
• A presentation of architectures used based on ML for detecting and classifying coffee beans and coffee leaves;
• A detailed exploration of the main limitations, challenges, and future directions related to the use of ML techniques for the classification and detection of coffee beans and coffee leaves;
• A summary of the databases used in research to detect and classify coffee beans and coffee leaves.

This work was divided into five more sections to contextualize the use of ML in coffee farming, specifically its use on coffee beans and leaves. Following this introduction, the second section describes the research strategy and details the important points of each selected study. The third section introduces the main ML techniques used for classification tasks. The fourth section summarizes and analyzes the findings, presents the importance of coffee classification in the coffee industry, and demonstrates how agribusiness uses ML techniques. The fifth section discusses the studies’ challenges, future trends, and limitations and summarizes the dataset for classification tasks. Finally, the sixth section presents the final considerations.

2. Methodology

This paper is a descriptive research of the comprehensive review type carried out through online access to the IEEE, Science Direct, and Springer databases. In the browsing for research articles, the descriptors "machine learning" and "coffee." For the inclusion of works in the present study, the following criteria were established: articles published in the last five years before the consultation, articles in which the topic of "machine learning" is related to coffee classification techniques, and articles available in full format in the databases above. Theoretical studies and systematic literature reviews were not included in the overview of synthesized research data.

The article analysis process followed these steps: 1) search for descriptors in the databases above; 2) exclusion of articles published more than five years ago; 3) exclusion of works with low relevance; 4) critical reading of the articles and checking whether they meet the proposed theme; 5) exclusion of theoretical articles and literature reviews; 6) final analysis. After selecting the articles using inclusion and exclusion criteria, tables containing the main information about the objectives and scenario of application, the methodology used for coffee classification analysis, and the results were created.

In the IEEE database, 114 articles were found with the words “machine learning” and “coffee”. Fifty-eight (58) articles were disregarded because they were unrelated to the coffee area; for example, the author’s last name is Coffee or the three searched words do not appear simultaneously in the text. Only "machine learning", or only "coffee". Ten (10) articles were disregarded because they were not related to classifying coffee beans, such as Predicting Coffee Crop Yield, detecting coffee trees, coffee price prediction, and agriculture 4.0. Finally, two (2) articles were disregarded in the analysis because it was a literature review. In the same way, in the Science database, 2,828 articles were found; 2,770 were disregarded because it was not related to the general theme; 34 were disregarded because it was not related to the specific theme; and two (2) articles were disregarded in the analysis because it was a literature review. In summary, 7,630 articles were found in the Springer database, but 7,617 were disregarded; 6 were disregarded because they were unrelated to the specific theme; and 1 article was disregarded in the analysis because it was a literature review. Figure 1 shows the research methodology used in this study. Figure 1 shows the total number of publications per database and the steps used to choose the selected publications.

In Figure 2, we can see a fluctuation in the number of publications over the last five (5) years, which various factors may have influenced. In general, the number of publications has increased over the years.

Additionally, Table 2 below shows the number of publications per database searched. As illustrated, the first column corresponds to the database, and the second column describes the number of publications.

2.1. Research Questions

This review was designed to answer twelve Research Questions (RQ). The RQ are as follows.

RQ1: What are the most used architectures based on ML for detecting and classifying defects in coffee beans?

RQ2: What are the most used architectures based on ML for detecting and classifying coffee beans roasting?

RQ3: What are the most used architectures based on ML for the sensory classification of coffee?

RQ4: What are the most used architectures based on ML for detecting and classifying the maturity of coffee beans?

RQ5: What are the most used architectures based on ML for detecting and classifying coffee diseases?

RQ6: What defects are considered when classifying the coffee bean in the studies?

RQ7: What stages are considered to classify coffee roasting in the works studied?

RQ8: What stages are considered to classify coffee maturity in the works studied?

RQ9: What are the different coffee diseases addressed in the works studied?

RQ10: What challenges exist, and how can research be improved in detecting and classifying coffee beans and leaves based on the studies reviewed?

RQ11: How diverse is the research on coffee bean and coffee leaves detection or classification, and for what classification purpose are ML applications the most?

RQ12: Which databases are available for research into detecting and classifying coffee beans and coffee leaves?

The questions RQ1, RQ2, RQ3, RQ4, RQ5, RQ6, RQ7, RQ8, and RQ9 aim to map the most used architectures for ML in coffee bean classification and, consequently, classify commonly used ML approaches to resolve them. Based on this review, these five questions identify the main ML algorithms and their main metrics used to achieve the classification objective. In these questions, it is possible to carry out a quantitative and qualitative analysis of the more effective ML algorithms to be applied to coffee bean classification. Through mapping, it is possible to identify how researchers are addressing the techniques that involve ML and consequently analyzing potential gaps related to the topic.

By answering RQ10, RQ11, and RQ12, we can understand the challenges, future trends, and final observations when implementing ML techniques in various coffee bean and coffee leaves classification scenarios.

Finally, RQ1 and RQ6 are answered in Subsection 4.1, RQ2 and RQ7 in Subsection 4.2, RQ3 in Subsection 4.3, RQ4 and RQ8 in Subsection 4.4, RQ5 and RQ9 in Subsection 4.5, and RQ10, RQ11 and RQ12 in Section 5.

3. Machine Learning Techniques for Coffee Classification

ML is an area of AI capable of generalizing a problem from a set of training data to correctly classify or predict data that has not been previously observed [19]. It is a tool that can be applied to any field that needs data analysis.

There are three ML techniques: unsupervised, supervised, and reinforcement learning. A training set is observed in supervised learning and comprises examples of input and output pairs. The machine learns a function that maps the input to the output, and this function is called the hypothesis function [20]. The supervised learning problem can be classified when the output value has finite values or regression when the output value has infinite values. The idea of classification involves grouping items into classes or groups. The difference between supervised and unsupervised learning can be seen in Figure 3 and Figure 4 below.

Computer vision is an area of AI that seeks to analyze, interpret, and extract relevant information from images so that decisions can be made. Object recognition in its complete generality is difficult [20]. Convolutional Neural Network (CNN) is a ML model that extracts features in various computer vision tasks like image classifications. Several CNN architecture models, such as AlexNet, VGGNet, ResNet, MobileNet, EfficientNet, and DenseNet, are used for image classifications. The Figure 5 shows a typical CNN.

Vision Transformer is a neural network with the potential for extensive use in image classification tasks [22]. The Figure 6 shows a illustration of the Vision Transformer Model.

ML techniques combined with computer vision have become decisive in developing agricultural efficiency and have been widely applied in research in various areas of agriculture [24], including coffee farming. Due to different algorithms, this research seeks to identify which models achieve greater accuracy in image classification.

4. Overview of Synthesized Research Data

This section summarizes the most significant studies on ML in coffee classification. It details each study, including the reference (authors and publication year), objectives and application scenarios, methodology (architectures and dataset), and main results.

We identified 72 selected studies that could be grouped into 5 common themes: classification of (1) coffee defects, (2) coffee roasting, (3) coffee aroma and flavor, (4) coffee maturity, and (5) coffee diseases. The 5 categories are listed and commented below.

4.1. Coffee Defects Classification

Defects and impurities in coffee beans reduce their quality. Defective beans affect the flavor of the coffee and, therefore, devalue it and affect the producer’s profitability. Several factors can cause these defects, such as problems in production and storage and physiological and genetic changes. Identifying defects makes it possible to improve management and prevent the occurrence of these defects, which depreciate the coffee.

As illustrated in Table 3 and Table 4, the first column corresponds to the reference of the literature review; the second column describes the main objectives of the paper; the third column lists the models and algorithms used in the paper; and the last column informs the results achieved.

Most coffee defect classification research consists of categorizing coffee according to effective beans of an intrinsic nature (imperfect beans), which are beans damaged by the imperfect application of agricultural processes such as improper drying or picking of overripe cherries [44]. These coffee beans are black, broken, and damaged by insects, including the coffee borer beetle.

The challenge in future research in defect classification is to cover all defects originating from coffee cultivation. Few studies have identified external defects in the coffee bean, such as the presence of sticks, stones, clods, bark, and small insects. These defects originate during harvesting due to coffee falling on the ground and poor fanning. In the same way, few studies have identified acidic coffees derived from prolonged fermentation.

The paper [30] classified 6 types of coffee intrinsic and extrinsic defects. The authors created a screening system with a user interface. The authors of [34] created a high-speed near-infrared (NIR) camera to capture images of coffee beans moving on a conveyor belt. [25] also proposed a conveyor for capturing and uploading images.

[26] used an automated system to classify good and bad coffee beans. The bad coffee beans are divided into three categories: broken, insect-infested, and mold. The system recognizes the coffee beans; if all classification results are negative, the coffee bean is considered good and kept. If at least one output indicates a defect, the coffee bean is labeled with the corresponding type of defect. [28] classified all defects categorized by the Specialty Coffee Association of America (SCAA).

Meanwhile, [27] developed a Python algorithm to extract texture, shape, and color characteristics from images of coffee beans. Specialty coffees go through a more laborious production process, require special care from planting to roasting, and are made of pure, unmixed, high-quality coffees. At the same time, traditional coffee has inferior and defective beans. This mixture reduces the quality of the coffee as it changes its flavor and reduces its cost. [35] classified coffee beans in different regions of Brazil as special or traditional. The papers [38] and [37] classify coffee bean species. [38] classified arabica and robusta coffee plants, and [37] classified espresso, Kenya, and Starbucks pike place species.

The networks tested for classifying coffee defects were Resnet-50, Densenet201, SVM, DNN, RF, GAN, MobilenetV2, Hough Circle Transformer, Mask R-CNN, NFNet-F3, SqueezeNet, Slim-CNN, InceptionResnetV2, ANN, InceptionV3 and CNN networks proposed by the researches authors.

Enhanced, AlexNet, ResNet-18, KNN, MLP, VGG-16, VGG-19, MobileNet were also tested, but obtained inferior results when compared to other networks.

The highest accuracy achieved was 99.58% for the ANN network proposed by the work (tsai, 2023). Tsai’s study combines ANN with Mass Spectrometry and analyzes 2 classes of single beans: civet coffee beans and regular beans. It is not possible to say which network model presents the best performance among all the researches analyzed because the databases tested were different. The works differ in the defects they seek to classify and consequently in the number of classes presented in each model. The researches used datasets from 100 images to around 20,000 images. The size of the dataset also influences the model result.

4.2. Coffee Roast Classification

Coffee roasting is one of the final stages of production and a determining factor in producing a good drink. Good roasting highlights the characteristics of each type of bean, which are flavor, aroma, and acidity. The same coffee bean may have different characteristics depending on the type of roast. The color of the coffee bean and the aroma can differentiate roast levels. The coffee beans change color as the temperature increases, and the roast can be light, medium, or dark. Each of them will have different flavor characteristics.

Roasting is defined by the time the coffee bean remains at a high temperature under the careful supervision of a specialized professional. Work that classifies the roasting level and analyzes the process until the coffee reaches the final carbonization stage contributes to automating the process without professional supervision.

As illustrated in Table 5, the first column corresponds to the reference of the literature review; the second column describes the paper’s main objectives; the third column lists the models and algorithms used in the paper; and the last column informs the results achieved.

The authors of papers [46,47] developed a mobile application that automatically classifies coffee bean quality based on its roast degree.

Additional work has been carried out on the automated diagnosis of coffee roasting level, achieving excellent results. [45] achieved a maximum 100% accuracy in classifying Arabica coffee into light, medium, and dark. To predict the roasting degree of coffee beans, the authors of [48] created a model that consists of 4 input data: temperature, humidity, place of origin, and temperature of the roasting curve sampled every 15 seconds.

The networks tested for coffee roast classification were SVM, ResNet-152, MobileNetV2, Fully Connected Neural Network, SPSO, DenseNet121, and CNN proposed by the research authors.

VGG16, MobileNetV1, NasNetMobile, Linear Regression, DT, RF, Support Vector Regression, and Fully Couped Neural Network were also tested, but obtained inferior results when compared to other network models.

The research by (Septiarini, 2022) achieved maximum accuracy to classify 3 roast levels using SVM. It is not possible to say which network presents the best performance among all the works analyzed because the databases tested were different. The works differ in the roast levels they seek to classify and consequently in the number of classes presented in each model. The studies studied used datasets from 160 images to around 11,000 images. The size of the dataset also influences the model result. More studies are needed to consider different coffee roasting levels.

4.3. Coffee Sensory Classification

During coffee production, the beans can go through different processes. While traditional coffee has undesirable characteristics such as firm bitterness, special coffees have a diverse range of flavors and aromatic notes and are better balanced. Sensory analysis of coffee involves identifying its properties in the drink. The method takes place when the taster separates, inhales, and ingests a small amount of roasted and ground coffee and evaluates the following sample factors: Fragrance, Aroma, Uniformity, Sweetness, Flavor, Acidity, Body, Finish, and Balance.

As illustrated in Table 6, the first column corresponds to the literature review’s reference; the second column describes the paper’s main objectives; the third column lists the models and algorithms used in the paper; and the last column informs the results achieved.

The models for finding the sensory characteristics of coffee (aroma and flavor) do not use images, these studies employ a dataset with coffee samples and can use an electronic tongue or nose. The electronic nose (e-nose) has gas sensors that can identify patterns when combined with MA algorithms. The dataset of these studies is composed of samples of coffee odors collected by the gas sensors present in the e-nose. The samples are submitted to machine learning algorithms to recognize and classify the coffee brand [58] and coffee bean types [56]. [53,54,56,58] classified coffee beans based on their aroma using an e-nose. The method proposed by [58] achieved 100% accuracy for classifying two distinct coffee brands using odor samples collected by e-nose gas sensors. The electronic tongue (e-tongue) has low-selective potentiometric sensors that respond to a wide variety of flavors. [55] classified the taste of coffee using an e-tongue and achieved excellent results. The dataset [55]’s research is composed of coffee cup samples prepared by a coffee machine, and the model identifies 21 varieties of coffee and achieves an average accuracy of 91.3%. The study [57] classified the quality of coffee drinks based on measurements made to almond and roasted coffee beans. [59] used a sensory evaluation of the coffee quality scores dataset with features related to aroma, flavor, acidity, body, balance, uniformity, and others using various linear regression models to predict coffee quality scores.

The networks tested for coffee sensory analysis were SVM, KNN, PLSR, LDA, ANN, SVM and various Regressions Models. It is not possible to say which network presents the best performance among all the works analyzed because the techniques used were different (e-nose and e-tongue) and the databases tested were also different.

4.4. Coffee Maturity Classification

Coffee maturation begins with the fruits turning light green, evolving into cherry fruits ready to be harvested. The previous year’s flowering influences fruit maturation and is directly associated with the quality of the drink. Identifying the maturity of the coffee is very important because the coffee must be harvested with the highest percentage of ripe ”cherry” fruits to provide the best quality drink.

Research in coffee classification must consider the ability to generalize to other coffee-producing regions and different coffee varieties. The characteristics extracted from the images may not capture all the nuances of coffee maturity.

As illustrated in Table 7, the first column corresponds to the literature review’s reference; the second describes the papers’s main objectives; the third lists the models and algorithms used in the paper; and the last column informs the results achieved.

CNN networks are used to classify the ripening stage of coffee beans by [60,62,65]. The paper’s authors [60] classified coffee cherry fruit into immature, semi-mature, mature, overripe, and dry. [61] classified coffee cherries based on skin color and shape characteristics. [63] classified coffee cherries based on spectral and textural characteristics.

The networks tested to identify coffee maturity were DenseNet201, KNN, ResNet50, Naive Bayes, and GoogLeNet. VGG16, VGG19, InceptionResNet-V2, Inception-v3, ANN, Deep-CNN, ShuffleNet were also tested, but did not achieve satisfactory results when compared with other networks. It is not possible to say which network presents the best performance among all the works analyzed because the databases tested were different. The researches differ in the maturity levels they seek to classify and consequently in the number of classes presented in each model. The studies studied used datasets from 600 images to around 5,000 images. The size of the dataset also influences the model result.

4.5. Coffee Disease Classification

Several diseases affect coffee farming. These diseases cause serious damage and reduce or nullify coffee productivity when poorly controlled. Studies on many coffee classifications aim to accurately and quickly identify diseases in coffee leaves for effective management and control.

As illustrated in Table 8, Table 9, Table 10, and Table 11, the first column corresponds to the literature review’s reference; the second column describes the paper’s main objectives; the third column lists the models and algorithms used in the paper; and the last column informs the results achieved.

The white stem borer disease is a major pest of Arabica [75,101] proposed a method to identify them. The beetle pierces the coffee stem and remains buried until the plant falls completely. As the disease is difficult to diagnose in the early stages of its infestation, it causes major problems for coffee producers. The authors of the paper [75] used the YoloV5 deep learning model to identify white stem borer disease using color, texture, and shape characteristics. The Jackal UGV robot navigated and mapped the simulated environment of an Arabica coffee plantation using a depth camera and state and location estimation algorithms.

Coffee Leaf Rust is a serious disease that affects many coffee-producing regions. It is caused by a pathogenic fungus that attacks the underside of coffee leaves and is described by yellow-orange, powdery dots. If left untreated, rust can cause a drop in coffee production. [76] carried out a comparative study designed to identify Coffee Leaf Rust with twenty-five models and achieved the best performance of 90% with models ResNet101V2, InceptionV3, ResNet50V2, Xception, and DenseNet169. [67] used a CNN architecture to spot rust infestation and identify rust in coffee leaves of different varieties and stages of development.

The paper [84] presented an application for identifying and quantifying diseases and pests in coffee leaves using smartphone images. The application consists of two main modules: semantic segmentation for severity and symptom classification.

The authors of the paper [66] proposed a method to capture the discolored distribution of coffee leaves, allowing her to identify the severity of the disease more easily.

In [68], the authors achieved 100% accuracy in identifying disease with no errors using VGG16 architecture, hence demonstrating that the appropriate configuration and optimization can generate efficient results. [81] also used VGG16 to classify diseases in coffee leaves. In this paper, the authors modified VGG16 and achieved an accuracy of 97.9% to classify healthy and infected coffee leaves.

The work in [77] proposed a hybrid feature fusion with MobileNetV3 that extracts local features, while Swin Transformer extracts high-level features classification performance and rust. [78] used two models. The first model identifies coffee leaves. This model used a CNN architecture with 10 convolutional layers, Adam optimization, softmax, and 100 epochs and achieved an accuracy of 95%. The second model was used to detect diseases from coffee leaves, and the CNN architecture has 20 convolutional layers, Adamax optimization, and 100 epochs, achieving an accuracy of 90%. In the same way, [79] proposed two architectures to classify diseases of coffee leaves. The first identifies the diseases of coffee in six classes. The second architecture is used to improve the classification performance made by the first stage.

On the other hand, [87] proposed three experiments that can learn to classify using just a few examples. Each experiment was tested with two architectures: Tripletnet and Protonet. The first two experiments were tested to classify the biotic stress of coffee leaves. The third experiment sought to automatically estimate the severity of the disease: low, very low, high, very high.

The paper [82] introduced a dataset with images captured in a controlled environment showing nutritional deficiencies in coffee leaves. The nutritional deficiencies are Nitrogen, Phosphorus, Potassium, Magnesium, Boron, Manganese, Calcium, and Iron. The nutritional deficiencies are detected and classified by analyzing the color and shape characteristics. Leaves with more than one deficiency can also be detected.

Moreover, [83] proposed a model with GoogLeNet and RestNet in the feature extraction phase and MLP with traditional ML classifiers such as KNN, DT, SVM, and RF for classification tasks.

Three CNN-based methods were proposed in [93]. ECNN focuses on the effective concatenation of five CNNs; HLGGM combines dimension-reduced Mobile-Net v3 features with handcrafted features; and HLGCM has reduced dimensionality and uses a DT.

The networks tested to identify coffee diseases were SVM, VGG16, ResNet50, Deep Belief Network, BPNN, MobileNetV2, ResNet101V2, TripletNet, ProtoNet, GooLeNet, ResNet, UNet, AlexNet, ANN, Mask R-CNN, XGBoost, RFT, NVIDIA Digits, Decision Tree, Enhanced EfficientNetV2-5, and DenseNet264. MobileNet, MobileNetV3, RNN, CNN, Swin Transformer, PSPNet, KNN, MLP, GradBoosting, RLM, OpenCV, LDA, NB, QDA, DenseNet121, were also tested, but did not achieve satisfactory results when compared with other networks. It is not possible to say which network presents the best performance among all the works analyzed because the databases tested were different. The works differ in the diseases they seek to classify and consequently in the number of classes presented in each model. The studies studied used datasets from 400 images to around 58,000 images. The size of the dataset also influences the model result.

The proposed methodologies for classifying coffee diseases can be applied to assist farmers in early detection and making decisions to protect coffee plantations. It is necessary to include more classifications for coffee diseases and expand the size of the dataset. Future studies can also compare other deep learning algorithms with other pre-trained models to achieve higher accuracy.

5. Challenges and Future Trends

ML, along with complementary technologies, such as Computer Vision, are revolutionizing coffee farming, propelling agribusiness towards unprecedented efficiency, speed, and precision in decision-making while reducing operational costs. Our review evidenced the existence of several computational models devised for coffee bean classification, addressing various attributes such as defects, roasting, diseases, aroma and flavor, and maturation of coffee, hence demonstrating the potential for continuous advancements. However, despite this progress, most of the existing AI-based models and tools predominantly reside in experimental phases, lacking widespread practical deployment.

They further often exhibit a restricted focus, tailored to particular regions [82,89] or coffee types [35,45,54,64,68,74,85,90,94,97] and further demonstrate a limited scope in identifying nuanced characteristics like roast degrees [47] and maturity stages [60,61]. The need for more universally applicable models is evident, especially in areas like aromatic profiling, where ML remains underutilized, revealing significant untapped potential for research and industry applications.

Furthermore, the scarcity of extensive image datasets is a critical bottleneck, underscoring the necessity for enriched coffee bean data to enable robust model generalization across diverse environmental and cultural contexts. Table 12, Table 13 and Table 14 provide a concise summary of the datasets used in the analyzed articles and are available for download. Additionally, the Tables list datasets that could be valuable for future coffee classification tasks using ML.

Although little explored, agribusiness professionals should prepare themselves for the future tools that will emerge from the research results presented in this article. These agribusiness professionals will encounter new challenges and changes in the way they work. They must adapt and relearn their roles, focusing on their analytical skills in data-driven agriculture.

Recognizing the urgent demand for refined classification tools to aid coffee growers, this work underscores the importance of machine learning-based image classification in supporting decision-making and enhancing overall industry efficiency.

Finally, the transformative power of ML and Computer Vision offers promising prospects for coffee bean diagnostics, promising to enhance productivity and management confidence significantly. Future research should, therefore, expand upon these findings, focusing on refining classification techniques for a broader and more accurate application in coffee bean classification.

As a result, Figure 7 lists the roadmap of future trends for resources in ML coffee classification. To maximize production, reduce costs, and increase precision in decision-making in coffee farming, future research needs to increase the capacity to implement higher-performance ML algorithms, increase the size of image datasets, and, finally, develop fewer generalist methods and more classes considered.

6. Conclusions

This comprehensive literature review examined the application of ML techniques to coffee bean and leaves classification, focusing on the last five years of research related to coffee defects, roasting, maturity, sensing, and diseases.

The synthesis of data extracted from the reviewed works offers valuable insights, presenting a comprehensive overview of contemporary day research, the most frequently used algorithms, and model comparisons.

Some of the research directions for the future include creating new data sets that are larger, less general, and with more classes, and, above all, the search for increasing the capacity to implement higher-performance ML algorithms.

Notably, there has been a significant increase in relevant scientific publications, reflecting the growing impact of machine learning-based image classification in advancing coffee agriculture. This surge in scholarly activity has catalyzed increased investments in developing new products and services, underscoring the promising trajectory for future ML research within the coffee domain.