Overview of Artificial Intelligence Applications in Roselle (Hibiscus sabdariffa) from Cultivation to Post-Harvest: Challenges and Opportunities

Abstract

Hibiscus sabdariffa (H. sabdariffa) is a high-value economic and functional crop, limited by agroclimatic conditions and low technological adoption. This systematic review examines the current state of artificial intelligence applications in agricultural management, analyzing 2111 records, selecting 82, and synthesizing 22 studies that meet the inclusion criteria. This review adopts a holistic framework aligned with three priority areas in agriculture—resource and climate management, crop productivity and quality, and sustainability—to explore how AI addresses key challenges in the cultivation and post-harvest processing of Hibiscus sabdariffa. The results show a predominance of classical machine learning techniques, with limited implementation of deep learning models. The most common applications include image classification, yield prediction, and analysis of bioactive compounds. However, limitations remain in the availability of open data, reproducible code, and standardized metrics. The narrative synthesis identified clear opportunities to integrate emerging technologies, such as deep neural networks and the Internet of Things (IoT), particularly in water management and stress monitoring. The review concludes that strengthening interdisciplinary research and promoting data openness is key to achieving a more resilient, sustainable, and technologically advanced crop.

1. Introduction

The cultivation of H. sabdariffa commonly known as hibiscus or roselle is of significant importance both economically and agriculturally, especially in tropical and subtropical regions. This plant is valued for its multiple uses, ranging from beverage production to medicinal and culinary applications. In countries like Mexico, H. sabdariffa is an economically important crop, with states such as Guerrero leading its production [1]. According to the Agri-Food and Fisheries Information Service (SIAP, 2023), Mexico has favorable agroclimatic conditions for the development of this crop, which is in high global demand. However, due to low yields, more than 50% of domestic consumption is supplied by imports.

Additionally, H. sabdariffa production faces various phytosanitary challenges that can affect its growth and quality. Among the main pests that affect this crop are the cutworm (Agrotis sp.), the cotton aphid (Aphis gossypii), the stem borer (Eutinobothrus brasiliensis), the whitefly (Bemisia tabaci), and the leaf cutter ant (Atta spp.). These pests can cause significant damage to plants, affecting both the quantity and quality of production [1,2,3,4]. In addition to pests, fungal diseases pose a considerable threat to H. sabdariffa cultivation. Pathogens such as Corynespora cassiicola and Coniella musaiaensis have been associated with calyx spotting, a condition that deteriorates the appearance and quality of the final product, reducing its commercial value [5].

The presence of weeds also presents a challenge as they compete with the crop for nutrients, water, and light, reducing yield. Proper weed management is essential to ensure optimal H. sabdariffa growth [1,6,7].

Climate has a significant impact on the ideal growth of H. sabdariffa, which thrives in tropical and subtropical regions. However, the increasing occurrence of extreme weather events, such as heat waves and torrential rains, can disrupt plant growth and affect calyx quality. Prolonged exposure to high temperatures can cause water stress, while excessive rainfall promotes the development of fungal diseases in the soil and leaves [8,9].

According to [10] from the Value Chain and Agricultural Cluster Development Project in Nicaragua, roselle or H. sabdariffa can be grown in tropical and subtropical climates at altitudes ranging from sea level to 1400 m. Optimal conditions for its development include temperatures between 22 °C and 25 °C and annual rainfall between 500 and 1000 mm. At higher altitudes, plant growth can be affected by lower temperatures and a more extended vegetative period, which can reduce productivity and alter the chemical composition of calyces, resulting in a decrease in the concentration of anthocyanins and other bioactive compounds.

Water availability is another limiting factor for H. sabdariffa cultivation, especially in regions with irregular rainfall patterns. In arid or semi-arid areas, the implementation of drip irrigation systems can be crucial to ensure adequate water supply, optimize resource use, and improve crop sustainability [11]. Recent studies have explored artificial intelligence-based predictive models to estimate plant water requirements, representing a promising tool to mitigate the effects of climate change in precision agriculture [12].

In light of these challenges, it is clear that advanced technologies need to be implemented to enable early monitoring and detection of pests, diseases, and weeds, as well as post-harvest tracking in H. sabdariffa. The use of tools such as computer vision offers promising solutions to efficiently identify and manage these problems, minimize losses, and optimize production. The adoption of these technologies not only improves crop profitability but also contributes to more sustainable and precise agricultural practices [13,14,15,16].

This review aims to analyze recent advances in the application of artificial intelligence to the cultivation and post-harvest of H. sabdariffa, identifying the methodologies used, the types of data processed, and opportunities to improve production and sustainability, as well as the literature gaps. Through a systematic analysis, it seeks to provide a comprehensive overview of the current state of the field and propose perspectives for future research.

Artificial Intelligence in Agriculture

The application of artificial intelligence (AI) techniques in image and data analysis has revolutionized the agricultural sector, providing efficient tools to improve crop management and optimize post-harvest processes. These approaches enable the automated evaluation of visual information, facilitating the early identification of factors that can affect production, such as the presence of pests or diseases. Thanks to these advances, it is now possible to optimize agricultural resources and improve decision-making with greater precision in all processes. Various approaches to visual data analysis have proven effective in detecting complex patterns under variable conditions, thus contributing to more sustainable and efficient agriculture [17,18,19].

In addition, supervised and unsupervised machine learning have been used to predict agricultural yields by evaluating factors such as soil quality, nutrient availability, and climatic conditions. For example, the study in [20] presented a deep learning approach designed to predict the levels of nitrogen, phosphorus, and potassium in soil by analyzing the growth patterns of cabbage plants. In [21], the authors proposed an innovative classification model called CNN + AFN to detect nutrient deficiencies in tomato crops using images of tomato leaves. This model combines convolutional layers for feature extraction with a supervised learning method known as the artificial hydrocarbon network (AHN) serving as the dense layer. Their experimental findings demonstrated that the CNN + AHN classifier effectively estimates low nutrient concentrations in tomato plants. Meanwhile, the work in [22] focused on identifying and classifying uniform production zones for common beans by applying machine learning to simulated yield data. The study aimed to determine the environmental variables that define these zones and to construct a spatial and temporal sowing calendar tailored to rainfed (wet and dry seasons) and irrigated (winter season) farming systems.

Supervised learning models can be trained to predict harvests based on historical data, while unsupervised models allow for the discovery of non-obvious relationships within agricultural datasets [23]. These applications have been essential for improving the efficiency of resources such as water and fertilizers, promoting more sustainable agriculture.

Another area of application is weed management, where image segmentation and classification algorithms enable the differentiation of weeds from crops [24,25,26,27,28,29,30]. This facilitates the control of targeted weeds, reducing the need for herbicides and their environmental impact. In general, computer vision and machine learning redefine the concept of precision agriculture by integrating advanced technology to maximize productivity, minimize costs, and mitigate environmental impact [31,32,33,34].

2. Synthesis of Agronomic Challenges Affecting Hibiscus sabdariffa

The production of H. sabdariffa faces multiple challenges that affect its yield, sustainability, and commercial quality. Although it is a plant of great economic and cultural significance in various regions of the world, its development is limited by several issues.

Based on the analysis of the included studies, this review systematically identified eight key challenges related to H. sabdariffa cultivation, which were grouped into three main categories: agricultural resource management and climatic conditions, crop productivity and quality optimization, and economic and environmental sustainability [35,36]. First, the management of agricultural resources and climatic conditions includes issues such as monitoring soil quality [37,38,39], which is crucial to ensure optimal growth conditions, efficient water resource management [40,41] in arid regions, and mitigation of the effects of climate change [42], which directly impact production. Second, optimizing crop productivity and quality encompasses strategies such as biological control of pests and diseases to reduce losses [43,44], improving harvesting processes to increase efficiency [45,46], and monitoring bioactive compounds, such as anthocyanins [47,48], which are key to the commercial quality of calyces. Finally, the economic and environmental sustainability of the crop focuses on the utilization of agricultural by-products [49,50,51] and the management of emerging diseases [52,53], two fundamental aspects to improve the economic viability of the crop and minimize its environmental impact. These interrelated areas provide a comprehensive framework for addressing challenges and maximizing the potential of this important crop.

Table 1. Challenges related to H. sabdariffa cultivation and post-harvest, their references, and their impact.

Challenge	Issues	Impact Description
Agricultural Resource Management and Climatic Conditions	Low fertility, salinity, and pH imbalance limit crop growth. Monitoring technologies can optimize soil fertility [54,55,56,57,58,59].	Poor soil quality directly affects yield and crop sustainability, increasing the need for external inputs such as fertilizers.
	Water scarcity in arid and semi-arid regions limits crop production. Smart irrigation systems optimize the use of this resource [60,61,62].	Water scarcity significantly reduces yields, especially in arid areas, affecting the economic viability of the crop.
	Climate change, with extreme temperatures, prolonged droughts, and torrential rains, affects production and complicates agricultural planning [63].	Water and heat stress caused by climate change reduce crop yields, impacting agricultural economies and food security.
Crop Productivity and Quality Optimization	Pests such as aphids and fungal diseases limit crop productivity. Biological and chemical control are key strategies for integrated management [64,65,66,67,68].	Pests and diseases increase production costs due to the need for constant controls, affecting both yield and calyx quality.
	Anthocyanin concentration, which determines product quality, is influenced by climate, soil fertility, and crop management [69,70,71,72,73,74].	Variability in anthocyanin concentrations affects commercial value and the functional properties of the final product, impacting market acceptance.
	Manual harvesting is inefficient and costly. Mechanization improves efficiency and reduces operating costs in large plantations [75,76,77,78,79,80,81,82,83,84,85].	Lack of mechanization increases production costs, making the crop less competitive in global markets.
Economic and Environmental Sustainability of the Crop	Diseases such as fungal and bacterial phytopathogens reduce productivity and calyx quality, threatening crop sustainability [86].	Emerging diseases cause significant losses by reducing both the quantity and quality of the final product while increasing management and control costs.
	Residues such as leaves and stems can be transformed into biofuels and value-added materials, promoting a circular economy [87,88,89,90,91,92,93,94,95,96,97,98].	Lack of by-product utilization leads to resource waste and missed opportunities to generate additional income.

presents the eight challenges divided into three groups of issues, along with their direct impact on H. sabdariffa cultivation and post-harvest. These challenges highlight the need for technological solutions and sustainable practices to address current and future problems. The table was compiled based on references that contribute to each challenge within the three groups of issues, which align with the core domains of precision agriculture: resource management and climate adaptation, productivity and quality optimization, and sustainability. These references were synthesized from the broader scientific literature and are not limited to the systematic review.

3. Methodology

We conducted a systematic review using the PRISMA 2020 methodology [99], which facilitated the identification, evaluation, and synthesis of the literature, employing applied inclusion and exclusion criteria.

The search for the article was conducted in ten different academic databases: ScienceDirect, IEEE Xplore, Taylor & Francis, Wiley Online Library, Nature, Springer, MDPI, PubMed, Google Scholar, and SciELO.

Table 2. Queries used in academic databases with Boolean operators and language adaptation.

Database	English	Spanish
ScienceDirect, Taylor and Francis, Wiley Online Library, PubMed, MDPI	(“Hibiscus sabdariffa” OR ”Roselle”) AND (“Artificial Intelligence” OR “Computer Vision” OR “Deep Learning”)	–
NATURE, IEEE Xplore	(“Hibiscus”)	–
Springer	(“Hibiscus sabdariffa” AND “AI”)
Google Scholar	(“Hibiscus sabdariffa” OR “Roselle”) AND (“Artificial Intelligence” OR “Machine Learning” OR “Deep Learning” OR “Computer Vision”)	(“Hibiscus sabdariffa” OR “Roselle”) AND (“Artificial Intelligence” OR “Machine Learning” OR “Deep Learning” OR “Computer Vision”)
IEEE LatinAmerica, SciELO	–	(“Hibiscus sabdariffa” OR “Roselle” OR “Flor de Jamaica”)

shows the queries used for the searches in the different databases. The article selection process consisted of three phases: title screening, abstract screening, and full-text evaluation. Inclusion and exclusion criteria were applied, focusing on the use of H. sabdariffa crops, post-harvest, and AI algorithms. Publications from 2006 to 2025 were considered, including works published in both Spanish and English. The last comprehensive literature search was conducted on 5 August 2025, across all selected databases to ensure inclusion of the most recent studies.

Across various databases, the number of documents retrieved are as follows: ScienceDirect yielded 104 documents, Taylor and Francis retrieved 20, Wiley Online Library found 105, Springer provided 157, NATURE offered 34, IEEE Xplore held 46, PubMED showed 4, MDPI presented 20, Google Scholar located 1549, SciELO discovered 75, and IEEE Latin America found no documents.

The criteria for exclusion were as follows: (1) any varieties aside from H. sabdariffa, and (2) documents that utilized statistical methods or software, including ANOVA, SRM, SAS, SuperPro Designer, SPSS, GraphPad Prism, Excel, OriginPro, Minitab, SigmaStat, or Origin. On the contrary, the inclusion criteria encompassed all documents employing any artificial intelligence techniques, as illustrated in Figure 1, embracing even traditional methods. Overall, 73 documents satisfied the inclusion and exclusion criteria.

When considering the extensive scope of AI, it is essential to recognize its vast and cutting-edge field, which is divided into numerous subfields. These share a similar methodology, aimed at tackling distinct tasks within specific areas. As noted in [100], AI can be classified into seven key domains: robotics, natural language processing, speech recognition, planning, computer vision, machine learning, and expert systems. In contemporary society, these technologies play a crucial role in addressing various issues, minimizing human involvement, and enhancing different operations. Agriculture, in particular, also reaps the benefits of these advancements. Figure 1 illustrates a diagram showing these paradigms along with their respective subfields.

As shown in Figure 2, a total of 2114 records were identified across various databases, including ScienceDirect, SpringerLink, IEEE Xplore, PubMed, Google Scholar, among others. Duplicate records were not removed at this stage. After the initial screening based on title and abstract, 2029 records were excluded, 1942 of which were related to species other than H. sabdariffa or did not involve the use of AI and 87 that used only conventional statistical methods without AI. Subsequently, 85 full-text publications were recovered, with no unrecoverable documents. During the eligibility assessment, 62 publications were excluded for not applying AI, duplication, or not being related to the H. sabdariffa variety, resulting in 23 final studies included in the systematic review. As part of the citation search process, 271 records were identified and screened. However, none met the inclusion criteria as 94 did not apply AI techniques and 177 referred to different hibiscus varieties. In total, these 23 studies form the basis of the review, corresponding to the number of publications considered in the analysis.

Figure 1. Artificial intelligence domains and subfields, inspired by [101].

Figure 1. Artificial intelligence domains and subfields, inspired by [101].

Table 3. Documents found and selected from the search.

Database	Literature
	Found	Selected	Included
ScienceDirect	450	104	9
Taylor and Francis	1581	20	0
Willy on Library	1300	105	1
Springer	180	157	2
NATURE	508	34	1
IEEE Explore	53	46	0
PubMed	17	4	0
MDPI	157	20	2
Google Scholar	5020	1549	8
SciELO	272	75	0
IEEE LatinAmerica	0	0	0
Total	9538	2114	23

presents a summary of the documents identified and selected during the literature search in various databases. In total, 2114 records were retrieved, distributed among databases such as ScienceDirect (104 documents), Springer (157), IEEE Xplore (46), and Google Scholar, which contributed the largest number with 1549 records. ScienceDirect contributed the highest number of included studies (9), followed by Google Scholar (8) and Springer and MDPI (2). Some databases, such as SciELO and IEEE Latin America, did not provide relevant studies for review. During the identification phase, specific search strategies were applied to ten databases and Google Scholar, using English and Spanish terms to maximize completeness. The details of the queries and Boolean combinations are presented in Appendix A 

Table A1. Detailed search results per query and database.

Database/Query	Date	Language	Results	Total
ScienceDirect
Hibiscus sabdariffa AND Artificial Intelligence	5 August 2025	English	58
Hibiscus sabdariffa AND Computer Vision	5 August 2025	English	23
Hibiscus sabdariffa AND Deep Learning	5 August 2025	English	74
Roselle AND Artificial Intelligence	5 August 2025	English	78
Roselle AND Computer Vision	5 August 2025	English	80
Roselle AND Deep Learning	5 August 2025	English	137
Total ScienceDirect				450
Taylor & Francis
Hibiscus sabdariffa AND Artificial Intelligence	5 August 2025	English	8
Hibiscus sabdariffa AND Computer Vision	5 August 2025	English	23
Hibiscus sabdariffa AND Deep Learning	5 August 2025	English	43
Roselle AND Artificial Intelligence	5 August 2025	English	86
Roselle AND Computer Vision	5 August 2025	English	399
Roselle AND Deep Learning	5 August 2025	English	1022
Total Taylor & Francis				1581
Springer
Hibiscus sabdariffa AND AI	5 August 2025	English		180
Wiley Online Library
Hibiscus sabdariffa AND Artificial Intelligence	5 August 2025	English	38
Hibiscus sabdariffa AND Computer Vision	5 August 2025	English	46
Hibiscus sabdariffa AND Deep Learning	5 August 2025	English	67
Roselle AND Artificial Intelligence	5 August 2025	English	108
Roselle AND Computer Vision	5 August 2025	English	305
Roselle AND Deep Learning	5 August 2025	English	736
Total Wiley				1300
Nature
Hibiscus	5 August 2025	English		508
IEEE Xplore
Hibiscus	5 August 2025	English		53
PubMed
Hibiscus sabdariffa AND Artificial Intelligence	5 August 2025	English	2
Hibiscus sabdariffa AND Computer Vision	5 August 2025	English	2
Hibiscus sabdariffa AND Deep Learning	5 August 2025	English	1
Roselle AND Artificial Intelligence	5 August 2025	English	7
Roselle AND Computer Vision	5 August 2025	English	2
Roselle AND Deep Learning	5 August 2025	English	3
Total PubMed				17
MDPI
Hibiscus sabdariffa AND Artificial Intelligence	5 August 2025	English	22
Hibiscus sabdariffa AND Computer Vision	5 August 2025	English	10
Hibiscus sabdariffa AND Deep Learning	5 August 2025	English	14
Roselle AND Artificial Intelligence	5 August 2025	English	47
Roselle AND Computer Vision	5 August 2025	English	22
Roselle AND Deep Learning	5 August 2025	English	42
Total MDPI				157
SciELO
Hibiscus OR Roselle OR Flor de Jamaica	5 August 2025	Spanish		272
IEEE Latin America
Roselle OR Hibiscus sabdariffa OR Flor de Jamaica	5 August 2025	Spanish		0
Google Scholar
(“Hibiscus sabdariffa” OR “Roselle”) AND (“Artificial Intelligence” OR “Machine Learning” OR “Deep Learning” OR “Computer Vision”)	5 August 2025	English + Spanish		5020
Grand Total				9538

, while the results obtained per database and the total number of records are summarized in

Table A2. Search strategies across all databases.

Database	Boolean Query	Date	Language	Filters	Notes
ScienceDirect	(“Hibiscus sabdariffa” OR ”Roselle”)AND (“Artificial Intelligence” OR “Computer Vision” OR “Deep Learning”)	5 August 2025	English	Language only	Full text
Taylor & Francis	(“Hibiscus sabdariffa” OR ”Roselle”) AND (“Artificial Intelligence” OR “Computer Vision” OR “Deep Learning”)	5 August 2025	English	Language only	Full text
Springer	(“Hibiscus sabdariffa”) AND (“AI”)	5 August 2025	English	Language only	Full text
Wiley Online Library	(“Hibiscus sabdariffa” OR ”Roselle”) AND (“Artificial Intelligence” OR “Computer Vision” OR “Deep Learning”)	5 August 2025	English	Language only	Full text
Nature	”Hibiscus”	5 August 2025	English	Language only	Simplified query
IEEE Xplore	”Hibiscus”	5 August 2025	English	Language only	Simplified query
PubMed	(”Hibiscus sabdariffa” OR ”Roselle”) AND (“Artificial Intelligence” OR “Computer Vision” OR “Deep Learning”)	5 August 2025	English	Language only	Full text
MDPI	(”Hibiscus sabdariffa” OR ”Roselle”) AND (“Artificial Intelligence” OR “Computer Vision” OR “Deep Learning”)	5 August 2025	English	Language only	Full text
SciELO	(”Hibiscus” OR ”Roselle” OR “Flor de Jamaica”)	5 August 2025	Spanish	Language only	Regional scope
IEEE Latin America	(”Roselle” OR ”Hibiscus sabdariffa” OR ”Flor de Jamaica”)	5 August 2025	Spanish	Language only	Regional scope
Google Scholar	(“Hibiscus sabdariffa” OR “Roselle”) AND (“Artificial Intelligence” OR “Machine Learning” OR “Deep Learning” OR “Computer Vision”)	5 August 2025	English + Spanish	Language only	Single query to minimize duplicates

.

Figure 2. PRISMA diagram showing the study selection process in the systematic review. A total of 2112 records were identified, of which 2029 were excluded during the initial screening and 60 after full-text review. In the end, 23 studies were included in the review.

Figure 2. PRISMA diagram showing the study selection process in the systematic review. A total of 2112 records were identified, of which 2029 were excluded during the initial screening and 60 after full-text review. In the end, 23 studies were included in the review.

Four researchers independently reviewed and reached a consensus on study selection without automation. The systematic review data, based on key results, are in

Table 4. Summary of primary outcomes extracted from the selected studies. The table defines key data fields used for analysis, including study characteristics, AI methodologies, data types, results metrics, and geographic context, to facilitate synthesis and comparison across studies.

Outcomes	Description
Year of publication	Analyze trends in the application of AI to H. sabdariffa.
Reference	Authors of the work to show trends by research groups and geographic areas.
Title	To easily locate the original research.
Objective	Clarifies whether the focus was pest detection, yield prediction, etc.
Methodology, AI techniques	Procedures and types of algorithms used.
Types of data	Specifies whether they are field data, satellite images, spectroscopy, among others.
Results based on accuracy	Metrics used in the studies and the results obtained.
Geographic location	May influence crop conditions or model generalization.

.

In addition to the primary results, the variables from the reviewed studies were collected and are presented in

Table 5. Summary of secondary outcomes extracted from the selected studies. The table defines additional contextual variables, including study source, preprocessing steps, real-world applicability, reported advantages, and limitations, to support qualitative analysis and interpretation of the evidence.

Outcomes	Description
Source of the study	Database where each article was found.
Reference	Authors of the work to show trends by research groups and geographic areas.
Preprocessing (PP)	Evaluate the quality, reproducibility, and robustness of the proposed methods.
Real-world application (RA)	Has it been tested in the field? Is it used by farmers? Does it require special infrastructure?
Study advantages	Positive aspects or benefits derived from the research that highlight its scientific, methodological, or practical contributions.
Study limitations	Challenges or constraints of the approaches.

. Reference and identification variables are used as the primary key to link Table 4 and Table 5.

Regarding missing or uncertain information, it was assumed that when studies did not report performance metrics, they were not available in the original articles. No estimations or imputations of missing data were performed.

In

Table 6. Summary of the systematic review evaluation based on PRISMA 2020 criteria, highlighting methodological choices, potential biases, and evidence certainty in light of the study’s characteristics and limitations.

Criterion	Description
Risk of bias assessment in individual studies	A formal risk of bias assessment was conducted using a modified JBI checklist comprising nine methodological criteria. This evaluation provided an objective measure of methodological quality across all included studies, where higher scores reflected lower bias risk. Study selection was carried out by nine authors through consensus, without the use of automated tools, ensuring adherence to predefined inclusion criteria.
Methods of synthesis	A narrative synthesis was conducted due to the heterogeneity among studies. They were grouped according to type of AI, data, and application. No missing data handling or meta-analysis was performed. Results were presented in tables and a PRISMA diagram. No heterogeneity or sensitivity analyses were conducted, but limitations were discussed qualitatively.
Publication bias assessment	A formal risk of bias assessment was conducted using a modified JBI checklist, and studies from multiple databases were included without restrictions by country or results, which helped to reduce potential bias. Grey literature was not considered.
Certainty of evidence assessment	A formal tool such as GRADE was not used, but qualitative criteria were applied uniformly, considering methodological clarity, consistency of metrics, and reproducibility of the studies.

, the systematic review addresses four essential aspects, as outlined in the PRISMA guidelines. A formal risk of bias assessment was performed using a modified JBI checklist with nine methodological criteria, allowing evaluation of methodological quality in the included studies. For synthesizing the results, a narrative strategy was used due to the substantial heterogeneity in AI methodologies and evaluation metrics, which precluded a meta-analysis.

4. Results

The results of this systematic review are presented through a narrative synthesis, given the heterogeneity of study designs, AI methods, data types, and outcome measures. Key findings, methodological trends, and limitations are summarized below.

The application of computer vision techniques in agriculture has resulted in substantial improvements in monitoring, analyzing, and optimizing various aspects of crop management. In the specific case of H. sabdariffa, several studies have employed computer vision and machine learning algorithms to address critical challenges, with the goal of increasing crop productivity and sustainability. In this systematic review, the most relevant studies that apply these methodologies were analyzed, evaluating their approaches, performance metrics, advantages, and limitations. Due to the heterogeneity of the methods used, the results are presented through a narrative synthesis in which the key characteristics of each study are tabulated and trends in the use of artificial intelligence for this crop are identified.

This section highlights the role of deep learning models, image processing techniques, and advanced data analysis methods that leverage deep learning. It also identifies research gaps and opportunities for future improvements in the application of computer vision to H. sabdariffa cultivation, which are detailed in Section 6.

Figure 3 illustrates the evolution of artificial intelligence methods applied to H. sabdariffa research from 2013 to 2025. Classical machine learning approaches, such as artificial neural networks (ANNs), partial least squares regression (PLS), and algorithms such as support vector machines and random forests, dominated early applications and still represent the majority of studies (approximately 70%). In contrast, deep learning methods, including convolutional neural networks such as ResNet, DenseNet, and VGG, have emerged more recently, accounting for around 30% of the reviewed works, with a noticeable increase after 2020 and significant adoption in 2024. This progression highlights a clear shift toward more robust, accurate and scalable AI solutions to address agronomic challenges in H. sabdariffa cultivation.

Figure 4 illustrates the geographical distribution of scientific publications applying artificial intelligence techniques to H. sabdariffa research. Research activity is concentrated in a few countries: Nigeria and India accounting for 17.4% of the publications, followed by Malaysia with 13% and Turkey, Indonesia, and Senegal with 8.7%. Other countries, such as China, Mexico, Saudi Arabia, Japan, Denmark, and Iraq, represent approximately 4.4% each. This distribution reveals a strong regional focus of AI applications in H. sabdariffa cultivation, primarily in tropical and subtropical regions, reflecting a growing global interest in the integration of emerging technologies for the sustainable development of this crop.

Figure 5 summarizes the characteristics of AI studies on H. sabdariffa using four pie charts. Panel (a) shows that structured/tabular data and images are the most frequently used inputs. Panel (b) indicates that ScienceDirect and Google Scholar comprise the main publication sources. Panel (c) highlights classification as the predominant task, with segmentation and detection also reported. Panel (d) shows that machine learning is the leading thematic area, followed by ML–DL combinations, with additional contributions from computational intelligence, computer vision, and deep learning.

Table 7. Part 1—Summary of primary outcomes from selected studies applying artificial intelligence to H. sabdariffa.

ID	Author (Year)	Country	Title	Objective	Methods	Dataset	Accuracy/Metrics
Database: ScienceDirect
9	Pushpa, B. R. et al. [102] (2024)	India	On the importance of integrating convolution features for Indian medicinal plant species classification using hierarchical machine learning approach	Develop a hierarchical classification model to categorize 100 species using feature fusion	Fusion model with hierarchical ML	Images	Validation: Acc = 93.96%, Test: Acc = 94.54%
4	Aydin, Ö. F. et al. [103] (2025)	Turkey	Smartphone-based app development with machine learning using Hibiscus sabdariffa extract for pH estimation	Develop a mobile solution for pH prediction based on colorimetric analysis	Random Forest, kNN, MLP	RGB images of extracts	RMSE = 0.12, R2 = 0.98
14	Huang, X. et al. [104] (2014)	China	Measurement of total anthocyanins content in flowering tea using near infrared spectroscopy combined with ant colony optimization models	Explore the feasibility of using NIR spectroscopy for rapid and non-destructive determination of total anthocyanin content in flowering tea	ACO-iPLS, GA-iPLS, iPLS, Full-spectrum PLS	Spectral	R = 0.9856; RMSECV = 0.1198 mg/g. ACO-iPLS performed best.
15	Horta-Velázquez, A. et al. [105] (2025)	Mexico	The optimal color space enables advantageous smartphone-based colorimetric sensing	Optimize the color space to improve smartphone-based colorimetric sensing, minimizing sensitivity to lighting variations	Region of Interest (ROI)	Images	DNR, RMSRE, and MAPE reported.
16	Adeyi, O. et al. [106] (2022)	Nigeria	Process integration for food colorant production from Hibiscus sabdariffa calyx: A case of multi-gene genetic programming (MGGP) model and techno-economics	Design an integrated HAE-BP process for crude anthocyanin powder (CAnysP), analyzing process variables and economic performance	MGGP	Structured data	APR: 99.98%, UPC: 98.47%.
3	Bankole, D. T. et al. [107] (2022)	Nigeria	Acid-activated Hibiscus sabdariffa seed pods biochar for the adsorption of Chloroquine phosphate: Prediction of adsorption efficiency via machine learning approach	The objective of the article is to investigate the adsorption of chloroquine phosphate onto acid-activated Hibiscus sabdariffa seed pod biochar and predict its efficiency using machine learning models	ANN	Structured data	($R^2$): 98.23%
11	Mavani, N. R. et al. [108] (2024)	Malaysia	Determining food safety in canned food using fuzzy logic based on sulphur dioxide, benzoic acid and sorbic acid concentration	Develop a fuzzy logic framework to determine the safety of canned food by evaluating preservative concentrations	Fuzzy logic, Mamdani inference	Structured data	R2 = 1.0000, MSE = 0.0007–0.0240, MAE = 0.0267–0.1150.
23	Periyappillai G. et al. [109] (2025)	India	Advanced ensemble machine learning prediction to enhance the accuracy of abrasive waterjet machining for biocomposites	Optimizar parámetros de Abrasive Water Jet Machining (AWJM) para composites de fibra de Roselle con cáscara de huevo, mejorando calidad, eficiencia y precisión	ANN, LSTM, RF, kNN combinados y RSM	Structured data	R2 = 0.9956 para SR, 0.9989 para MRR, 0.9680 para Kerf Angle; error absoluto promedio menor a 2%
Database: Nature
12	Laskar, Y. B. et al. [110] (2021)	India	Hibiscus sabdariffa anthocyanins as potential modulators of estrogen receptor alpha activity with Favourable Toxicology: A Computational Analysis Using Molecular Docking, ADME/Tox Prediction, 2D/3D QSAR and Molecular Dynamics Simulation	Determine anthocyanin activity and toxicology using computational techniques	PLS, PCR, kNN	Structural data	$q^2$ = 0.7043 (2D QSAR); 0.4106 and 0.4769 (3D QSAR).
18	Verma, T. N. et al. [111] (2021)	India, Saudi Arabia	Experimental and empirical investigation of a CI engine fuelled with blends of diesel and roselle biodiesel and Roselle Biodiesel	Analyze performance of roselle biodiesel in CI engines	ANN	Structured data	R = 0.9990 ± 0.0005, R2 = 0.9980 ± 0.0011

and

Table 8. Part 2—Summary of primary outcomes from selected studies applying artificial intelligence to H. sabdariffa.

ID	Author (Year)	Country	Title	Objective	Methods	Dataset	Accuracy/Metrics
Database: Springer
22	Ishola, N. B. et al. [112] (2019)	Nigeria	Process modeling and optimization of sorrel biodiesel synthesis using barium hydroxide as a base heterogeneous catalyst: appraisal of response surface methodology, neural network and neuro-fuzzy system	Model and optimize hibiscus oil conversion into hibiscus methyl esters (HSME)	RSM, ANN, ANFIS	Structured data	ANFIS: R2 = 0.9944; ANN: 0.9875; RSM: 0.9789.
7	Mustafa, M. S. et al. [113] (2020)	Malaysia	Development of automated hybrid intelligent system for herbs plant classification and early herbs plant disease detection	Develop system for classification and early disease detection in herbs	NB, PNN, SVM	Images	Hybrid method with FIS achieved 99.3% accuracy.
Database: MDPI
19	Tsuchitani, E. et al. [114] (2022)	Japan	Recording the Fragrance of 15 Types of Medicinal Herbs and Comparing Them by Similarity Using the Electronic Nose FF-2A	Record and compare the fragrance profile of 15 medicinal herbs using an electronic sensor	Ward’s method	Structured data (numerical)	–
21	Juhari, N. H. et al. [115] (2018)	Malaysia, Denmark	Physicochemical Properties and Oxidative Storage Stability of Milled Roselle (Hibiscus sabdariffa L.) Seeds	Analyze physicochemical properties and oxidative stability of seeds under different storage conditions	Principal Component Analysis (PCA)	Structured data (multivariate)	–
Database: Wiley Online Library
17	Omerogullari B. Z. et al. [116] (2023)	Turkey	Investigation and feed-forward neural network-based estimation of dyeing properties of air plasma treated wool fabric dyed with natural dye obtained from Hibiscus sabdariffa	Predict dyeing properties of wool fabric dyed with hibiscus extract using FFNN	Feed-Forward Neural Network (FFNN), Levenberg-Marquardt	Spectrophotometric data	R2: L = 0.95797, a = 0.95284, b = 0.96574, K/S = 0.94478.
Database: Google Scholar
1	Faye, P. et al. [117] (2024)	Senegal	Water Optimization in Digital Farming	Establish an agricultural calendar to optimize planning based on climatic variations, soil types, and irrigation systems	Decision Tree (DT)	Structured data (agro-climatic)	NR
6	Faye, P. et al. [118] (2024)	Senegal	Machine Learning for a Better Agriculture Calendar	Propose an AI-based agricultural calendar adaptable to climate change	KMeans	Structured data (agro-climatic, soil)	Elbow Method selected 4 clusters.
5	Reddy, M. et al. [119] (2015)	India	Assessing Climate Suitability for Sustainable Vegetable Roselle (Hibiscus sabdariffa) Cultivation in India Using MaxEnt Model	Evaluate climate suitability for sustainable cultivation of roselle using MaxEnt	Maximum Entropy (MaxEnt)	Structured data (georeferenced bioclimatic)	AUC: 0.993 (training), 0.992 (testing)
8	Bastian, A. et al. [120] (2023)	Indonesia	Roselle Pest Detection and Classification Using Threshold and Template Matching	Design a system to detect and classify pests in roselle plants to reduce crop failure risk	Thresholding, Template Matching	Images	Accuracy: 75%
13	Qader, H. et al. [121] (2013)	Iraq	Simultaneous Spectrophotometric Determination of Binary and Ternary Mixtures Using Chemometric Techniques	Apply spectrophotometric and AI methods for compound determination in roselle extracts	ANN, Partial Least Squares (PLS)	Spectrophotometric	Recovery: 94.31% (GA), 93.21% (AA)
10	Musyaffa, M. et al. [122] (2024)	Indonesia	IndoHerb: Indonesian Medicinal Plants Recognition Using Transfer Learning and deep learning	Identify medicinal plants using transfer learning CNNs	ResNet, DenseNet, VGG, ConvNeXt, Swin Transformer	Images	ConvNeXt: 92.5% accuracy
20	Ilomuanya, M. et al. [123] (2020)	Nigeria	Development and Optimization of Antioxidant Polyherbal Cream Using Artificial Neural Network Aided Response Surface Methodology	Develop and optimize a polyherbal cream with hibiscus extract	MFFF, MNFF	Structured data (numerical)	Viscosity: 8355.86 cP, Spreadability: 47.01%, Particle size: 0.12 µm
2	Bankole, D. T. et al. [124] (2022)	Nigeria	Modeling and Adsorption Studies of Selected Pharmaceuticals and Food Dyes Onto Chemically Modified Agrowaste Biomass	Predict adsorption efficiency using hibiscus-based adsorbents	ANN	Structured data	MSE: 4.43, R2 = 0.9906; Removal: 95.78–98.79% (BSP1-CQP, EGB1-CQP)

present the primary outcomes for each study. The tables show nine different outcomes. First, the ID, which corresponds to the unique identifier of the study, allows the primary outcomes to be located and cross-referenced with the secondary outcomes; the year of publication and reference; the methodology used in the study (AI algorithms only); the type of data used in the study; the results obtained; the title of the study; the objective; and the country of the authors. Finally, each study is grouped under the publishers, which—after applying the inclusion criteria—were reduced to seven: Google Scholar, ScienceDirect, Springer, MDPI, Nature, Taylor and Francis, and Wiley Online Library.

Table 9. Part 1—Summary of secondary outcomes from selected studies applying AI to H. sabdariffa: preprocessing, real-world application, advantages, and limitations.

ID	Authors (Year)	PP	AR	Advantages	Limitations
Water Resource Management
1	Faye, P. et al. [117]	✓	–	Interpretability, ability to handle non-linear relationships, fast training without the need for normalization.	Overfitting, sensitivity to correlated data, and lower predictive capability compared to more advanced models.
2	Bankole, D. T. et al. [124]	✓	–	Use of agricultural waste as adsorbents for water purification; the adsorption process demonstrated in the study is easy to operate and shows high efficiency in contaminant removal.	The study is based on laboratory tests; despite exploring various experimental conditions, the complexity of interactions in natural environments may limit applicability of the results in multicomponent systems.
3	Bankole, D. T. et al. [107]	✓	–	An ecological and cost-effective approach using agricultural waste, with high predictive reliability through artificial intelligence, facilitating the design of water treatment processes.	Validation under real conditions has not yet been conducted, and practical aspects such as adsorbent regeneration and large-scale cost considerations remain unaddressed.
Soil Quality Monitoring
4	Aydın, Ö. F. et al. [103]	✕	✕	Innovative and accessible natural indicator, minimal dependence on traditional costly and complex techniques, integration of image processing and AI algorithms, smartphone-based pH measurement.	Model sensitivity to color and lighting variations; algorithm performance depends on image quality; limited generalization.
Climate Change Effects
5	Reddy, M. et al. [119]	✓	✓	Provides accurate predictions on roselle distribution, efficiently handles scarce data, identifies potentially suitable cultivation areas. The methodology is globally applicable and generates climate suitability maps in regions lacking precise crop data.	The model may not capture all relevant factors, and precise coordinates of occurrences may not be available in all contexts. Could also contribute to pest and disease control and soil quality monitoring.
6	Faye, P. et al. [118]	✓	✓	Groups regions with similar characteristics, identifies patterns in temperature and UV radiation; scalability makes it ideal for classifying areas by productivity and needs; improves agricultural management and decision-making.	Lack of mechanization and climate dependency are limitations.
Pest and Disease Control
7	Mustafa, M. S. et al. [113]	✓	–	Hybrid system, automated approach, speed and precision of the process. Combines computer vision with odor analysis; practical for field use.	High computational time due to integration of multiple algorithms, dependency on expensive equipment, sensitivity to environmental conditions, which may affect detection accuracy.
8	Bastian, A. et al. [120]	✓	✓	Implementation of a pest detection and classification system, includes real-time notification technology via IoT, allowing farmers to receive rapid alerts about pest presence.	Image quality, lighting conditions, lack of complete and well-designed datasets for roselle pests, generalization and classification of all pest types.
9	Pushpa, B. R. et al. [102]	✓	✓	Innovative hierarchical classification approach, improves accuracy, relevant dataset accessible via mobile application.	Model generalization, real-world conditions, limited number of species. Also related to emerging diseases.
10	Musyaffa, M. et al. [122]	✓	–	Innovative methodology, accurate identification of medicinal plants in Indonesia, creation of a valuable dataset and achievement of high accuracy levels, contributes to ethnobotanical knowledge preservation.	Dependent on the quality and characteristics of the dataset.

and

Table 10. Part 2—Summary of secondary outcomes from selected studies applying AI to H. sabdariffa: preprocessing, real-world application, advantages, and limitations.

ID	Authors (Year)	PP	AR	Advantages	Limitations
11	Mavani, N. R. et al. [108]	✓	–	Develops a fuzzy logic framework enabling fast and accurate assessment of food safety in canned products, reduces time and resources needed to verify food compliance with safety regulations.	Covers five food categories; thus, it cannot be applied to a wider range of products; depends on the availability of laboratory data.
Anthocyanins Monitoring
12	Laskar, Y. B. et al. [110]	✓	–	Innovative methodological approach, combines computational techniques, systematic analysis, development of new modulators, contributing to cancer therapy research.	Based on simulations and predictions, lacks experimental validation.
13	Qader, H. et al. [121]	✓	–	Fast and effective approach, uses environmental chemistry techniques, minimizes use of toxic solvents; ANN models demonstrate good precision and recovery.	Requires further optimization regarding learning rate and variable selection, results based on samples under controlled conditions.
14	Huang, X. et al. [104]	✕	✓	NIR spectroscopy combined with ACO-iPLS allows rapid, simple, non-destructive determination, reducing time and cost.	Lacks field studies validating the method’s effectiveness; methodology may have technical requirements that limit its implementation in broader production environments.
15	Horta-Velázquez, A. et al. [105]	✓	✓	Smartphone-based technologies enable accessible and reliable colorimetric analysis, with a wider measurement range than traditional methods and greater resistance to lighting variations thanks to the use of the CIELAB color space.	Variability between smartphone models affects reproducibility of the method; despite color corrections, lighting dependency and potential inaccuracy at high concentrations limit general applicability.
16	Adeyi, O. et al. [106]	✓	–	Innovative approach to optimize anthocyanin production, high accuracy in techno-economic predictions, comprehensive data analysis, and a scalable process design applicable to the natural colorant industry.	Dependence on experimental data and simulations limits direct field applicability as it does not address practical implementation in real-world environments. Also related to pest and disease control and soil quality monitoring.
17	Omerogullari Basyigit, Z. et al. [116]	✓	✓	The use of natural dyes from H. sabdariffa combined with plasma treatment offers a more eco-friendly and efficient process, reduces dependence on synthetic chemicals, and improves dye property estimation, promoting sustainable practices in the textile industry.	Requires specialized infrastructure for plasma treatment. Also related to climate change effects and water resource management.
By-Product Utilization
18	Verma, T. N. et al. [111]	✓	✕	Innovative approach to using roselle biodiesel, technical feasibility in compression ignition engines, artificial neural network model to predict engine behavior.	Based on laboratory tests, economic implications and availability of necessary infrastructure are not discussed.
19	Tsuchitani, E. et al. [114]	✓	✓	Innovative methodology for objective and quantitative characterization of herb aromas, ability to generate useful data for AI models, advantages over traditional methods.	The approach is limited to aroma analysis, does not consider other key factors such as chemical content or organoleptic properties, and reliance on specialized equipment may restrict applicability in rural or resource-limited settings.
20	Ilomuanya, M. et al. [123]	✓	✕	Use of ANN, significant results for antioxidant properties, positive correlation between formulation variables and expected outcomes, robust and efficient evaluation of ingredient interactions.	Study is laboratory-focused; generalization of results to practical situations is not addressed.
21	Juhari, N. H. et al. [115]	✕	✕	Detailed analysis of physicochemical properties, identifies optimal storage conditions to maximize seed stability and quality.	Does not consider practical use or adoption of the seeds by farmers and omits external factors that may affect quality during real-world storage.
22	Ishola, N. B. et al. [112]	✓	✓	Advanced modeling methods provide high accuracy in biodiesel yield prediction; biodiesel yield is favorable.	Practical applicability is not discussed; no mention of real-world implementation or use by farmers; the process may require infrastructure and costs that limit adoption at small scale or by local producers.
23	Periyappillai G. et al. [109]	–	✕	El trabajo demuestra un uso efectivo de modelos ensemble de machine learning para predecir parámetros claves de AWJM, logrando alta precisión y fiabilidad, lo cual mejora la eficiencia operativa y la calidad del producto en la fabricación de composites. Además, se identifican los parámetros con mayor influencia, facilitando la optimización del proceso.	El estudio se limita a datos experimentales controlados y no es claro si sus modelos han sido validados en operación real o por usuarios finales como agricultores o industrias en campo, lo que podría limitar la aplicabilidad práctica inmediata.

present a summary of studies related to the use of artificial intelligence in H. sabdariffa cultivation, highlighting their advantages, limitations, data preprocessing, field testing, or use by farmers or requiring special infrastructure, as well as their authors. Four of the twenty-one studies may also be associated with other challenges or issues ID 4, 8, 15, and 16.

The synthesis of results was conducted using a narrative approach as the included studies exhibited substantial heterogeneity in terms of methodologies, data types, and outcomes. No meta-analysis or formal statistical pooling was performed. While a systematic assessment of risk of bias and sources of heterogeneity was not conducted, key methodological differences were identified, and major limitations were qualitatively discussed. Sensitivity analyses were also not performed.

No meta-analyses or quantitative syntheses were conducted in this systematic review; therefore, formal assessments of risk of bias due to missing results were not performed. However, to minimize publication bias, studies from multiple databases were included, with no restrictions based on country, language, or type of reported outcome.

No formal confidence levels were assigned to the results. Nevertheless, qualitative criteria such as methodological clarity, the use of quantifiable metrics, and study reproducibility were applied to assess the overall robustness of the findings.

In [125], it is noted that operational decisions in agriculture are usually planned for the short term, while structural transformations and R&D strategies require multi-year or even long-term planning. Figure 6 shows the six strategic research lines in H. sabdariffa, classified according to their applicability in the short, medium, and long term.

Table 11. Part 1—Systematic comparison of region, data type, resolution, training set size, and main metrics used in the studies.

ID	Region	Data Type	Resolution	Training Set Size	Accuracy or Main Metric
1	Tropical (Senegal)	Real-time sensory data (moisture, temperature, infiltration)	Not applicable	Periodic measurements at 5 cm, 2 months	No metrics; use of decision trees and ML
2	Tropical (Nigeria)	Experimental lab data (adsorption)	Not specified	Varies: BSP-CQP (916, 870, 1061)	R2 > 0.98 (Logsig: 0.9854)
3	Tropical (Nigeria)	Experimental lab data	Not specified	1180 (train) + 252 (test) + 252 (validation)	R2: 0.998/MSE: 8.01
4	Mediterranean/Temperate (Turkey)	Images of indicator solutions	Standard 200 × 200 px	94 samples (75 for training)	MAE = 4.65–9.33%, CVRMSE = 6–7%, RMSE = 3.94–10.8%
5	Tropical/Subtropical (India)	Climate data (WorldClim)	Based on WorldClim products	23 points (75% for training)	AUC: 0.993 (train), 0.992 (test)
6	Tropical (Senegal)	Agro-meteorological and agroecological data	Variable	Historical data and official databases	Qualitative accuracy in agricultural calendar prediction
7	Tropical (Malaysia)	Leaf images + odor data (electronic nose)	Not specified	1000 samples (10 species)	Accuracy: 97–98% vision; 96–98% odor
8	Tropical (Indonesia)	Images with pests and diseases	Not specified	Proprietary dataset, no exact count	Accuracy: 75% pest detection
9	Tropical/Subtropical (India)	Medicinal plant images	3120 × 4160 px	13,536 images (100 species)	Accuracy: 94.54% (GSL100)
10	Tropical (Indonesia)	Medicinal plant images	128 × 128 px	12,000 images + augmentation	Accuracy: 92.5% (ConvNeXt)
11	Tropical (Malaysia)	Preservative concentration data (SD, BA, SA)	Not specified	50 values + industrial samples	MSE and MAE ≈ 0; high R2

and

Table 12. Part 2—Systematic comparison of region, data type, resolution, training set size, and main metrics used in the studies.

ID	Region	Data Type	Resolution	Training Set Size	Accuracy or Main Metric
12	Tropical/Subtropical (India)	Computational data (docking, ADME, Tox, QSAR)	Varied (depending on method: 2D, 3D QSAR)	30 (number of compounds used for QSAR models)	$r^2$ (above 0.79 in QSAR models) and $q^2$ (around 0.41–0.48)
13	Arid/Desert (Iraq)	Spectrophotometry data and AI models	High resolution (spectrophotometry)	Not specified (typically small to moderate in similar studies)	RMSE (e.g., 2.20% for salicylate and thiocyanate)
14	Subtropical (China)	NIR spectra	10,000–4000 cm−1	120 samples (72 for calibration + 48 for prediction) + 60 independent tests	R = 0.9856, RMSECV = 0.1198 mg/g
15	Tropical/Subtropical (Mexico)	Color data (RGB, HSV, CIELAB images)	Not numerically specified; smartphone images under varying lighting conditions	Not exactly specified; pigment data in dilutions	Accuracy in pigment concentration estimation (compared with spectrophotometry, LOOCV)
16	Tropical (Nigeria)	Experimental data and simulations	Not specified	Not specified (lab study and predictive models)	R2 = 0.9643 for CAnysP APR and R2 = 0.9847 for UPC
17	Mediterranean/Temperate (Turkey)	Spectrophotometric data (CIELab and K/S values)	Not applicable/spectrophotometry	Not specified (experimental dataset)	R2 (regression): 0.94478 to 0.95797
18	Tropical/Subtropical (India)	Combustion and emission experimental data	Not specified	84 conditions (70%) for training, 18 (15%) for validation, 18 (15%) for testing	r = 0.9996 (Pearson correlation for BTE); r-squared = 0.9980 ± 0.0011
19	Subtropical (Southern Japan)	Aroma data measured by electronic sensors	Not specified (10-dimensional measurements over time)	Not specified (15 herbs and several tea and wine samples)	No explicit accuracy metric reported; similarity and correlation evaluated
20	Tropical (Nigeria)	Cream formulation and response data	Resolution not indicated	Not specified, 70% of data for training, 15% for validation, 15% for testing	R2 (coefficient of determination), RMSE, AAD, MAD
21	Tropical (Malaysia)	Volatile profiles (GC-MS) and statistical analysis	PCA: numerical resolution not specified, based on peak areas	Not specified	Variance explained by PC1 (48%) and PC2 (14%)
22	Tropical (Nigeria)	Process data (transesterification conditions)	Not specified	Not specified (experimental data)	R2 (coefficient of determination) for models: ANFIS = 0.9944, ANN = 0.9875, RSM = 0.9789
23	Tropical/Subtropical (India)	Datos experimentales AWJM de composites Roselle	Not specified	27 samples	R2 ajustado 0.9984 para MRR, Error promedio relativo en predicción <2%

present a systematic comparison that synthesizes the most relevant methodological and contextual aspects of the 21 analyzed studies, considering key factors such as geographic region, type of data used, data resolution, training set size, and reported evaluation metrics. A considerable diversity is observed in the sources of information employed—ranging from spectrophotometric data and environmental sensors to RGB images and chemical profiles—as well as in data resolutions and sample sizes, reflecting the heterogeneity of approaches in the use of artificial intelligence applied to the cultivation or post-harvest of H. sabdariffa. This methodological and regional diversity highlights not only the broad range of potential applications but also the complexity of establishing direct comparisons between models or generalizing their results to other agroclimatic regions. In particular, differences between tropical and subtropical zones may affect model transferability due to variations in climate, soil, and cultivation conditions. Therefore, it is recommended to consider cross-validation tests or region-specific adjustments in future developments. Overall, the information presented enables the visualization of relevant patterns, contrasts, and limitations in current approaches, providing a systematic foundation for identifying areas of opportunity in future research.

A formal risk of bias assessment was conducted using a modified JBI checklist with nine methodological criteria, generating both individual scores and aggregated visualizations (Figure 7 and Figure 8). This approach allowed the identification of methodological strengths and weaknesses across studies, where higher scores indicate lower bias risk. Figure 7 presents the risk of bias assessment for the 21 studies included in this review using a modified JBI checklist. Each study was evaluated across nine methodological criteria, scored as 1 (yes), 0.5 (partial), or 0 (no). The global risk of bias was calculated as the average of all domain scores and categorized as Low (green), Moderate (yellow), or High (red). This approach provides a concise yet informative overview of the methodological quality and reliability of the evidence base. The nine domains evaluated were (1) clearly defined objectives, (2) appropriate study design, (3) robust sampling strategy, (4) valid measurement tools, (5) detailed data analysis, (6) ethical considerations, (7) reproducibility, (8) relevance of findings, and (9) clarity in conclusions.

Table 13. RoB and average score for each included study. RoB indicates the risk of methodological bias, while Mean corresponds to the average compliance score across nine evaluated criteria.

ID	1	2	3	4	5	6	7	8	9	10	11
RoB	0.444	0.222	0.278	0.333	0.444	0.222	0.333	0.278	0.167	0.111	0.111
Mean	0.556	0.778	0.722	0.667	0.556	0.778	0.667	0.722	0.833	0.889	0.889
ID	12	13	14	15	16	17	18	19	20	21	22
RoB	0.000	0.278	0.167	0.222	0.167	0.278	0.167	0.278	0.222	0.222	0.444
Mean	1.000	0.722	0.833	0.778	0.833	0.722	0.833	0.722	0.778	0.778	0.556
ID	23
RoB	0.111
Mean	0.889

presents the average methodological risk of bias (RoB) and model accuracy by country. The RoB value was computed as the mean of the individual study risks, where each study’s risk was derived from the proportion of unmet criteria across nine items adapted from the JBI checklist for scoping reviews. A higher RoB value indicates a greater potential for methodological bias, while lower values suggest stronger adherence to quality standards. This representation enables a combined analysis of methodological reliability and the technical performance of AI models reported in the included studies. For interpretability, RoB values were categorized as follows: low risk (RoB < 0.20), moderate risk (0.20 ≤ RoB ≤ 0.40), and high risk (RoB > 0.40).

Additionally,

Table 14. Average RoB and accuracy by country. RoB values indicate methodological reliability; lower values suggest lower bias, while Accuracy represents the mean performance reported across studies.

Country	India	Indonesia	Mexico	Nigeria	Turkey	Senegal	Malaysia	Japan	Iraq	China
Risk of Bias	0.185	0.194	0.222	0.267	0.306	0.444	0.222	0.278	0.278	0.167
Accuracy (%)	81.48	80.56	77.78	73.33	69.44	55.56	77.78	72.22	72.22	83.33

presents the average methodological RoB and the reported model accuracy by country. China shows the lowest bias level (RoB = 0.167) and the highest accuracy (83.33%), while Senegal and Turkey exhibit the highest bias values (0.333 and 0.306, respectively) and the lowest model performance (66.67% and 69.44%). These results suggest that studies conducted in China, India, and Indonesia demonstrate stronger methodological compliance, whereas some countries may require improvements in study design and reporting quality.

Figure 8 illustrates the geographical distribution of studies applying AI techniques to H. sabdariffa research, considering only countries located in tropical and subtropical regions to ensure a more accurate bias analysis. Although the review initially included publications from non-tropical regions, these were excluded from the visualization since their environmental and agronomic conditions differ significantly from those of primary production areas.

In this map, darker shades represent higher reported model accuracy, while the labels display both the accuracy percentage and the corresponding risk of bias score. Lower RoB values indicate a lower risk of bias and thus higher methodological quality, while higher RoB values suggest a greater potential for methodological bias. For instance, China reported the highest accuracy (83.3%) with a low RoB (0.17), while Senegal exhibits the highest bias level (RoB = 0.33) and the lowest model performance (66.7%). This visualization facilitates a combined interpretation of methodological rigor and technical performance across different regions.

5. Discussion

This review shows most studies use classical machine learning like ANN, PCA, k-means, decision trees, and SVM for agricultural tasks H. sabdariffa [102,106,112,114,119,126]. Deep learning use is limited despite success in other crops like maize, rice, and tomato [113,122,127,128]. Advanced architectures like ResNet and VGG are rarely used in studies H. sabdariffa, highlighting a technological gap compared to the rapid deep learning advancements in agriculture [128]. Despite rising global demand for H. sabdariffa, deep learning application to this crop is minimal. Challenges include the need for technical expertise, computational resources, and interdisciplinary collaboration in hibiscus-producing regions, which still lean on traditional methods.

A key limitation of the reviewed evidence is the limited availability of public data and reproducible code [120,121,124]. As shown in the tables, most studies do not report the availability of datasets or source code, which hinders comparison, replication, and cross-validation of results. Moreover, a significant portion of the articles reviewed come from sources such as Google Scholar, while the representation in high impact journals is lower, which may be related to access restrictions or publication bias [121,122,124]. Furthermore, many studies do not report standardized accuracy metrics or formal statistical analyses, which limits objective assessment of the performance of the implemented models.

This systematic review, while conducted with rigorous criteria, has methodological limitations. It lacked a formal risk of bias assessment and sensitivity analysis due to the diverse nature of included studies. Although multiple databases were searched, gray literature and pre-prints were not included [121,122]. The focus on English and Spanish publications might have excluded studies in other languages. These limitations should be considered when assessing the conclusions.

The review highlights the use of advanced deep learning for disease detection, quality classification, and yield prediction in cultivation H. sabdariffa [103,106,111,113,127]. This enhances decision-support accuracy for producers and technicians [127]. Policies should promote interdisciplinary research combining computer vision, deep learning, and agricultural science [126,128,129], while prioritizing standardized data sharing. Future studies should test models in real conditions and report results with consistent metrics.

The modified JBI checklist revealed that most studies had moderate to high methodological risk, with main issues being the lack of cross-validation, inadequate reporting of sample size justification, and unclear data selection and processing methods. Few studies showed strong compliance in all areas. This highlights the need to standardize evaluation and validation in future research on H. sabdariffa. Figure 7 visually summarizes these findings.

Most studies on H. sabdariffa are located in tropical and subtropical regions, matching the species’ natural zones. This regional focus restricts model transferability to other agroclimatic areas due to factors like soil variability, climate, and local production systems affecting AI model performance and generalizability. Figure 8 shows regional differences through accuracy metrics and average methodological risk per country. These findings highlight the need for models adaptable to diverse ecological and economic contexts.

This review deliberately concentrated on H. sabdariffa to provide a structured and in-depth synthesis of AI applications for this high-value crop. However, we recognize the importance of situating these findings within a broader agricultural context. Future research will extend this analysis to include related varieties, such as Hibiscus, where similar AI methodologies have been applied, enabling cross-crop comparisons and the identification of scalable and climate-resilient solutions.

6. Challenges and Future Perspectives in the Cultivation and Post-Harvest of Hibiscus sabdariffa

In light of current evidence and identified gaps, the following section outlines critical challenges, potential opportunities, and strategic directions for future research and innovation in H. sabdariffa cultivation.

The cultivation and post-harvest of H. sabdariffa faces various technological, climatic, and agricultural adoption limitations that restrict its productivity and global expansion. These barriers are particularly critical outside tropical and subtropical regions, where climatic conditions hinder optimal crop development. This section discusses the main current challenges, emerging trends, and future research directions aimed at advancing towards more sustainable and intelligent crop management.

One of the main challenges for the sustainability of H. sabdariffa post-harvest or cultivation is the efficient use of water resources. Implementing intelligent systems that integrate computer vision, sensors, and predictive models would enable early detection of water stress, precise estimation of irrigation needs, and diagnosis of diseases associated with water use. These solutions should be incorporated into precision agriculture frameworks with a sustainable focus, taking into account adaptation to tropical climates and resilience to climate change scenarios.

Based on the conducted review, six long-term priority research lines have been identified: (1) development of genetically modified microbial strains and transgenic varieties to enhance the synthesis of anthocyanins and procyanidins; (2) design of AI-assisted automated systems for the extraction, purification, and characterization of bioactive compounds; (3) metabolic studies in model organisms to understand the absorption kinetics of these compounds; (4) isolation and identification of gut microflora involved in their metabolism; (5) identification of secondary metabolites with beneficial effects on human health; and (6) applied research for the development of functional foods, supplements, and clinical products derived from H. sabdariffa.

The use of advanced technologies such as deep learning, multispectral imaging, convolutional neural networks, and IoT platforms offers a promising path toward digital and sustainable agriculture for H. sabdariffa. Integrating these tools can optimize monitoring, improve resource use efficiency, and increase crop yields. It is necessary to promote public policies that encourage the adoption of these technologies in producing regions, as well as foster collaboration among research institutions, producers, and technology developers to facilitate knowledge transfer and the practical implementation of innovative solutions.

7. Conclusions

This review outlines the current state of AI in the cultivation and post-harvest of H. sabdariffa, a valuable plant with underutilized technological potential. Most studies use classical machine learning techniques, with little use of advanced deep learning models, despite their known effectiveness in precision agriculture. Challenges include a lack of open datasets, limited reproducible code, diverse evaluation metrics, and no standardized methods. Nevertheless, there are significant opportunities to enhance the yield, sustainability, and resilience of H. sabdariffa cultivation through remote sensors, deep neural networks, and IoT platforms. These tools can aid in water stress monitoring, disease detection, and resource optimization, fitting into sustainable digital agriculture, particularly in tropical and subtropical areas where this plant grows.

Adopting these approaches offers both technical advancements and strategic opportunities to align H. sabdariffa production with modern demands for efficiency, traceability, and climate adaptation. Developing replicable models, open data sharing, and interdisciplinary collaboration are crucial for impactful future research. The integration of artificial intelligence, sustainable agriculture, and biotechnology is key to transforming H. sabdariffa crop management into a smarter, resilient, and competitive model. A total of 23 relevant publications were reviewed: 16 peer-reviewed journal articles ensuring rigor, two conference papers highlighting recent developments, two high-potential preprints, and one undergraduate thesis providing academic insight.