To develop the macro-process (Figure 2), we consider the six steps for conducting a systematic review as established in the methodology proposed by Kitchenham (2004). The steps are the planning phase (SR-0), research identification (SR-1), primary study selection (SR-2), study quality assessment (SR-3), the relevant information is extracted from the preliminary studies (SR-4), and synthesis of the results found in the primary studies (SR-5). In the SR-0 phase, research questions and protocol design are established. In SR-1, the search strategy for the systematic review is generated, publication bias is identified, the bibliography management process is determined, and the search documentation mechanism is established. Additionally, in SR-2, inclusion and exclusion criteria are set for study selection. In SR-3, quality thresholds are defined, and instruments for their assessment are designed. In SR-4, relevant information is extracted from the primary studies; the formats established in the review planning are utilized to achieve this. Finally, in SR-5, a synthesis of the results found in the prior studies is carried out for a case study. The extracted information is tabulated in a way that consistently answers the research questions posed in the previous stages.

According to the review objectives, we present the plan to build that below.

Considering the information extracted, the development of the MF is constructed following the Business Process Notation and Modeling - BPMN. The Bizagi software (Bizagi, 2020) is employed to achieve this.

For the evaluation of the MF, the expert weighting method was employed, which considers the following steps under Ishizaka and Nemery (2013):

Applying the protocol outlined in Figure 2 and considering the elements established in the systematic review planning, articles about knowledge management in agriculture were assessed between 2013 and 2023. A total of 481 articles were initially identified, resulting in a final count of 37 articles after removing duplicates, applying the defined exclusion and inclusion criteria, and conducting a quality assessment of the studies (Figure 3). This structure conforms to the requirements for an indexed journal submission.

Following the data extraction process, various types of knowledge, extraction methods, representation methods, their limitations, central areas of application, and the tools employed were identified. Regarding knowledge representation methods, it was observed that 40.4% of the studies utilized knowledge graphs, followed by ontologies at 34.6% and production rules at 25% (Figure 4).

On the other hand, Table 1 identifies the techniques employed in the data extraction process, noting the utilization of manual procedures such as interviews or the application of surveys with experts, alongside Natural Language Processing (NLP) techniques oriented toward entity recognition and relation extraction in unstructured data. Some of the tools employed for knowledge extraction and representation were also identified. There was a notable prevalence of the “Web Ontology Language - OWL,” used for knowledge representation in the Semantic Web, and the rule-oriented programming language CLIPS or one of its adaptations, such as Jess Rule, for knowledge representation through rules. Furthermore, in knowledge graphs, the Resource Description Framework (RDF) was identified as the primary means of representation. Additionally, the query language SPARQL was highlighted as essential for accessing and extracting information from RDF datasets.

Additionally, the main areas of intervention within the field of agriculture were identified, with pest and disease management accounting for 53.8%, comprehensive crop management at 19.2%, and nutritional management at 11.5% (Figure 5).

Regarding the data structure, 75% of the articles contemplate using unstructured data, encompassing text, images, audio, and video. 39% consider semi-structured data, and 12% pertain to structured data. Furthermore, the two types of knowledge considered in the knowledge management process were identified, with explicit knowledge comprising 92% of the studies and tacit knowledge accounting for 36% (Figure 6).

Finally, Table 2 presents some of the limitations of knowledge representation methods. At a general level, limitations were identified, such as the size of the knowledge base, the impact of the quality of input data on the reliability of the represented knowledge, the specificity of knowledge, which constrains its scalability, and the high requirement of experts for the creation and updating of the knowledge base. In ontologies, resistance may arise from formalizing specific agricultural domain knowledge, highlighting the challenge of representing knowledge with spatiotemporal characteristics.

Based on the information gathered during the systematic review process, the Knowledge Management Framework (MF) was proposed for subsequent integration into data analytics. Initially, the MF was proposed using the flow diagram standard, and subsequently, the refined process involved applying the Business Process Model and Notation (BPMN).

Knowledge Characterization and Knowledge Extraction (KC and KE): in the initial phase of the proposed MF, the characterization process of the data scarcity issue was considered, along with an assessment of the required knowledge type and the identification of available knowledge sources. These sources may contain either implicit or explicit knowledge. Therefore, a selection process was defined through a gate establishing an inclusive flow, meaning that both types may be found within the same knowledge source.

In cases where the source contains tacit knowledge, an elicitation process was outlined to extract unstructured data, which is subsequently stored in a data repository. Next, an activity was defined to extract implicit knowledge from the unstructured data, utilizing the identified extraction methods. These methods align with the Natural Language Processing techniques described in Table 1 and any others that may be placed. This process yields semi-structured or structured data. In the event of semi-structured data, a normalization and transformation process were established to convert it into structured data. Conversely, if the extraction yields structured data, it is directly stored in a data repository.

Finally, in cases where the assessed knowledge is explicit, the type of data structure to be processed was determined through an exclusive gate. Depending on the structure, the same processes defined earlier will be followed.

Knowledge Representation (KR): in the second phase, corresponding to the knowledge representation process, the selection of the knowledge representation method was defined. This decision was informed by the data repository containing the various representation methods and their respective limitations (Table 2). These limitations were identified for each method (Table 2). Subsequently, the representation method was implemented, considering the data object containing the available tools (Table 1). Following this, an exclusive gate was set up to evaluate the response of the knowledge representation method to the defined data scarcity issue. Should the implementation of the representation method appropriately address the problem, the represented knowledge is then stored in a data repository.

Knowledge Integration (KI): in the third phase of the MF, corresponding to the process of knowledge integration in data modeling, the task of integrating the knowledge represented in one or more phases of data modeling was established. This selection will depend on the model optimization objectives. This task makes use of the data warehouse containing the represented knowledge. Similarly, the knowledge integration process is defined by an inclusive gate, allowing the integration of knowledge in one or more phases simultaneously. In this sense, the represented knowledge can facilitate business or data understanding, support the data preparation process, optimize the modeling process, or support the evaluation process of the generated models. Following the model evaluation, compliance with the established requirements for the model was defined through an exclusive gate to proceed with its deployment or iterate the evaluation process of the integration phase(s).

Finally, in Figure 7, the Methodological Framework is presented, articulating the three phases that support the knowledge management process and its integration into data analytics models.

In this phase of the MF development process, the framework underwent evaluation and refinement conducted by four experts in the field of data analytics. Eight evaluation criteria were employed, encompassing aspects of clarity and comprehensibility, relevance and pertinence, adaptability and flexibility, and feasibility of implementation.

Following the evaluation conducted by experts (Figure 8), the highest weighted scores were assigned to criteria 3 (Lyubchich et al., 2019), 4 (4.4), 5 (4.4), and 8 (4.4), reflecting aspects of adaptability, flexibility, relevance, and reliability in implementation. On the other hand, criteria 1 (3.4), 2 (3.4), 6 (3.2), and 7 (3.8) yielded lower averages, although not falling below the mean evaluation level. These criteria are associated with the clarity, comprehensibility, and relevance of the Methodological Framework (MF). The lowest weighted score was attributed to criterion 6, which pertains to identifying the advantages and benefits of applying MF in data analytics. Furthermore, considering the indicators used to evaluate the consistency among experts, the ICCa was satisfactory, with a value of 0.41. Additionally, the average Spearman coefficient among all experts was 0.85, indicating a high level of concordance.

Furthermore, in addition to the assigned rating for each established criterion, the experts provided recommendations to be considered in addressing the weaknesses identified in the MF (Table 3).

Based on the consolidated information, modifications were made to the MF (Figure 9):

The integration of knowledge into data modeling allows for a reduction in data dependence, an improvement in the precision and robustness of models, and, in some cases, confers physical meaning to the obtained results (Willard et al., 2020). Also, knowledge management strategies are critical for making decisions in climate change mitigations and adaptations to ensure better practices in small farming (Chisita and Fombad, 2020). In this context, some authors propose general frameworks for knowledge integration in data modeling, such as in Von Rueden et al. (2019), where the information flow in a process called informed machine learning is defined. This process generally involves problem identification and the search for a joint solution where data and prior knowledge are integrated, presenting some mechanisms for representing knowledge and its integration into data modeling. Similarly, in Roscher et al. (2020), an approach is proposed where the integration of domain knowledge is considered to improve the explainability of data models. Additionally, in Karpatne et al. (2017), despite not presenting a guide for knowledge management or its integration into data modeling, the paradigm of theory-guided data science is referred to, where the use of explicit and tacit knowledge is considered for refining the results of data models to be consistent with the understanding of physical phenomena.

Similarly, the proposed Methodological Framework (MF) is based on the general approach of integrating knowledge into data modeling. However, it delves into the processes by presenting specific activities to support the characterization, extraction, and representation of knowledge and its subsequent integration into data modeling. It considers the type of knowledge required, the type of data structure, and methods of knowledge extraction and representation, allowing for the support of the optimization of data models in their different development phases.

Regarding the characterization and extraction of knowledge, according to its origin and considering the types usually defined in the knowledge management area, it is classified as explicit and tacit knowledge (Hajric, 2018). Explicit knowledge, formalized and encoded, is called “Know-What.” This type of knowledge is found in the content of indexed journals, databases, public documents, reports, videos, and images, among others. Explicit knowledge is contained in files with different formats of structure, known as structured, semi-structured, or unstructured data, and treated by various methods to carry out the extraction process. For the extraction process of explicit knowledge, the MF considers the identification of the type of data structure where it is contained. Data extraction and direct storage are proposed when dealing with structured formats, thinking they possess a formal structure (Hajric, 2018). The performance of a normalization and transformation process from semi-structured to structured data is presented for semi-structured data. It is suggested by some authors who carry out the knowledge extraction process from HTML formats refined for the identification of concepts with the help of experts (Ahsan et al., 2014; Bonacin et al., 2016), the use of web crawlers to extract information directly from pages in HTML or XML formats (Baumgartner et al., 2005; Lin et al., 2018; Zhai et al., 2021) or the process of manual error correction, normalization, and standardization of semi-structured data to structured data suggested by Chenglin et al. (2018). This procedure is necessary to obtain data in a formal structure to be worked with using knowledge representation methods.

Similarly, tacit knowledge, known as “Know-How,” corresponds to that found in the minds of individuals and has not been quantified or represented in any accessible format. It is manifested through practices and experiences in the application domain (Rhem, 2005; Becerra-Fernandez and Sabherwal, 2014) and possesses defining characteristics such as difficulty in communication, practicality, experiential nature, unconsciousness, and personalization (Pérez-Fuillerat et al., 2019). In the MF, when the process of extracting tacit knowledge is carried out, knowledge elicitation methods are used, which involve storing extracted knowledge from experts in non-structured formats (Jakus et al., 2013). The use of elicitation methods depends on the characteristics of the users with whom the process will be developed. In the case of agriculture, some studies suggest the application of techniques such as knowledge harvesting (Frappaolo, 2008), storytelling (Whyte and Classen, 2012; Prasarnphanich et al., 2016; Zammit et al., 2018), interviews (Ferrari et al., 2016), and video sharing (Zammit et al., 2018).

Subsequently, when knowledge is contained in a non-structured data format, either through elicitation or explicit knowledge in this structure, the MF proposes a process of extracting knowledge considered implicit knowledge. It refers to patterns or relationships between data that are not evident to humans (Frappaolo, 2008). For this purpose, a tacit knowledge extraction task is established and supported by a data object containing extraction methods identified in the agriculture domain. Among the recognized methods, some studies report the use of manual tasks to carry out the extraction and categorization of data (Devraj and Deep, 2015; Bonacin et al., 2016; Goldstein et al., 2019; Admass, 2022). Similarly, some authors mention Natural Language Processing (NLP) techniques, such as the “Stanford Dependency Trees” structure used for extracting entities from the agricultural knowledge domain (Devi and Dua, 2017), Named Entity Recognition (NER) used for identifying and classifying entities from text into predefined categories (Chenglin et al., 2018), Predicate-Argument Structure (PAS) used to represent relationships between the predicate and its arguments (nouns, prepositional phrases, etc.) in a sentence (Chatterjee et al., 2019), Conditional Random Field (CRF), which corresponds to a probabilistic graphical model used for sequence labeling, and Syntactic Tree-based Relation Extraction, which uses syntactic trees to extract relationships between named entities in text (Xiaoxue et al., 2019). There are also tasks proposed by Gharibi et al. (2020), such as POS tagging, chunking, and Stanford Parser, which allow the identification of relevant words, their grouping into meaningful phrases, and the provision of a syntactic structure for understanding relationships between words. Finally, other authors mention the use of neural networks such as Lattice Long Short-Term Memory (LSTM), Structured Perceptron, or Bidirectional Gated Recurrent Init (bi-GRU; Kung et al., 2021; Zhu et al., 2021), used to process sequences of data in texts. These methods are necessary to identify patterns that are not explicit to humans and are present in the unstructured data used in the knowledge characterization process.

On the other hand, following Bergman (2018), knowledge representation is the description of an object through different elements. Knowledge representation comprises three main aspects: concepts as basic units of knowledge, associations or relationships between concepts, and a dynamic structure built by the concepts and their associations (Gutiérrez, 2012). Knowledge representation methods are applied to logical language resources, that is, formal and explicit language. Therefore, the Methodological Framework (MF) establishes a series of activities to extract and transform knowledge from unstructured and semi-structured data into a set of structured data that possess the required characteristics to implement representation methods (Staab and Studer, 2009). In this context, the MF delineates activities for selecting the knowledge representation method and its subsequent implementation. It is supported by data objects containing representation methods, their limitations, and the tools available to carry out the process.

The methods identified are production rules, generally used for procedural knowledge representation (Yingying et al., 2017), i.e., methods or processes for performing a task (Gutiérrez, 2012). In agriculture, it has been widely used, especially for supporting pest and disease management (Balleda et al., 2014; Devraj and Deep, 2015; Kalita et al., 2017; Yingying et al., 2017; Admass, 2022). In some cases, production rules are used with other knowledge representation methods, as in Yingying et al. (2017), where rules are combined with knowledge graphs to design an expression and reasoning model for diagnosing diseases in tomato cultivation. In Afzal and Kasi (2019), a knowledge model based on ontology was developed to support rice production, using rules to keep the reasoning process of the knowledge base created through ontology. Additionally, in Sottocornola et al. (2023), rules are employed to support the explanation process of diagnosis in treating diseases in apple cultivation.

However, the use of production rules in agriculture has limitations, such as the relatively small size of the constructed knowledge bases (Balleda et al., 2014); moreover, when working with semi-structured or unstructured data, metadata may be derived from unreliable sources due to incomplete or incorrect information (Gomez-Perez et al., 2017). Rule-based systems are limited to the dataset used for constructing the knowledge base, which may not represent all the dynamics of the addressed problem (Godara and Toshniwal, 2020). Similarly, rule-based models are limited by the quality of the rules and require extensive expert intervention in the domain for rule maintenance and updating. They also face challenges when attempting to scale to other problems, either within the same domain or outside of it (Nismi Mol and Santosh Kumar, 2023).

On the other hand, ontologies can be defined as a formal and explicit specification of a set of related concepts (Jakus et al., 2013). Some studies in agriculture have proposed the use of ontologies to improve semantic interoperability between developed systems and data sources (Bonacin et al., 2016; Stucky et al., 2018), design and build a knowledge base to support query systems (Devi and Dua, 2017; Aminu et al., 2019; Jearanaiwongkul et al., 2019), support the development of knowledge graphs serving as a design layer (Chenglin et al., 2018; Xiaoxue et al., 2019), provide lexical modeling and conceptualization to extracted knowledge (Yanchinda, 2019) or propose a semantic representation of IoT device data to reduce the need for human intervention (Afzal and Kasi, 2019). However, ontologies have limitations in their application, such as linguistic disambiguation. Expert keyword selection and query formulation may affect the quality of results, requiring a high availability of experts for any system scaling process. Many resources are needed for knowledge base maintenance.

Additionally, there is a low standardization of concepts in the agriculture domain, affecting ontology understanding and consistency, along with language barriers in which concepts used for ontology construction are found (Ahsan et al., 2014; Bonacin et al., 2016; Goldstein et al., 2019; Fahad et al., 2021; Kung et al., 2021; Malik et al., 2021). These limitations have led to the development of graphs as a novel mechanism for knowledge representation. It involves the extraction of entities, attributes, and their relationships, integrating knowledge through entity alignment and association with ontologies. Moreover, it facilitates the completion of the knowledge update and retrieval processes (Xiaoxue et al., 2019).

Like ontologies, knowledge graphs serve as a structured semantic knowledge base that describes concepts and their relationships in symbols (Xiaoxue et al., 2019). In this sense, graphs can be represented with varying levels of formalization, depending on whether one desires a lighter and more flexible representation or aims for knowledge representation with semantic consistency, integrating with an ontology that serves as a design layer for the knowledge graph (Chenglin et al., 2018). Under this, some authors have proposed the use of knowledge graphs with semantic support through ontologies to assess the impacts of agriculture and climate change on water resources (Bonacin et al., 2016), represent knowledge at a general level in the agricultural field (Ahsan et al., 2014; Devi and Dua, 2017; Chatterjee et al., 2019; Malik et al., 2021), automatically generate agrometeorological reports (Chenglin et al., 2018), address fertilization and soil management in corn cultivation (Aminu et al., 2019), support decision-making in pest and disease management (Goldstein et al., 2019; Jearanaiwongkul et al., 2021), and precision agriculture (Fahad et al., 2021).

Similarly, studies have been proposed to consider using knowledge graphs to support the wine sector, employing a lighter and more flexible representation, i.e., without being supported by an ontology (Abbal et al., 2016; Groumpos and Groumpos, 2016). Finally, like ontologies and production rules, knowledge graphs present similar limitations, such as low scalability to other knowledge domains and even to different areas within the same knowledge domain. The quality of the represented knowledge depends on the input data to the system (Chenglin et al., 2018), the need for labeled data for the application of machine learning models for entity and relationship extraction, and the necessity of domain knowledge experts for identifying or verifying meaningful semantic relationships among extracted concepts, which can consume significant resources (Chatterjee et al., 2019); moreover, graphs must undergo constant maintenance and updates, requiring a substantial allocation of resources due to the need for a high level of expertise in the knowledge domain (Xiaoxue et al., 2019). When selecting a knowledge representation method, the limitations of its application must be considered to address the problem appropriately.

Integrating knowledge into data modeling is of great interest, particularly in scenarios where data might be inaccessible, unavailable, or of low quality (Porth et al., 2019). Knowledge integration can occur at any phase of modeling (Von Rueden et al., 2023). Therefore, within the MF (Methodological Framework), a flow is established through an inclusive gate that allows the inclusion of knowledge represented in any data modeling phase. The conditions set in the inclusive gate include data generation (data understanding), model evaluation (model assessment), parameter adjustment (model development), scientific consistency (business understanding and model evaluation), and attribute selection (data preparation). In this regard, studies have been proposed related to knowledge integration in the data acquisition phase (Hain et al., 2011; Wang et al., 2017; Read et al., 2019; Yu et al., 2019; Clemens and Viechtbauer-Gruber, 2020; Downton et al., 2020; Zhao et al., 2020; Sepe et al., 2021; Yu et al., 2021; Raymond et al., 2022; Schröder et al., 2022), data preparation phase (Froehlich, 2020; Mudunuru and Karra, 2021; Bajracharya and Jain, 2022; Fuhg and Bouklas, 2022; Kohtz et al., 2022), optimization process of machine learning algorithms (Anoop Krishnan et al., 2018; Azari et al., 2020; Chadalawada et al., 2020; Huang et al., 2020; Qian et al., 2020; Sun et al., 2020; Tartakovsky et al., 2020; Jurj et al., 2021; Lu et al., 2021; Soriano et al., 2021; Kim et al., 2022), and as support for explaining data model results (MacInnes et al., 2010; Read et al., 2019).

Ontologies and knowledge graphs can support interoperability among knowledge domain datasets, verify the quality of extracted data, classify data, extract attributes or relationships, or facilitate working with heterogeneous data (Robinson and Haendel, 2020; Sahoo et al., 2022; Mummigatti et al., 2023). Furthermore, axioms established in an ontology can support constructing new ontologies by inducing the reuse of existing knowledge or verifying the consistency of the new ontology (Smith et al., 2007; Mungall et al., 2011). They can also expand or enrich the characteristics used in a machine learning model without finding relationships from the data, ensuring consistency or coherence through context rules (Kulmanov et al., 2021; Shrivastava and Deepak, 2023). Similarly, ontologies and graphs can be used for task prediction (Mazandu et al., 2017; Chen et al., 2021), text clustering (Wei et al., 2015; Ruas and Grosky, 2018; Mehta et al., 2021), or to support attribute reduction or selection (Garla and Brandt, 2012). These integrations are typically achieved through entity similarity or embedded entity methods (Deepa and Vigneshwari, 2019; Sun et al., 2020; Mežnar et al., 2022). Therefore, ontologies and knowledge graphs are highly useful in supporting the development of data models, especially in contexts such as small-scale agriculture, where historical data series are mostly unavailable, or the available data is of low quality.

On the other hand, in some cases, knowledge can be explicitly represented, allowing its integration into data modeling phases without any characterization or extraction process. In this regard, hybrid models that integrate results from mechanistic or empirical models have been developed, either for generating training data or for model evaluation data (Ji and Lu, 2018; Feng et al., 2019; Maya Gopal and Bhargavi, 2019; Saha et al., 2020; Sansana et al., 2021). Additionally, the integration of algebraic or differential equations has been proposed, which can be used to condition policy in learning, modify the error function, function parameterization, or as restrictive functions (Mangasarian and Wild, 2008; Karpatne et al., 2017; Lu et al., 2017; Muralidhar et al., 2019; Ramamurthy et al., 2019; Asvatourian et al., 2020; Gupta and Das, 2020; Meng et al., 2022). Similarly, invariance properties have been proposed to enhance the performance of machine learning models (Ling et al., 2016; Wu et al., 2018). Lastly, expert knowledge has been incorporated to ensure that results generated by machine learning models have scientific consistency (Brown et al., 2012; Choo et al., 2013; Spinner et al., 2020). Thus, knowledge integration depends on the improvement objectives sought concerning data models.

The increase in variability and climate change, diseases, and pests, among other problems, negatively impacts agriculture, particularly affecting small-scale producers who are highly vulnerable and have low resilience. Additionally, food security relies on the adaptive capacity of small-scale producers to address such events (Hatfield et al., 2020). A significant amount of research has proposed data methods to contribute to solving these problems (Xie, 2011; Ghahari et al., 2019) (Dalhaus et al., 2018; Wang et al., 2018; Mangani and Kousalya, 2019; Roznik et al., 2019; Shirsath et al., 2019; Boyd et al., 2020; Zhang et al., 2020). However, the information used has different temporal and spatial resolution, affecting its correct application at the local level. At the local level, farmers possess knowledge about practices and techniques; however, this local knowledge can vary from one agricultural region to another. In this context, knowledge extraction and representation can be useful for storing knowledge from heterogeneous sources and sharing it with farmers (Jearanaiwongkul et al., 2019; Haider et al., 2021). Furthermore, there is an exponential amount of data about farm management and system conditions, necessitating proper methods to represent and share this data to support farmers’ activities (Aminu et al., 2019; Goldstein et al., 2019; Bhuyan et al., 2022). For this reason, in the context of small-scale farming, it is necessary to complement data analysis with knowledge that can support model development, considering data scarcity and heterogeneity.

Another problem where the Framework can be useful is addressing the lack of financial data to support risk management in the context of financial inclusion. In this sense, knowledge about system conditions or agronomic management may be necessary for develop new instruments for improvement. The Methodological Framework (MF) can facilitate the extraction and representation of knowledge from various sources to build new tools, such as credit scoring, while considering the heterogeneity of diverse agricultural systems (Simumba et al., 2018; Bunnell et al., 2021).

In the context of agricultural insurance, the management and integration of knowledge in data modeling will enable the proposition of agricultural insurance design solutions, facilitating the reduction of aggregation bias by considering specific characteristics of the production system. These include crop phenology, access conditions or availability of primary resources and implementing techniques or practices that enhance or diminish producers’ adaptive capacity. Additionally, it may facilitate the integration of area-related knowledge, such as agroecological classifications or soil types. This adjustment would fine-tune the utilization of the proposed parametric index, consequently mitigating idiosyncratic risks. It also aims to minimize gaps in insurance acquisition stemming from poor design comprehension or a weak correlation between premium payments and individual-level losses (Berg et al., 2009; Ramasubramanian, 2012; Thompson, 2017; Fonta et al., 2018; Madaki et al., 2023).

The optimization of data models through knowledge can provide producers with more adaptive tools to enhance their resilience against variability and climate change events, diseases, pest control, and all agronomic management factors contributing to food security and economic growth in small-scale agriculture.