This subsection reviews key perspectives from Information and Communication Technology for Development (ICT4D), a field examining how digital tools are introduced and sustained in developing countries, to explain why technology adoption in rural communities often faces significant contextual challenges. The persistent failure of technology-based interventions in rural areas is rarely a result of technical deficiency; rather, it stems from a design-reality gap that overlooks the human aspect of system interaction (Heeks 2018). Within ICT4D research, this gap occurs when implemented systems are designed based on urban or institutional assumptions that mismatch the localised decision-making realities of rural users. Successful implementation requires contextual integration, where technical success is defined by how effectively the technology couples with existing community behaviour. This gap is further widened by human factors, specifically how individual characteristics shape risk perception and technology trust (Walsham 2017). In rural coastal settings, the adoption of an EWS might be influenced by the recipient’s behavioural profile, including age, fishing experience and their specific role within the fleet. The theoretical foundation of the design-reality gap identifies seven dimensions of distance between system design and local reality: Information, technology, processes, objectives, staffing, management and other resources (Heeks 2018). In rural disaster management, this gap is most prominent in the information dimensions. A centralised design often assumes standardised information quality.

While governments and agencies increasingly deploy ICT-enabled warning platforms, rural users often underutilise them because of limited contextual fit (Baudoin et al. 2016). Disaster warnings that do not match local environmental understanding are perceived as unclear or irrelevant, resulting in delayed or non-response (Mercer et al. 2012). These findings indicate that the success of ICT interventions in rural contexts is not determined by technology availability alone but requires alignment with local knowledge, cultural practices and information interpretation mechanisms. Therefore, this study argues for context-specific system design that integrates human aspects, including cultural dimensions, alongside technical features (Heeks 2018; Walsham 2017). This argument provides the basis for adopting a human aspect in disaster ICT development, where technology must be designed in relation to local knowledge systems and community practices.

The smart village concept offers a foundation for community resilience by empowering communities to utilise local strengths alongside appropriate technology. Rather than replacing traditional practices, this paradigm views technology as an ecosystem that supports bottom-up innovation and community-based problem-solving (Alhari et al. 2022; Mukti et al. 2022; Susanti et al. 2023). This approach marks a critical shift from top-down technological imposition to community-led digital transformation, where rural innovation emerges through the meaningful adaptation of technology to local needs (Dembovska et al. 2023; Li & Zhong 2022; Lombardo, Saeli & Campisi 2023). The concept of Smart Village can be seen in Figure 1.

In top-down approach, the design is often rigid and centralised, leading to interpretability failures in existing national systems. Conversely, the smart village approach positions local knowledge as a strategic development asset, acknowledging that rural communities generate context-specific insights that guide decision-making under environmental uncertainty. This ensures that the system provides the specific recommendations and action-anticipation cues that users currently find lacking in standard platforms.

In this context, LIK is not merely cultural folklore but a form of local innovation that strengthens digital inclusion by ensuring technologies reflect local realities and remain socially acceptable. By prioritising community participation, the smart village approach ensures that disaster governance is co-produced with the people it is intended to protect. This alignment ensures that LIK is not treated as an add-on to disaster technology but is incorporated as a core knowledge component that guides warning interpretation, decision-making and community response.

Local indigenous knowledge refers to context-specific environmental knowledge, practices and interpretations that evolve through long-term interaction between communities and their ecological surroundings according to United Nations Educational, Scientific and Cultural Organization (UNESCO). Transmitted across generations, LIK forms a community-based knowledge system that supports daily survival and disaster recognition. In rural communities with limited access to formal infrastructure, LIK serves as an autonomous early warning mechanism derived from direct environmental sensing (Mercer et al. 2012). Because it is accessible, observable and culturally trusted, LIK plays a significant role in strengthening community resilience. Current research advocates for hybrid strategies where LIK and scientific systems are combined to enhance decision-making (Akanbi & Masinde 2018; Hiwasaki et al. 2014; Limpo et al. 2022; Wang et al. 2019). Despite this potential, LIK remains undervalued in formal disaster management and is often documented as narrative observations rather than actionable knowledge (Hiwasaki et al. 2014). Existing research on LIK–EWS integration remains largely conceptual and fragmented, and most studies focus on documentation rather than demonstrating how knowledge can be systematically embedded into technological architectures. Consequently, LIK often remains external to system design. Two methodological gaps hinder effective integration: Firstly, the validation gap, many efforts rely on expert-driven interpretation of a small number of informants, risking subjective bias. Without broader community validation, integrated LIK may lack community acceptance. This study addresses this by identifying six specific trust parameters of LIK from previous literature and one added parameter found through interview, which are utilisation, continuity, combination, duration, frequency, accuracy and response to quantify community consensus. By measuring these factors, LIK is transitioned from a subjective narrative to objective data points suitable for a trust index. This quantitative grounding ensures that only rules with high community reliability are prioritised for system integration. Secondly, the transformation gap, prior frameworks lack methodological clarity on converting tacit, narrative knowledge into structured, machine-readable formats. Literature rarely offers operational steps for converting LIK into rule-based logic capable of supporting automated early warning decision models. To address these gaps, this study proposes a socio-technical integration framework that treats LIK as a structured knowledge resource. By emphasising a validation phase, the framework ensures LIK is credible, community endorsed and suitable for computational translation. This process serves as a bridge between local wisdom and digital architecture by providing a standardised evidence base for automated decision support.

A practical implementation gap remains in transforming narrative-based LIK into structured decision rules suitable for digital platforms. Recent research suggests that for LIK to move beyond narrative forms, it must be represented in a format compatible with automated decision models. While this study establishes a trust index through community validation, empirical data indicate that the hierarchical importance of LIK signs is not static, and it fluctuates across different geographical contexts and specific parameters. This dynamic variance presents a complex challenge for traditional, static systems.

Machine learning (ML) is identified as the necessary next-stage pathway for this research to automate the reconciliation between the seven trust parameters and specific behavioural profiles of the community across different geographical contexts. Following methodologies in recent multimodal data behaviour research, Random Forests or Gradients Boosting (XGBoost) can be utilised to identify nonlinear relationships between these parameters. Specifically, by applying SHAP (Shapley Additive exPlanations) values, the model can quantify the feature importance each parameter contributes to safety or unsafe (inaction) responses (Li et al. 2025). This enables the identification of specific parameter combinations that lead to behavioural unsafe (inaction) even when high-accuracy signs are present. By positioning these validated rules as training features for future ML models, the framework provides a computational foundation for hybrid, personalised warning systems. This approach allows for the determination of optimal alert thresholds, calculating whether the interaction between LIK trust parameters and behavioural traits necessitates an extra warning to overcome the risk of unsafe (inaction) behaviour. Consequently, the framework moves beyond simple documentation to become a predictive system for optimising disaster response within the specific coastal fishing communities.

The primary motivation for this study is the design-reality gap in rural disaster management, where top-down EWS fail to achieve long-term adoption because of a lack of contextual alignment. Sentiment analysis of national EWS platforms revealed pervasive concerns regarding information quality, specifically citing data delays, forecast inaccuracies and a lack of actionable recommendations.

The initial inquiry confirmed a significant geographic variance in technology utility. In Aceh, formal ICT-based adoption was measured at 70.8%, yet the perceived ICT benefit was notably lower at 60.4%, indicating that nearly 10% of users found the technology ineffective for their needs. In Yogyakarta, while ICT adoption and perceived benefit both reached 100.0%, deeper analysis through the System Usability Scale (SUS) and User Experience Questionnaire (UEQ) revealed that this high usage masks significant frustration. Results identified marginal usability (SUS < 68) and negative dependability. In contrast, LIK remains universally trusted (100% adoption) across all regions. The identified problem is the lack of an interpretive bridge that can successfully translate technical signals into the social logic required for community action.

The objective of this research is to develop a socio-technical integration framework that enables the systematic transformation of tacit LIK into a formalised, machine-readable format. The solution must meet two primary criteria: (1) it must provide a structured metadata schema capable of transforming narrative LIK into computable logic, and (2) it must establish a community-based validation to ensure integrated knowledge is endorsed by the communities. To achieve this, the solution identifies seven specific trust parameters: Utilisation, continuity, combination, duration, frequency, accuracy and response. By meeting these objectives, it provides a computational foundation for future hybrid EWS that integrate technology with indigenous environmental sensing.

This stage focuses on the creation of the LIK metadata schema, a design artefact developed through thematic analysis of qualitative narratives. The study was conducted in three rural coastal regions of Indonesia characterised by high exposure to hydrometeorological hazards and strong LIK traditions: Pangandaran (West Java), Lhoknga (Aceh) and Depok (D.I. Yogyakarta). These locations were selected based on their direct exposure to the Indian Ocean and historical experiences with major disasters, such as the 2004 and 2006 tsunamis. Table 2 shows the details of study location and the different characteristics it possessed.

Data acquisition involved 17 in-depth interviews with local fishermen to elicit detailed narratives of environmental sensing and hazard interpretation. This was followed by a focus group discussion (FGD) with eight heads of fishermen communities to validate and refine the emerging insights, ensuring the data reflected shared community consensus rather than isolated observations. The resulting schema structures LIK into five distinct components: Attribute, Object, Value, Effect and Action anticipation. The core innovation of this artefact is the action anticipation, which captures the specific community-endorsed protective measures triggered by an environmental sign.

The artefact’s (LIK metadata schema) were evaluated through a quantitative process. Firstly, the acquisition and formalisation method was demonstrated by converting raw qualitative interview transcripts into 33 distinct LIK. Secondly, the artefact was subjected to an empirical evaluation through a survey of 438 fishermen to quantify community consensus. To ensure evaluation rigour, we implemented an operationalisation protocol where participants anchored responses in core memories of high-stake disaster events to mitigate recall bias. The reliability of the validation instrument was confirmed via a test–retest protocol with 20 fishermen, yielding an intraclass correlation coefficient (ICC) > 0.8, which indicates high temporal stability in the community’s assessment of sign reliability and behavioural triggers.

Ethical clearance to conduct this study was obtained from the Gadjah Mada University Medical and Health Research Ethics Committee under request number (Ref. No. 882609/UN1/FTK.2/DTETI/KM/2025).

This study initially addressed the problem identification phase of the DSRM by quantifying the gap between existing technical solutions and community needs. The survey results provided confirmation of the design-reality gap. A sentiment analysis of general Indonesian users of national EWS platforms like InfoBMKG revealed pervasive concerns regarding information quality, specifically citing data update delays, forecast inaccuracies and lack of interpretability where technical alerts failed to provide clear, actionable meaning for the user. While these sentiments reflect a broad national user base, they signify a critical information gap for rural communities that require local, real-time data for appropriate decision-making. While LIK adoption was found to be universal, with 100% of fishermen confirming its use and perceived benefit, the adoption of formal ICT-based disaster systems was significantly lower.

Furthermore, empirical data from 438 fishermen from coastal fishing communities reveal that this gap is compounded by a significant geographic divide in technological adoption. While ICT usage reached 100.0% in D.I Yogyakarta, it dropped to 70.8% in Aceh, illustrating that the technology dimension of the gap is highly context dependent (Figure 2). Critically, even where adoption is present, the perceived benefit of ICT is notably lower at 60.4% in Aceh.

While the survey indicated a 100.0% ICT adoption and benefit rate in D.I Yogyakarta, a deeper investigation was conducted to determine whether this high usage masked underlying frustrations. Further surveys utilising the SUS and UEQ in the coastal fishing communities of D.I Yogyakarta region revealed more data regarding the actual quality of interaction. This analysis confirms a significant disconnect, the platform received a SUS score below 68, indicating marginal usability, while UEQ scores for dependability and efficiency were categorised as neutral to negative. These metrics identify a reality where the system exists but fails to achieve its objectives because it provides warning information without the recommendations or action-anticipation cues required for rural decision-making. When these gaps, whether in access, interpretability or actionable information, are not addressed, the result is a partial failure where technology is present but ignored in favour of trusted LIK. These data prove that an integration approach is required to bridge high-trust indigenous logic with under-performing technical infrastructure.

To enable the future integration of LIK into technology-based EWS, the knowledge must be transformed from narrative descriptions into a structured, machine-readable format. Field analysis reveals that the LIK used by fishermen is not unstructured; rather, it follows a consistent internal logic where each observation contains core components explaining what is detected, how it is interpreted and what it implies for safety.

To address RQ1, this subsection demonstrates how the developed metadata schema formalises tacit indigenous narratives into structured logic. The schema was developed through a two-stage process involving a literature review followed by a qualitative analysis of 17 in-depth interviews with local fishermen and a FGD with eight heads of fishing communities.

This iterative process identified five essential components required to bridge oral tradition with computable logic. The LIK metadata schema is displayed in Figure 3.

While the initial literature review provided a basic structure for environmental observation, it was the qualitative field inquiry that revealed the necessity of a second, more advanced structure. Specifically, the interviews and FGDs highlighted that for LIK to be actionable, it must include action anticipation, the community-endorsed safety response triggered by a sign.

These components allow LIK to be represented in a form that follows logical reasoning, making it compatible with knowledge bases, decision rules and ML. The structure can be generalised as:

IF [Object: X] has [Value: A]

AND [Object: Y] has [Value: B]

THEN [Effect: Disaster]

ACTION RECOMMENDATION [Action Anticipation]

For example:

IF [Object: Tide movement] has [Value: Calm]

AND [Object: Water level] has [Value: Sudden recession]

THEN [Effect: Tsunami]

ACTION RECOMMENDATION: [Action Anticipation: stay away from coastline]

In many cases, fishermen emphasised that a single environmental indicator is insufficient to trigger a safe response (taking action anticipation). Instead, they cross-validate multiple LIK signals before making safety decisions. For example, while domestic animals appearing restless or agitated are a known indicator, it is rarely enough on its own to prompt an evacuation to higher ground. It must be accompanied by secondary signs, such as the appearance of sudden, anomalous waves during clear weather. Similarly, a notably calm sea state since morning or experiencing no typical tidal movement is not perceived as an immediate threat until it is corroborated by a sudden recession of the sea level or a felt earthquake. This indicates that LIK frequently operates through multi-condition logic. During analysis, several such combinations were identified across the three study locations, many of which are utilised consistently by all fishing communities. This cross-regional consistency suggests a strong collective validation of compound LIK rules. From a design perspective, this structured representation enables the systematic conversion of narrative local knowledge into rule-based logic. These validated, multi-factor rules provide the necessary foundation to be encoded into ML models for EWS, acting as the essential computational bridge between raw indigenous observations and automated disaster logic.

To systematically identify and structure LIK into a machine-readable format, the framework utilises a process called LIK acquisition. This stage focuses on identifying indigenous knowledge through expert consultation, including 17 in-depth interviews with local fishermen and a FGD with eight heads of fishing communities. The acquired data are then organised into a format that allows it to be processed and utilised by digital systems. This acquisition phase is divided into two distinct sub-stages: LIK identification and LIK formalisation. Local indigenous knowledge identification involves documenting and classifying various forms of local knowledge, such as animal behaviours and environmental cues, through the involvement of domain experts. As shown in Table 4, tacit narratives like ‘A sudden recession of the sea level’ or an ‘Pets and Domestic animals appear restless (or agitated)’ are extracted from interview transcripts and assigned specific codes (e.g. Ts-3 to Ts-6) based on their geographical relevance in West Java, Aceh or Yogyakarta. This LIK identification example can be seen in Table 4.

Local indigenous knowledge formalisation subsequently structures this identified knowledge into a structured format following LIK metadata schema. Table 3 presents the five-component logic (Attribute, Object, Value, Effect, and Action anticipation), which is demonstrated through the narrative transformations shown in Table 5. For instance, a narrative about a sudden recession (Ts-3) is formalised into specific variables where the ‘Attribute’ is Sea, ‘Object’ is Water Level, ‘Value’ is Sudden Recession, ‘Effect’ is Tsunami and the ‘Action anticipation’ is to stay away from the coastline. This structured representation provides the necessary bridge between narrative LIK and computable disaster logic.

Following the evaluation phase of the DSRM, the formalised rules were subjected to a community-consensus test to answer RQ2. To transition from expert-driven identification to a data-driven representation of the broader population, the LIK validation stage was conducted through a survey of 438 fishermen across the study locations. This stage verifies the identified indigenous knowledge through community consensus, moving beyond subjective interpretation, while the other stages such as rule-based knowledge and LIK integration are reserved for further research. The entire stages can be seen in Figure 4: LIK acquisition and LIK validation stage.

By applying a trust index to measure reliability, we calculated mean scores across seven trust parameters found from the literature review and interview: Utilisation, Continuity, Duration, Frequency, Accuracy, Response and Combination. Table 6 details the definition of each parameter and the questions that were asked during the survey.

The results demonstrated that these seven parameters significantly re-order the hierarchy of LIK signs beyond simple accuracy. In terms of duration (Figure 5), Ts-6 (Pet Distressed) was ranked higher than Ts-4 (Sudden, anomalous waves), suggesting it is perceived as a more persistent signal. However, the order reversed when evaluating frequency (Figure 6) and response (Figure 7), where Ts-4 was ranked higher; this indicates that while Ts-6 is more persistent, Ts-4 is more effective at actually triggering protective actions. These shifts suggest that even though both indicators were universally validated with 100% accuracy (Figure 8), their perceived reliability varies significantly across different trust dimensions.

This quantitative validation proves that reliability is not a single-variable metric; instead, it converts raw observations into rule-based knowledge. By identifying which signs are most reliable across specific parameters, the framework provides the necessary evidence to determine which LIK rules are trusted enough for automated system integration and which require additional cross-validation.

Another significant finding of this evaluation is the impact of geographic settings and behavioural variance on LIK trust. While tsunami-related rules (Ts-1 to Ts-6) maintained universal accuracy across all sites, weather-related LIK signs such as Wn-1 (descending clouds) showed regional variance, with 91.2% accuracy rating in Aceh compared to 80.4% in Yogyakarta. This variance confirms that the framework serves as a diagnostic tool to pinpoint critical, context-specific rules for different coastal regions rather than relying on a generalised model.

Furthermore, the data suggest that actionable utility (defined by the parameter of response) is significantly influenced by the fisherman’s behavioural profile. Key factors include age, fishing experience, fishing role, fishing location, number of experiences with different types of disasters and number of interactions with disasters. These factors dictate response readiness. For example, the data identified a specific overconfident profile, typically characterised by older fishermen with extensive sea experience, who may ignore high-accuracy LIK signs.

The regional findings highlight that the study locations possess distinct behavioural profiles, which further influence the varying levels of trust in LIK signs. As illustrated in Figure 9, the comparative data for Aceh and Yogyakarta, factors such as age distribution, fishing experience and LIK usage duration show clear differences between the two fishing communities.

Specifically, the Aceh profile is characterised by a higher mean age and more extensive fishing experience compared to D.I. Yogyakarta. These differences in experience and age contribute directly to how each community prioritises LIK parameters and responds to environmental triggers. Furthermore, the fishing location significantly dictates the specific types of LIK utilised by each community. For instance, fishing communities in Aceh often operate in locations far from the shore, requiring them to stay at sea for several weeks. Because they must observe signs while in the middle of the sea, often at night, they rely more heavily on celestial indicators rather than animal behaviours or coastal environmental cues that are difficult to monitor in deep-water or nighttime settings. Figure 10 shows that the celestial indicator only appears in Aceh.

Because of these distinct behavioural landscapes, the selected study locations serve as a significant choice for testing the framework. They allow for a comparative analysis of how socio-demographic backgrounds shape the reliability and actionable utility of LIK. These findings lay the groundwork for future research involving ML. By integrating the weighted trust index with these specific behavioural profiles, a multi-dimensional decision matrix can be developed. This matrix will enable future EWS to provide customised notifications, determining whether to issue a standard warning or an extra warning based on the interaction between the sign’s reliability and the unique behavioural characteristics of the recipient.