The CSA for this research is Vale do Sousa, located in northwestern Portugal, approximately 50 km east of Porto. This region extends over 28,941 hectares and includes joint collaborative management areas (ZIFs): Entre-Douro-e-Sousa (north of the Douro River) and Paiva (south of the Douro River). The primary land use in Vale do Sousa is forestry (Figure 2). The predominant species are eucalypt and maritime pine (Eucalyptus globulus Labill and Pinus pinaster Ait) while other species include pedunculate oak, chestnut and cork oak (Quercus robur L, Castanea sativa Mill, and Quercus suber L). These species, along with shrublands, cover an area of 20,318 hectares in Vale do Sousa, which is the focus of this study.

Vale do Sousa has faced significant wildfire challenges. Notable extreme wildfire events include those in 2016, which burned 2,923 hectares (10.1% of the total area), and in 2017, which burned 7,428 hectares (25.7% of the total area). These recurrent wildfires underscore the critical need for effective forest management and wildfire risk reduction strategies in the region.

The area is characterized by a fragmented forest ownership structure, with predominantly small, privately-owned forest holdings. Previous studies (Borges et al., 2017; Marques et al., 2021) have highlighted a strong stakeholder interest in wildfire risk reduction, making Vale do Sousa an ideal location for this research. The local stakeholders include a diverse group of actors with varied interests and goals, ranging from timber production to wildfire risk reduction. This fragmented ownership and diversity of interests complicates forest management and necessitates collaborative approaches.

For this study, we focused on 2,429 afforested management units (MUs) within a total of 3,205 units in the landscape. These MUs serve as the spatial basis for decision-making, and they are designed based on homogeneity of species, age, rotation, site index and topographic conditions. These MUs represent distinct forested areas within the CSA, allowing for the prioritization of fuel management efforts based on multiple ecological and socio-economic criteria. The identification and assessment of these units are fundamental to the MCDA approach applied in this research.

To address these challenges, the forest association ‘Associação Florestal do Vale do Sousa’ (AFVS) was established 29 years ago and leads the development of forest management plans within a complex decision-making context. This research took advantage of the existence of a Community of Wildfire Innovation (CWI) that includes a diverse group of stakeholders with representatives from the two AFVS local firefighters’ brigades, eight municipalities, NGOs, forest industry, government and non-governmental organizations (D9.4 Technical Periodic Report 1, FIRE-RES Project).

As part of the quest to elaborate a consensual order of priorities, in which the opinion of all participants is taken into account, but seeking a group opinion, the preferences of each participant and their performance in the pairwise comparison are analysed.

To assess the quality and alignment of participants’ inputs, we used a combination of Euclidean distance (ED), Spearman rank correlation (S), and CR, similarly to Srđević et al. (2008). Each metric captures a different dimension: ED measures divergence from the priorities of one individual to each of the others, S captures ordinal agreement in priorities, and the CR evaluates the internal coherence of judgments. This combined approach provides a more nuanced evaluation than any single metric alone. Additionally, we calculated the consensus convergence algorithm (Regan et al., 2006) implemented with Python for the dataset (Supplementary material 5), and the results closely matched those obtained from our selected measures, supporting the robustness and reliability of this approach.

The spatial data processing work took advantage of Geographical Information Systems (GIS), and Decision Support Systems (DSS) functionalities. Data acquisition involved a forest field inventory and publicly available spatial information for the Vale do Sousa classification into MUs. Biometric variables were processed by DSS growth and yield functionalities (e.g., Nunes et al., 2022). Additionally, the ‘FlamMap’ software (Finney, 2006) was used to estimate two fire behaviour variables using available biometric information, predominant wind directions, and the Portuguese fuel model (Fernandes et al., 2009). The decision support software CDP was used to build a hierarchical model to characterize the problem and the structure of its solution. The software facilitated the structuring of the model by organizing its overall objective (or goal), criteria, and sub-criteria levels, the alternatives (the MUs), and participants’ performance weights. This approach ensured that each participant’s preference was reflected in the MCDA model, rather than simply averaging the group’s preferences. The EMDS system, integrated within ArcGIS Pro (Reynolds et al., 2023), applied the MCDA model developed in the CDP software to each afforested MU. This process generated priority maps for each criterion, as well as a final composite map that integrates all criteria.

The symbology chosen for these maps utilized five equal intervals, spanning the minimum (‘Very low’) to maximum (‘Very high’) priority values assigned to each MU. Each class represents a constant data range, ensuring a straightforward and consistent interpretation of priorities (Milic et al., 2019).

Moreover, representative maps were developed for each of the 17 sub-criteria included in the study. The symbology of these maps was generated by using the Jenks Natural Breaks algorithm, which creates class breaks that group similar values together while maximizing differences between classes. This method enhances visualization by avoiding unrealistic class distributions caused by extreme values in the dataset (Chen et al., 2013).

To evaluate the reliability and robustness of the group-based prioritization, two complementary analyses were conducted: a consistency check comparing individual and group outputs, and a sensitivity analysis exploring the impact of weight changes on the prioritization results.

For the consistency analysis, the spatial priority maps generated using the weight vectors provided by each of the eight individual stakeholders were compared to the map obtained from the aggregated group weights. Priority values were grouped into five equal-interval classes, consistent with the visualization approach used in the results. For each MU, the number of individual maps that assigned the same class as the group scenario was recorded. The average number of matches per MU across all participants was calculated to quantify the overall level of agreement between individual and group prioritizations.

Sensitivity was assessed through a local perturbation approach in which each weight was adjusted independently by a fixed percentage (±10%), with the remaining weights proportionally rescaled to preserve the overall structure. This approach aligns with the crossover value suggested by Saaty (1994) to assess criticality, and applied in similar studies (Marto et al., 2018; Abelson et al., 2021). For each perturbed scenario, the Mean Absolute Deviation (MAD) of the priority scores across all MUs, relative to the base (group) scenario, was computed to provide a summary measure of average change per unit. In addition, the number of classification changes per MU across all scenarios was recorded.

A total of 22 people were contacted, with 19 ultimately participating in the first stage of the study. The participants represented a broad range of entities, including four town councils, the National Republican Guard, two public agencies responsible for environmental protection, a company specializing in forest fire protection and rural firefighting, two pulp and paper companies, an electric utility company overseeing powerline management, two local forest associations, three local firefighting organizations, and the National Authority for Emergency and Civil Protection. The responses from the 19 participants were evaluated, and the final criteria and sub-criteria were selected as outlined in the methodology.

The criteria are composed of four groups, originating from the initially proposed ones, all of which were accepted by the respondents in the first survey. A total of 17 measurable sub-criteria were selected (see Supplementary material 1 for individual sub-criteria maps), reduced from the 22 originally proposed. The five sub-criteria that were not retained were excluded based on limited support from stakeholders during the first survey. At the same time, other sub-criteria were added based on their suggestions, such as the wildfire resistance indicator and shrub biomass. These new sub-criteria were analysed, discussed during the FG meeting, and included in the study once it was determined that spatial data could be obtained to characterize them.

The criterion ‘Fuels’ assesses the status of the vegetation in the area by analysing the elected sub-criteria. Understanding the current vegetation condition, specifically those variables that can be modified by human actions (e.g., canopy cover, crown base height) allows for effective management. To obtain the values for the five sub-criteria selected, a set of previous studies, software and models were used (Nunes et al., 2022; Faias et al., 2012; Gómez-García et al., 2016; Mönkkönen et al., 2014).

The criterion ‘Potential for extreme fires’ incorporates maps from the ICNF (ICNF, n.d.-b): ‘Territories with potential for large fires’ (areas of more than 500 hectares that have not burned for more than 10 years in the ‘High’ or ‘Very High’ danger classes) and the ‘Short-term danger map of rural fires’. It also includes two other sub-criteria: a simulated spread rate map using FLAMMAP software (Finney, 2006) and a sub-criterion that addresses the resistance of the forest to the fire when it occurs, that is Wildfire Resistance Indicator (Ferreira et al., 2015).

‘Suppression drivers’ refers to characteristics of a MU that would maximize the combination of accessibility, availability of suppression infrastructures, optimal locations or adequate topographic conditions. The sub-criterion accessibility from the forest road network (ICNF, n.d.-a)., refers to the distance firefighters would need to cover on foot to reach the area from their vehicles. The slope (Agência para a Modernização Administrativa, n.d.), distance to water points (ICNF, n.d.-a) and proximity to the designed firebreak network (ICNF, n.d.-a) are also included, as they are considered to have an influence in the effectiveness of suppression efforts.

The Vulnerability criterion attempts to identify areas that are particularly threatened by fire and that hold significant cultural, social, or environmental value. These are often referred to as ‘high values at risk’. Specifically, the sub-criteria include urbanized areas (Direção-Geral do Território, 2020), which focus on protecting humans and their infrastructure. Another sub-criterion targets conservation corridors (ICNF, n.d.-c), which addresses key environmental assets, and the aspect (Agência para a Modernização Administrativa, n.d.), which represents specific in-risk locations based on their orientation. The inclusion of aspect was controversial during the participatory process, and its relevance was discussed during the FG. Nevertheless, other related studies have incorporated this type of sub-criterion (Abedi Gheshlaghi et al., 2020; Arca et al., 2020; Gigović et al., 2018; Sivrikaya and Küçük, 2022; Van Hoang et al., 2020). The fourth is a map with population density by municipality (Eurostat, 2021).

During the FG discussion, nine participants representing a range of organizations and expertise—including ICNF-GFR (forest and fire policy), Navigator (industrial forest management), CIM Tâmega e Sousa & CM Penafiel (local wildfire prevention and land management), AFEDV & AFVS (private forest owners and technical assistance), BV Penafiel (wildfire suppression), AGIF (national fire management strategy), and NPC-GNR Penafiel (law enforcement and environmental monitoring)—came together to discuss the relevant criteria and sub-criteria for the study. In our model, these sub-criteria are also the attributes that read spatial and tabular data, and the “parameters” we refer to are in fact the bounds used to define CDP utility-function curves, which translate raw attribute values into a 0–1 priority score.

Although the online survey had engaged 19 participants, having nine for the FG was considered appropriate, since every stakeholder sector was present for this phase. The FG took place in a town within the CSA. Most of the parameters for the sub-criteria were assigned there, and the remaining ones were determined post-FG by consulting local datasets (Table 1).

It is necessary to explain the definition of certain sub-criteria parameters. For example, in the case of the ‘Aspect’ sub-criterion, the parameters are expressed by classes based on slope orientation, with Class one representing North, Class two representing South, Class three representing West, and Class four representing East. Additionally, some sub-criteria have parameters expressed as percentages. For instance, ‘In touch with fire breaks’ refers to the percentage of the area of a Management Unit that overlaps with firebreaks. Similarly, for the ‘Resprouters’ sub-criterion, the parameter represents the percentage by area of resprouting species relative to seeders. This distinction is significant, as studies, including Botequim et al. (2015), have demonstrated that biomass accumulation following disturbance is slower in areas where the percentage of resprouting species is higher compared to seeders.

After the FG and the definition of utility function parameters for the sub-criteria, the remaining step was to assign to each a weight to each criterion and sub criterion (the relative importance of each to setting the priority for management). That information was obtained from a survey utilizing the AHP’s pairwise preference elicitation approach (Supplementary material 4), collected during the last step of the FG. The responses from eight out of the nine FG initial participants were collected. First, the responses were used to calculate each participant’s weight related to their performance and alignment with the group as measured by their CR, ED and S (Section 3.4). Then, the final weights were obtained for the overall group (Tables 2, 3).

The CR values for each participant were calculated to illustrate their consistency in making pairwise comparisons (Figure 3). Considering the findings of Saaty (1980), some participants exhibited CR values slightly above the 0.1 CR threshold, highlighting potential challenges in accurately comparing the criteria or sub-criteria.

To account for these inconsistencies, performance weights were assigned to participants based on the normalized inverse values of their CR scores (Equation 1). Participants with lower CR values, (indicating higher consistency) had a greater influence on the final prioritization. Conversely, participants with higher CR values were appropriately down weighted to ensure the robustness of the aggregated results (Table 4).

The S and ED measures highlighted the strong alignment of participant one’s preferences with those of the others (Table 5). Participants seven and eight also scored positively on these measures at different stages of the exercise. Noteworthy are participants three, four, five and six, whose S and ED scores highlight discrepancies in opinion on the weights of importance with the rest of the group. This means that the opinion of these participants, when combined with their CR, is taken into account but has less relevance in the final priority scores.

The hierarchical model (Figure 4), developed in the CDP software, integrates participants’ preferences while adjusting for the CR, S and ED of their assessments. No participant’s input is excluded, but their individual CR and level of alignment with the group’s preferences values are used to adjust the weight of their contributions. Therefore, the weights for their participation (Table 5) are included in the CDP model that contains the MCDA. Figure 5 illustrates a simplified version of the model with the criteria and sub-criteria.

The model begins by incorporating each participant’s evaluation of the relative importance of the main criteria groups. For example, the weight of the criterion ‘Fuels’ is calculated as a weighted summation of each participant’s preferences, adjusted by their individual weights of performance.

The same approach applies to the sub-criteria. Participants’ evaluations of relative importance of the sub-criteria within each criterion group are weighted based on their CR, S and ED weights, reflecting the level titled “Participant performance for sub-criteria weighting.” The final weights for each sub-criterion are calculated by aggregating the weighted inputs from all participants, ensuring that both the criteria and sub-criteria reflect a balanced consensus, while accounting for individual inconsistencies or large differences with the group’s preferences. This phase-specific approach ensured that different preferences in one phase did not disproportionately impact their overall contribution to the prioritization process.

The final output of the MCDA process is a set of spatially explicit priority maps, generated to aid decision-making in the CSA. These maps were produced using ArcGIS Pro and the EMDS software, applying the CDP model described in the previous section individually to each MU. For each MU, its lowest criteria values are converted to normalized priorities (scores) using AHP’s direct method based on the scale parameters (Table 1).

Four maps for the criteria ‘Fuels’, ‘Suppression drivers’, ‘Potential for extreme fires’, and ‘Vulnerability’ (Figure 6), were obtained, pointing out the five classes of priority for management based on each sub-criteria defined in the hierarchical model. A final map of priorities for management aligning with the objectives of this study is produced from the weighted combination of the other four criteria maps (Figure 7). From this map we can extract the following information regarding the area distribution for each priority class (Table 6).

Using equal intervals to visualize the maps of priorities we see that a smaller number of MUs fall in the higher classes regarding priority for management. For a MU to score in the ‘Very High’ class, it has to have high priority scores across all four criteria groups, and it can be observed how rarely that occurs (Table 6). It can be seen how most of the areas with very high management priority are located in the northwestern part of the CSA, creating a sort of priority band stretching from the Douro River to the far north. This aligns with the priority of three of the four priority maps by criteria, where we can see that both ‘Fuels’, ‘Suppression Drivers’ and ‘Potential for Extreme Fires’ have the highest priority areas in those zones. In Supplementary material 1, it is possible to see the 17 maps of each sub-criterion that characterize these four criteria maps and that visually illustrate the basis of the result obtained.

While the resulting priority map identifies areas of strategic interest for fuel management, it is important to note that this framework is prescriptive rather than predictive. Its purpose is not to forecast wildfire occurrence, but to support decision-making by structuring and spatially representing stakeholder preferences, expert knowledge, and contextual criteria. Therefore, traditional statistical validation techniques such as ROC curves or F1-scores are not applicable. Instead, a practical form of validation was conducted through visual inspection of high-priority areas using satellite imagery, confirming consistency with expected criteria combinations (e.g., fuel accumulation near populated zones and road networks). This validation was documented in a video shared with stakeholders for confirmation.

The analysis of agreement between individual stakeholders and group prioritization revealed an average match of 4.79 out of 8, indicating that, for each MU, nearly 60% of the individual classifications aligned with the group-based result.

To assess sensitivity, 42 weight perturbation scenarios were tested: one ±10% perturbation for each of the 17 sub-criteria and 4 main criteria. For each scenario, the MAD was computed across all MU relative to the group scenario. The MAD values ranged from near-zero up to a maximum of 0.00425, corresponding to less than 0.5% of the full priority scale. The five scenarios that resulted in the highest deviations corresponded to adjustments in the following weights: Potential for Extreme Fires (±10%), Fuels (−10%), Rate of Spread (+10%), and Above Ground Biomass (+10%). The number of MUs that experienced a class change under the most sensitive scenario was 78, representing 3.2% of the 2,429 MUs in the study, and highlighting the model’s general stability under perturbation.