From Land to Water: The Impact of Landscape on Water Quality Through Linear Models

Abstract

This work explores the relationship between landscape metrics and surface water in the Galicia-North Portugal Euroregion, employing 6,220,767 linear regression models through Python scripts to predict surface water quality. The Iberian Biological Monitoring Working Party (IBMWP) index, based on benthic macroinvertebrate communities from 40 sites across Portugal (PT) and Galicia (GL), served as the biological indicator. The models were initially selected based on linear regression assumptions (17 tests), and validated against real-world data, evaluating statistical performance through indicators such as R-squared, mean absolute percentage error (MAPE), and percentage bias (PBIAS). Results indicated that GL had a higher macroinvertebrate abundance, whereas Portugal showed greater diversity and family richness. Statistical analysis revealed that landscape significantly influenced water quality, with land use composition and configuration driving differences in ecological conditions between regions. The best-performing models demonstrated a high R-squared value of 0.7, a MAPE of 27% for calibration (PT) and 10% for the validation (GL), indicating a strong predictive relationship. The models provide valuable insights into the complex interactions between landscape patterns and water quality, highlighting how variations in landscape structure can directly affect ecological integrity. These findings reinforce the need for strategic land management to preserve water quality and emphasizing the importance of transboundary governance across the Euroregion to foster sustainable development.

1. Introduction

Water is crucial for human survival and development. The management of water resources has been alongside civilization for thousands of years [1], with legal frameworks dating back to 1750 B.C. [2]. Currently, managing water resources is a complex and dynamic process [3] that requires continuous adaptation. This complexity has increased with rising water demand driven by population growth [4], and climate change [3,5] which amplify both floods [6] and droughts [7]. As knowledge of pollution sources and their impacts on human health [8] and ecosystems [9,10] advances, concerns over the availability of consumable freshwater resources become even more pressing.

To guide water governance, the European Union (EU) implemented the Water Framework Directive (WFD) in the early 21st century (Directive 2000/60/EC) [11]. This directive establishes a comprehensive framework for managing water resources in terms of quality and quantity across the EU. Its implementation is monitored through River Basin Management Plans (RBMP), which propose measures to achieve WFD objectives [12] including the goal of attaining good or higher ecological status of surface waters by 2027 [13]. Achieving these goals requires systematic research and monitoring [14] to guide decision-makers [15].

Surface waters are complex systems influenced by interactions with groundwater [16] and climate [17]. They reflect both natural and anthropogenic processes occurring across connected catchments [18,19], as substances from soil are transported into rivers through runoff [20]. Pollution sources are generally classified into point source and diffuse emissions [21]. Point sources primarily include wastewater discharges from urban and industrial treatment plants [22,23]. Although these facilities aim to reduce contaminant loads, they can still degrade surface water quality [24], and even legally compliant concentrations may harm ecological integrity [25,26,27]. Diffuse emissions, on the other hand, originate from contaminants spread over large areas and transported by runoff and percolation [28]. In agriculture, excessive use of fertilizers and pesticides [29,30] poses significant threats as irrigation and rainfall carry pollutants to the surface and groundwater [31]. Livestock fields similarly contribute to organic pollutants and accelerate soil erosion [32]. Artificial surfaces exacerbate pollution through low infiltration rates, accumulating contaminants during dry periods that are washed into sewers and rivers during rainfall events [33].

Forests are the only land-use type that consistently benefits water quality [34]. When strategically located, forests act as barriers, retaining pollutant flows [35], particularly as riparian vegetation along river margins [36]. Therefore, effective forest management is crucial for preserving water resources [37]. Land-use composition and configuration also significantly influence water quality [38]. Their effects are often assessed through relationships between surface water parameters and landscape metrics [39]. These metrics quantify landscape features such as proportion, shape, and spatial distribution [40]. Given the multiple threats to water quality, multidisciplinary research is essential to comprehensively address the complex interactions affecting water systems and to support evidence-based decision-making [41].

To manage freshwater ecosystems effectively, various tools and techniques have been developed to assess water quality, analyze impacts on watershed security, and support decision-making [42,43]. Notable water quality models include the Soil and Water Assessment Tool (SWAT) [44,45], the Water Quality Simulation Program (WASP) [46,47], and the QUAL2K model [48,49]. These models are widely used to simulate water flow, nutrient transport, and pollutant dispersion. Statistical methods, such as regression analysis, also play a critical role in predicting water quality, sediment analysis, and assessing the impact of land cover on ecological integrity [50,51,52].

The integration of Geographic Information Systems (GIS) has further advanced water quality management by enabling the visualization and analysis of spatial and temporal pollution patterns [53,54]. Coupling GIS with statistical analysis and machine learning methods enhances both accuracy and adaptability in tackling water management challenges [55,56]. However, the use of these techniques is still limited by data availability, computational demands, and the need for specialized expertise [57,58].

A particularly effective approach involves the use of automatic regression algorithms to evaluate and predict water quality [59]. These algorithms efficiently explore combinations of variables, ensuring the assumptions of linear regression are met. By employing a “brute-force” strategy, the algorithm produces robust and accurate models for assessing and predicting water quality [60,61].

Building upon the critical need for effective water resource management outlined above, this study investigates the intricate relationships between landscape configuration and composition with water quality within the Galicia-North Portugal Euroregion. Integrating land-use composition and configuration into water quality assessments is increasingly recognized as vital. This approach helps to better understand and mitigate the impacts of human activities on freshwater ecosystems [38,39]. By leveraging the Iberian Biological Monitoring Working Party (IBMWP) index, a robust biological indicator of ecological status, this study bridges spatial and biological data to unravel the interactions between landscape patterns and riverine health. The application of Linear Regression Models, implemented in Python 3, not only ensures statistical rigor but also offers predictive insights into surface water quality. This aligns with the growing reliance on data-driven methods to inform sustainable decision-making in water governance [43,60].

The novelty of this study lies in its interdisciplinary approach, which combines advanced statistical tools with spatial and biological datasets to address regional challenges. The focus on the Galicia-North Portugal Euroregion is particularly significant, as it represents a transboundary area with shared hydrological networks and environmental policies under the Water Framework Directive (WFD). By analyzing two study areas, one in Spain and the other in Portugal, this research offers a comprehensive view of the determinants of water quality across political and ecological boundaries. Moreover, the findings contribute to the Euroregion’s shared objectives of achieving good ecological status for surface waters by 2027, as mandated by the WFD.

2. Materials and Methods

2.1. Study Area

The study was conducted in the Northwestern Iberian Peninsula, encompassing the Galicia-North Portugal Euroregion, a transboundary territory of approximately 51,000 km² (Figure 1). This region features a complex geological history dominated by igneous and metamorphic rocks of the Hesperian (Iberian) Massif and soils derived from granite and schist [62]. Land use, according to the Land Occupation Information System (SIOSE) [63], is diverse, with forests (49.41%), grasslands and shrublands (29.61%), agricultural areas (12.64%), and artificial surfaces (7.86%) and water bodies (0.49%). Forests, primarily maritime pine (Pinus pinaster Aiton 1789) and eucalyptus (Eucalyptus globulus Labill., 1800), often replace native oak species (Quercus rubra L., 1753), while riparian zones host willows, alders, poplars, ash trees, and elms, depending on soil and microclimatic conditions, the water regime of the water line, and human intervention. The area’s Atlantic climate, influenced by maritime and continental winds, supports productive ecosystems but faces seasonal water scarcity due to shallow aquifers and reduced storage capacity. Hydrological features include a dense river network, with drainage patterns shaped by tectonic structures and characterized by seasonal runoff variability [64,65,66].

The key river basins analysed (Figure 1) include the Minho, Lima, and Cávado in the Portuguese territory, as well as the Oitavén, Verdugo, Lérez, and Umia in Galicia, which discharge into the highly productive Rías Baixas—coastal systems characterized by deeply indented shores shaped by gradual subsidence [67]. The Euroregion is home to several protected areas, including the Peneda-Gerês National Park, the North Litoral Natural Park, Serra do Cando, several protected landscapes and special protection zones, RAMSAR sites (e.g., the Umia-O Grove Intertidal Complex), UNESCO biosphere reserves, and Geosites. These areas are critical for conserving biodiversity and ensuring ecological connectivity. However, human activities such as agriculture, livestock farming, and forestry have significantly altered native vegetation, leading to ecosystem degradation [68]. Despite its environmental and socio-economic challenges, the Euroregion offers unique opportunities for studying interactions between landscape metrics and water quality to support sustainable management strategies.

2.2. Data and Variable Selection for Water Quality Analysis

2.2.1. Water Quality Assessment Using Benthic Macroinvertebrate Assemblages

Benthic macroinvertebrates constitute a taxonomically and functionally diverse group of aquatic organisms, including insect larvae such as Ephemeroptera (mayflies), Plecoptera (stoneflies), and Trichoptera (caddisflies), as well as Diptera (e.g., Chironomidae), crustaceans, molluscs, and annelids (e.g., Oligochaeta). These organisms inhabit the substrates of rivers and streams and respond sensitively to changes in water quality, habitat structure, and hydromorphological conditions. Due to their relatively limited mobility and life cycles spanning several months, macroinvertebrate assemblages provide an integrated measure of ecological conditions over time, reflecting both natural and anthropogenic influences. Macroinvertebrate communities were sampled in a total of 40 sites, 7 in the Galicia region (sampled during the years 2010 and 2016) and 33 in Entre Douro e Minho (sampled in the years 2010 and 2015–2016). The higher number of Portuguese sites reflects broader data availability from national and regional WFD monitoring networks. In Galicia, only sites with complete information on IBMWP values and corresponding landscape metrics for the same sampling years were included, ensuring data consistency and regional representativeness.

In Portugal, the sampling was conducted during the spring/summer (dry season), in accordance with the WFD regulations, which standardize the timing of ecological assessments. In Galicia, the data corresponds to regional monitoring programs conducted throughout all four seasons. Although these campaigns were not specifically designed for direct comparison with Portuguese sites, they provide enough information to apply the IBMWP index and evaluate ecological quality, enabling meaningful regional comparisons.

Benthic macroinvertebrate sampling followed a comparable protocol in both Portugal and Spain, ensuring data harmonization across the transboundary study area. In both countries, sampling was based on a multihabitat kick-sampling approach using a 500 µm mesh D-frame net (Entomopraxis, Barcelona, Spain). Habitat types were selected proportionally to their coverage in the reach. Only habitats occupying at least 5% of the streambed were included. In Portugal, macroinvertebrates were collected in a 50 m reach centered around a riffle (25 m upstream and 25 m downstream), using six sampling units of 1 m × 0.25 m (total sampled area ≈ 1.25 m²). These units were proportionally distributed across available microhabitats. Inorganic substrates included bedrock/boulders, stones, gravel/pebble, sand, silt and clay, while organic substrates comprised macrophytes and algae, and coarse particulate organic matter. The resulting composite sample was preserved in 99% ethanol (AGA, São Brás de Alportel, Portugal). In Spain, sampling was conducted over a 100 m reach, using 20 standardized kick samples of 0.5 m × 0.25 m (0.125 m² each), covering a total area of 2.5 m². Sampling effort was allocated proportionally to the estimated habitat cover, and the same net type and upstream sampling direction were used as in Portugal. Despite minor differences in reach length and total sampled area, both protocols are officially recognized and routinely used in their respective countries for the Iberian Biological Monitoring Working Party index (IBMWP). Therefore, both methods are valid and reliable, providing comparable results without compromising data accuracy or ecological integrity.

Despite the implementation of the WFD and the development of new biological indices based on macroinvertebrates, the Iberian Biological Monitoring Working Party index (IBMWP) [69,70] remains widely used. Due to its proven reliability and widespread use in both Portugal and Spain, this IBMWP index was chosen. Adapted from the United Kingdom BMWP index [71], it incorporates specific Iberian families, modifies some tolerance scores, and correlates its values with five levels of contamination [69]. This makes it highly suitable for regional ecological assessments. Therefore, both methodologies were implemented according to WFD requirements and are fully comparable for supporting ecological status assessments under this regulatory framework.

The IBMWP method evaluates water quality based on the different tolerance levels of aquatic invertebrates to organic pollution. Sensitive species like mayflies and stoneflies, which indicate cleaner water, are given a high tolerance score of 10, while less sensitive species such worms (Oligochaeta) receive a score of 1. The presence and diversity of macroinvertebrates are key indicators. Higher water quality supports a greater variety of sensitive species, which signals fewer pollutants. The score of the IBMWP index (a single metric index, [72,73]) is calculated by summing the sensitivity weightings (tolerance scores) of the taxa present in a sample. The IBMWP score ranges from 0 to >100. Its interpretation is as follows: very good quality (IBMWP score > 100), good quality (IBMWP score between 61 and 100), fair quality (IBMWP score between 36 and 60), bad quality (IBMWP score between 16 and 35), and very bad quality (IBMWP score < 16).

2.2.2. Landscape Metrics

For each sampling site, the corresponding drainage area was delineated using ArcGIS ArcHYDRO Tools 10.8 (Esri, Redlands, CA, USA) and Copernicus DEM 25 × 25 m (Copernicus/European Environment Agency, Copenhagen, Denmark), as input data [74]. Flow direction, flow accumulation, and drainage lines were first calculated. Next, the drainage area from the upstream point to the sampling location was delineated using ArcHydro [75].

Landscape metrics (LSM) were calculated from the CORINE Land Cover (Copernicus/European Environment Agency, Copenhagen, Denmark, 2018) to quantify landscape configuration and composition for each drainage area using various indices. The ZonalMetrics Toolbox (Adamczyk & Tiede, University of Salzburg, Austria) [76] introduced in ArcGIS 10.8 (Esri, Redlands, CA, USA) was used for this purpose. It provides an intuitive graphical user interface and allows batch calculations. Metrics were calculated for each land use type: agriculture (AGR), forest (FOR), and artificial areas (ART), as outlined in

Table 1. Calculated landscape metrics, and respective nomenclature. ART stands for artificial surfaces, AGR for agricultural areas, and FOR for forests and semi-natural areas. ALL stands for landscape metrics that include all land uses in the calculation of the metric.

Nomenclature	Description	Units	Land Use Category
			ART	AGR	FOR	ALL
			Artificial	Agriculture	Forest	All
ppc_(LUA)	percentage of land use patches	%	LUA	LUA	LUA
pz_(LUA)	area proportion occupied by LUA	%	LUA	LUA	LUA
lpi_(LUA)	area percentage occupied by the large land use patch of LUA	%	LUA	LUA	LUA
sol_(LUA)	percentage of land occupied by the of LPI for each land use type.	%				LUA
shdi_(LUA)	Shannon’s diversity index of LUA		LUA	LUA	LUA	LUA
ed_(LUA)	edge length LUA/total area	m/km2	LUA	LUA	LUA	LUA
ed_spec_(LUA)	edge length LUA/area LUA	m/km2	LUA	LUA	LUA
edp_(LUA)	edge length of LUA by total edge length	m/m	LUA	LUA	LUA
cce_(LUA)/(LUB)	edge length shared between LUA and LUB divided by the edge length of LUB	m/m	LUA and LUB	LUA and LUB	LUA and LUB
cedp_(LUA)/(LUB)	edge length shared between LUA and LUB divided by the edge length of all land uses	m/m	LUA and LUB	LUA and LUB	LUA and LUB
cedd_(LUA)/(LUB)	edge length shared between LUA and LUB divided by total area	m/km2	LUA and LUB	LUA and LUB	LUA and LUB

. Additionally, some metrics, such as the Shannon diversity index, were calculated across all land use types. In total, 42 metrics were derived to characterize the land use configuration and composition. These metrics were subsequently used as candidate regressors in the models.

2.3. Data Analysis

2.3.1. Linear Regression Models for Water Quality Prediction

Once the input data from the 33 Portuguese samples were prepared, linear regression models were estimated using a brute force approach [59] in Python 3 (Python Software Foundation, Wilmington, DE, USA), Supplementary Materials S1. The objective was to identify all possible combinations of landscape metrics (independent variables) that could predict a surface water quality parameter IBMWP (dependent variable) (Figure 2). The algorithm used is thoroughly described by Fernandes et al., 2023 [59]. In summary, running the algorithm requires defining several inputs: candidate regressors (LSM), the dependent variable (IBMWP), the minimum and maximum number of regressors, the statistical significance threshold, and maximum Variance Inflation Factor (VIF). For this study, the number of regressors was set between 1 and 6, a statistical significance level of 0.05 and the maximum VIF among regressors was set at 5 to ensure minimal multicollinearity and reliable estimation of individual effects on the IBMWP index.

The algorithm generated all possible combinations of regressors within the thresholds. For each combination, it computed the linear regression model, VIF for each regressor, R², adjusted R², the significance of the regressors, and results of heteroscedasticity and error normality tests. If the defined thresholds are not met, the algorithm proceeded to the next combination of candidate regressors, continuing until all combinations were generated and evaluated. The final output is a list of linear regression models, including all relevant statistical test results and regression coefficients (Supplementary Materials S2).

The algorithm does not perform autocorrelation tests, as they are typically required when each sample corresponds to a different timestep. In this study, each sample represents a drainage area. Therefore, spatial autocorrelation of the error term was separately assessed using Moran’s I test [77] to ensure that the residuals were not spatially clustered (Supplementary Materials S2).

The resulting models were then analyzed considering the assumptions of linear regression. They were calibrated with data from Portugal (33 samples) and validated with data from the Galicia region (7 samples).

2.3.2. Model Assessment

The output models generated by the algorithm adhere to the predefined statistical thresholds. However, selecting a single “best” model is not straightforward, as multiple factors need to be considered. Key features of suitable models include high R² values and high p-values for heteroscedasticity and error normality tests, as the regression model is valid when the null hypothesis is accepted. Additionally, models with low Zscores were considered preferable. Model performance was further evaluated using Mean Absolute Percentage Error (MAPE), Percentage Bias (PBIAS) and Normalized Root Mean Square Deviation (NRMSD) [59,78,79] These metrics were calculated for predictions in the Galicia drainage areas and compared with the observed values (

Table 2. Rating of water quality prediction models, based on the Percentage Bias (PBIAS) and Mean Absolute Percentage Error (MAPE), adapted from Salis et al., 2019 [79] and Fernandes et al., 2023 [59].

Rating	|PBIAS|	MAPE
High	0–10%	0–10%
Good	10–15%	10–20%
Reasonable	15–25%	20–50%
Low	>25%	>50%

).

3. Results and Discussion

3.1. Macroinvertebrate Diversity and Water Quality

In this study, 1,926,343 benthic macroinvertebrate individuals belonging to 106 families were identified across the 40 studied sites (7 GL and 33 PT), coordinates are provided in the Supplementary Materials S2. Macroinvertebrate density varied significantly. The highest density was recorded at station GL_SS_1 (23,662.4 ind/m²) and the lowest at PT_T22 (25.6 ind/m²). Mean density also differed markedly between regions: GL had a mean of 9713.9 ind/m², while PT averaged 815.5 ind/m². Despite GL sites having higher macroinvertebrate abundance, PT sites exhibited greater family richness, with 103 families compared to 75 in GL. The Shannon-Wiener diversity index in GL ranged from 1.95 to 2.83, with the highest value at GL_SS_4 and the lowest at GL_SS_1 (1.36). These diversity values were moderate, reflecting a community dominated by a few families (Chironomidae, Simuliidae, Elmidae, Baetidae, Hydropsychidae, Leptophlebiidae, and Nemouridae). This dominance results in lower overall diversity. In contrast, PT sites exhibited a broader range of Shannon-Wiener diversity values, from 0.17 at PT_T8 to 3.08 at PT_T55. Overall, PT demonstrated higher diversity, particularly at sites such as PT_T55, PT_T50, and PT_T3, where values exceeded 2.5. This aligns with the earlier finding that PT sites, despite their lower macroinvertebrate abundance, exhibit greater family diversity and richness. Although some methodological differences exist between countries, both protocols are based on multihabitat kick-sampling. They were designed to capture the ecological heterogeneity of stream reaches. This common focus ensures comparability, especially for the application of the IBMWP index, which relies on the presence of key macroinvertebrate families rather than absolute abundance. The lower abundance values recorded in Portuguese sites likely reflect the smaller sampled area rather than ecological differences alone. However, because both methodologies prioritize the proportional representation of dominant habitats, they remain fully compatible and valid for assessing ecological quality under the IBMWP framework.

The significant differences observed between GL and PT regions align with ecological patterns documented in other studies. For instance, noted a similar trend in river systems, where sites with higher abundance often showed dominance by a few tolerant species, resulting in reduced diversity. This pattern is evident in the GL region, where the high abundance is accompanied by moderate diversity due to community dominance by a limited number of families. Conversely, the findings in the PT region [80], with its greater richness and diversity (103 families), align with studies like Zhang et al., 2019 [81]. These studies showed that regions with more diverse and heterogeneous habitats tend to support richer macroinvertebrate communities. The Shannon-Wiener index values further substantiate this trend, with PT sites exhibiting higher diversity compared to GL. Similarly, studies such as Dolédec et al., 2006 [82], highlight that high diversity, even in the context of lower abundance, reflects a more balanced and ecologically resilient structure, a pattern evident in the results of the present study.

The high IBMWP scores across both regions, with 85% of sites showing very good biological quality, align with previous research linking biological monitoring indices to water quality assessments. Scores were consistently high across Portuguese and Spanish sampling sites, with no significant differences detected between regions (Kruskal–Wallis: KW-H (1;10) = 0.1169, p = 0.7324; ANOVA: F(1;8) = 0.064, p = 0.8066) (Figure 3). This pattern reflects preserved or moderately impacted ecosystems in both areas, as noted by Menezes et al., 2010 [83].

Of the 40 sites studied, 34 (85%) were classified as having very good biological quality encompassing all GL sites (100%) and most sites in the Entre Douro e Minho region (82%). However, in the remaining 18% of Entre Douro e Minho sites, water quality was more varied, spanning three quality classes: good (9%), fair (6%), and bad (3%). Notably, the variability in IBMWP scores was higher for Portuguese sites compared to GL ones, where the scores exhibited minimal variability, suggesting more consistent conditions in the GL region (Figure 3).

The absence of significant differences between annual IBMWP values in the GL region over the seven-year study period (KW-H (6;78) = 7.5447, p = 0.2734; F(6;71) = 1.4357, p = 0.2132) further underscores the stability of water quality in these ecosystems (Figure 4). Annual trends showed high variability in earlier years (2010–2011), followed by a moderate decrease in variability in later years (2013–2016). This trend, possibly influenced by environmental and sampling factors, suggests that while the median IBMWP values showed a slight downward trend, these fluctuations did not significantly impact water quality based on biological indicators (macroinvertebrates) (Figure 4). These results, combined with statistical tests showing no significant differences between regions or across years, align with findings by Feio et al. 2014 [84], who reported that biological indices remain stable across regions with diverse land uses, provided ecosystems are balanced. Despite differences in macroinvertebrate abundance and richness, the comparably high IBMWP scores in both regions reflect the overall health and stability of these aquatic ecosystems.

These findings have direct implications for management and conservation strategies. The higher family richness and diversity observed in Portuguese sites suggest that maintaining habitat heterogeneity is crucial for supporting resilient macroinvertebrate communities. Conversely, the dominance of a few tolerant species in GL sites, despite higher abundance, indicates that management should prioritize measures to increase habitat complexity and connectivity to promote ecological balance. In practice, monitoring efforts can focus on key strategic locations across the Euroregion to identify potential stressors and implement targeted restoration actions, such as riparian buffer enhancement or reforestation, which can help maintain or improve water quality.

3.2. Landscape Metrics Assessment

The analysis of landscape metrics highlights differences in land management practices between Galicia (GL) and Portugal (PT). These differences are particularly notable for agricultural and forested areas. A total of 42 landscape metrics, grouped into 11 categories, were calculated for different land uses. The values compared between the two regions (n = 33 for PT and n = 7 for GL) using the Shapiro–Wilk test (Figure 5). Comparisons were made using the Mann–Whitney U test for non-normal distributions or Student’s t-test for normal distributions. Significant differences (p < 0.05) were found in 13 metrics, highlighted with an asterisk (*) in Figure 5.

GL had a higher percentage of agricultural patches ppc_(AGR)*. In contrast, PT had higher percentages of forested patches ppc_(FOR)* and greater diversity indices in forested areas (shdi_(FOR)*). Discrepancies in the range of values between the two regions are evident in boxplots. GL shows a narrower range for some metrics, despite its smaller sample size. These findings suggest that management practices differ between the two regions, affecting landscape structure and ecological functions. PT shows greater diversity of land use types, with a mix of agricultural and forested patches. This diversity may contribute to higher biodiversity, consistent with findings from studies by Bieger et al., 2017 [85] and Dala-Corte et al., 2020 [86]. In contrast, the dominance of agricultural patches in GL reflects land use stress. This may explain its lower macroinvertebrate richness, a pattern similarly reported in agricultural watersheds by Marzin et al., 2012 [87]. This study highlights the loss of forested land to agricultural expansion, as noted by other studies in the area [88]. Regulatory frameworks also influence these observed differences. In Galicia, the Hydraulic Public Domain Regulation [89] and the Forest Law of Galicia [90] historically allowed forested land to be converted to agricultural use. These regulations often prioritized agricultural productivity over forestry. As a result, forested areas have decreased, negatively affecting connectivity and riparian biodiversity [88]. Reversing agricultural land to forestry is more complex. It is only permitted on abandoned agricultural land registered in a land bank for at least two years and requires prior notification to the forestry authority. Reforestation is only allowed under strict conditions. These include proximity to existing forested areas or the creation of up to 5-hectare enclaves using deciduous broad-leaved trees. This framework was in effect until its recent amendment under the Agricultural Land Recovery Law of Galicia [91], which came into force after the period analyzed in this study. In contrast, PT exhibited more diverse agricultural and forested landscapes, which are likely to support a richer ecological community and enhanced water quality regulation. However, the smaller sample size for GL limits the generalizability of these findings, as highlighted by Turner and Pham, 2015 [92]. Further studies with larger datasets are needed to confirm these patterns.

The analysis also underscores the impact of land ownership fragmentation, particularly in GL. As highlighted by Corbelle-Rico & López-Iglesias 2024 [93], fragmented ownership increases risks such as forest fires and flooding, complicating sustainable land management. These challenges emphasize the need for integrated governance strategies that balance agricultural and forestry interests [94], promoting a balance between agricultural production and biodiversity conservation, particularly in ecologically valuable areas.

From a management perspective, these landscape differences underline the importance of tailored land-use planning. In GL, strategies to reduce agricultural expansion or to restore forest patches in riparian zones could enhance connectivity and buffer capacity, mitigating runoff and improving water quality. In PT, preserving the diversity of forested and agricultural patches is essential to maintain ecosystem services. These insights support integrated, cross-border management policies that address both biodiversity conservation and sustainable land use, ensuring the long-term protection of aquatic ecosystems.

3.3. Global Model Results

From the python script, a total of 6,220,767 linear regression models were tested to predict IBMWP from LSM in Portugal (n = 33). Among these, only 630 (0.01%) met all linear regression assumptions, and their statistical and performance metrics are shown in Figure 6. Individual metric values are provided in Supplementary Materials S3. According to the predefined thresholds of the script, the maximum VIF among regressors was below 5, while the p-values for model explanatory power and regression coefficients were all < 0.05. The minimum p-values for error normality and heteroskedasticity were above 0.05, indicating that residuals followed a normal and homoscedastic distribution. The spatial autocorrelation of residuals was also close to zero, suggesting no spatial dependence in the model errors.

MAPE for IBMWP predictions using Portuguese data (calibration) ranged from 22% to 52%, revealing a reasonable calibration performance (Mape-Cal, Figure 6). For Galician data (validation), MAPE, PBIAS, and NRMSD exhibited a broader range of values, yet still achieved a high-performance rating according to Table 1.

Principal Component Analysis (PCA) further clarified the relationships among statistical variables, explaining 31% and 17.3% of total variability through its first two axes (Dim1 and Dim2), respectively (Figure 7). Clusters dominated by R², adjusted R² (R2_adj), and VIF indicated strong positive correlations, whereas error metrics (MAPE, NRMSD, and PBIAS) formed a distinct group. Contrasting alignments (e.g., between p_he and Zscore) revealed negative correlations, while suggested vectors (e.g., between p_he and R²), suggested weak or negligible associations.

Overall, these results demonstrate that landscape metrics can effectively predict water quality, providing a robust basis for data-driven management decisions. From a practical perspective, managers can apply these models to identify areas where landscape modifications, such as increasing forest patch connectivity or managing fragmented agricultural land, are likely to yield the greatest ecological benefits. The consistent performance across calibration and validation datasets indicates that certain models may be transferable between regions, supporting coordinated cross-border water management.

Finally, the mechanisms underlying these results, whereby landscape composition and configuration regulate hydrological buffering, sediment retention, and nutrient dynamics, have been previously demonstrated in Iberian catchments [95]. These processes underpin the predictive performance of landscape-based models for water quality, reinforcing their utility for regional-scale management and spatial planning.

3.4. Best Models

Based on the results from Figure 6, 12 models were selected as the best ones, within 630 from the python script. These were selected based on the 25th–75th %iles of the statistical/performance metrics. The respective selection criteria were: an adjusted R² > 0.55; a MAPE-Cal <32.0%; MAPE-Val <11.6%; a |PBIAS|< 9.9 and a NRMSD (%)> 14.5%. For a min (p value)—error normality tests and min (p value)—error heteroskedasticity, besides the output models are already above 0.05 as required for linear regression analysis, was selected values > 0.31 and > 0.17, respectively. For Moran’s I Z- score for the error term, it was chosen only models between −1.65 and 1.65, since these indicate a random spatial autocorrelation. In

Table 3. Performance of the selected models.

Model ID	36611	36698	153591	153678	153698	229124	341269	341765	356391	481290	481293	481296
nº of regressors	4	4	5	5	5	5	5	5	5	5	5	5
Max (VIF)	1.77	2.38	2.27	3.29	3.28	2.55	1.90	1.73	2.28	1.56	1.65	1.58
R2	0.63	0.62	0.65	0.66	0.68	0.63	0.65	0.66	0.64	0.64	0.63	0.63
R2 adjusted	0.57	0.56	0.58	0.60	0.62	0.56	0.59	0.59	0.58	0.57	0.56	0.56
max (p-value)—regression coefficients	0.01	0.05	0.02	0.03	0.03	0.03	0.03	0.05	0.02	0.05	0.03	0.04
min (p-value)—error normality tests	0.39	0.50	0.74	0.34	0.49	0.46	0.65	0.57	0.43	0.65	0.70	0.73
min (p-value)—error heteroskedasticity	0.42	0.32	0.34	0.24	0.21	0.23	0.51	0.18	0.18	0.57	0.68	0.33
MAPE (Calibration) [%]	26.80	29.92	26.73	28.67	27.08	31.78	31.09	32.01	31.95	29.39	29.03	29.39
MAPE (Validation) [%]	8.14	9.63	9.14	10.47	10.45	10.94	10.91	11.17	10.85	8.64	9.23	8.99
PBIAS [%]	3.74	1.33	8.27	5.05	7.82	8.27	9.04	9.36	7.70	6.98	6.61	7.68
NRMSD [%]	12.28	10.37	14.11	12.10	13.51	12.77	12.38	12.53	12.86	13.24	14.12	13.49
Zscore	0.71	1.60	0.40	1.23	1.65	1.06	1.08	1.00	1.05	0.86	0.90	1.25

, performance metrics are shown for each selected model, and Figure 8, the scatter distribution between observed and forecasted values of IBMWP can be seen. Still, in the Supplementary Materials S4, individual plots for all the 630 models can be found.

The landscape metrics retained in the best-performing models are biologically meaningful. Thus, higher forest cover and greater landscape heterogeneity are associated with improved riparian habitat quality, reduced sediment and nutrient runoff, and increased availability of allochthonous organic matter. These conditions favor sensitive macroinvertebrate taxa. Conversely, metrics reflective of agricultural dominance and fragmentation are linked to increased diffuse pollution and habitat simplification, promoting assemblages dominated by tolerant taxa [95,96]. These patterns reinforce the ecological basis of the statistical models and highlight the strong link between land management and riverine biotic integrity across the Euroregion.

The algorithm to calculate linear regression models was executed for a maximum of 6 regressors; still, the best models had from 4 to 5 regressors, resulting in a total of 12 models,

Table 4. Regression coefficients of the selected models.

Model	36611	36698	153591	153678	153698	229124	341269	341765	356391	481290	481293	481296
Nº of regressors	4	4	5	5	5	5	5	5	5	5	5	5
intercept	344.7	459.1	465.5	572.7	574.1	403.7	379.7	399.4	409.0	461.5	452.8	442.1
ppc(ART)			−2.9	−2.4	−2.6
pz_(ART)						−13.9
ppc_(AGR)	1.9	1.3				2.4	2.2	1.5	2.5
ppc_(FOR)			−1.8	−1.6	−1.6					−2.0	−2.1	−2.0
lpi_(ART)							−26.4	−23.7		−26.9	−31.8	−28.3
lpi_(FOR)			−1.1	−1.9	−1.9	−1.8	−1.6		−1.8
sol_(ALL)	−1.0	−1.6						−1.6		−1.0	−0.9	−0.9
shdi_(ART)	−73.2		−62.7							−66.5	−72.7	−63.8
shdi_(ALL)		−196.1		−214.9	−211.8
6_edp_(AGR)	−4.4	−3.2	−2.9	−2.4
6_edp_(ART)									−5.1
cce_(AGR)_(ART)						−1.0	−0.8	−0.7	−1.0
cedp_(AGR)_(AGR)										−3.1
cedp_(FOR)_(AGR)					−2.6	−5.4			−5.7		−3.1
cce_(AGR)_(FOR)							−2.5	−2.0				−1.5

. In common, these models have an average intercept of 447, which is the base IBMWP. The regressors have an equal sign across all models, which represents consistency across the selected models. The only positive regressor is ppc_(AGR), the percentage of agriculture land use patches, meaning that having fragmented agriculture areas safeguards hydric resources, as it creates a lower runoff contaminant loads comparatively linked agriculture areas in the same land use patch. However fragmented forested ppc_(FOR) decrease the capacity to retain contaminants in land to water [97]. The signs of the regressors is accordant to the correlation between the LSM and IBMWP (Supplementary Material S4), except for lpi_(FOR) which is poorly correlated with IBMP (Pearson’s r = 0.09), which has negative regression coefficients. Besides the signs are mostly consistent, it should not be blindly interpreted that a negative LSM represents a negative phenomenon on water quality, since the negative/positive sign results in a statistical approach to fit the model. Also, the LSM can better portray phenomena that decrease water quality, when used under transformations, or under other nonlinear models.

3.5. Final Considerations

Nevertheless, the primary innovation of this method lies in its ability to merge traditional ecological concepts with modern analytical technologies, driving a substantial advancement in conservation and sustainable management practices. Thus, the integration of statistical models and the use of the IBMWP index represent important advances in unifying evaluation criteria for river basin management across the Euroregion. This approach represents a milestone in the field of biomonitoring, expanding its applicability to a wide range of environmental scenarios.

The study also emphasizes the need for cooperative governance, as current disparities in land and water management practices between Spain and Portugal hinder effective and sustainable resource management. The results align with findings by White et al., 2021 [98], who underscore the importance of robust statistical frameworks for environmental modeling, and Gutiérrez-Rial et al., 2023 [99], who advocate for common indices in cross-border assessments.

The outcomes underscore the critical need to balance land use management with ecological and water quality goals. The conversion of forested land to agriculture, often driven by socio-economic needs, must account for its long-term impacts on river ecosystems, including biodiversity loss and increased flood risk. At the same time, the study highlights the importance of implementing adaptive measures to mitigate the effects of climate change and to protect the resilience of these systems. Moving forward, the integration of these models into a regional governance framework, as proposed by Fernandes et al., 2023 [59] will be essential for advancing sustainable land and water management strategies that reflect the shared environmental challenges and opportunities of the Galicia-North Portugal Euroregion. Despite these challenges, the study provides valuable insights into the relationships between model evaluation metrics, statistical indicators, and spatial characteristics, offering a solid foundation for water quality prediction in varying landscapes.

Overall, this study provides actionable insights for policy makers and water managers. By linking landscape configuration with macroinvertebrate-based water quality indices, we offer evidence-based guidance for prioritizing interventions, such as riparian reforestation, connectivity restoration, and adaptive land-use policies. Our findings align with previous evidence that landscape configuration (forest cover, patch connectivity, fragmentation) exerts a primary control on stream ecological integrity and can therefore guide targeted restoration and land-use planning measures [43,100,101]. Integrating these strategies within a coordinated Euroregion governance framework can enhance resilience to climate change, protect biodiversity, and maintain water quality across shared river basins.

4. Conclusions

This study demonstrates the importance of integrating landscape metrics and biological indicators to assess water quality in the Galicia–North Portugal Euroregion. Our results reveal that land management practices directly influence riverine ecological health: GL sites, with higher macroinvertebrate abundance but lower diversity, reflect the dominance of agricultural landscapes, whereas PT sites, with greater diversity and family richness, illustrate the benefits of heterogeneous land-use mosaics for biodiversity and water quality regulation.

Linear regression models and landscape metric analyses confirm that forest loss and agricultural expansion reduce ecological richness, while diverse and connected landscapes enhance ecosystem resilience. Despite regional differences, consistently high IBMWP scores indicate generally robust ecological conditions across the Euroregion.

These findings have clear practical implications for water and land management. They highlight the need for integrated, adaptive governance strategies that maintain habitat heterogeneity, support biodiversity, and mitigate the impacts of land-use change. Monitoring and management efforts can focus on key sites to implement targeted restoration actions, such as riparian buffer enhancement or reforestation, thereby safeguarding water quality and ecological resilience in transboundary river systems.