#### 3.1. Spatial Data

Figure 3 illustrates the spatial distribution of pressures in the Ave River Basin.

Figure 3A depicts the land use map of 2015. It can be noted that 50% was occupied by the dominant land use, forest and seminatural areas. Agricultural areas occupied 30% of the river basin, 20% was artificial surfaces, and less than 0.5% was water bodies. The values of discharged nitrogen from livestock, forest, and agricultural areas in the river basin catchments can be seen in

Figure 3B,D. The scattered effluent discharge sites are represented in

Figure 3C. Among a total of 60 locations, 24 were discharge sites of industrial treatment plants, while 36 were from domestic sewage treatment plants.

The twelve river sites where the IPtI

_{N} was measured are represented in

Figure 1, numbered from 101 to 112.

Table 2 depicts the IPtI

_{N} values and respective classification. During the winter of 2017, the ecological status was classified as “Excellent” in two sites, “Moderate” in 4, “Poor” for another 4 sites, and “Very Poor” in 2. Overall, the values increased from winter to summer, as none of the sites was classified as “Very Poor” in summer, 6 were classified as “Poor”, 3 as “Moderate”, and 3 as “Excellent”. In the locations 103, 105, 111, 112, 107, and 108, the classification changes were minimal. Besides, in site 111 the classification changed from “Moderate” to “Poor” because the IPtI

_{N} value was very close to the class threshold. In site 106 the classification changed from “Poor” to “Moderate”, but in sites 109, 110, 103, and 104 the decrease in ecological status was startling because there was a significant decrease in IPtI

_{N} and subsequent classification of a less-rich class. For site 101 there was a significant increase, from “Poor” to the maximum class “Excellent”. These changes of values were dependent upon seasonal effects but also on the pressures in surface waters.

#### 3.2. Interpretation of a PLS-PM Example Model

The output models of SmartPLS were all similar to the one represented, as an example, in

Figure 4. Each measured variable (MV) is represented as a yellow rectangle, and latent variables (LVs) as blue circles. Inside the LVs preceding other LVs, the R-squared value is portrayed. In the example model, only “Ecological Integrity” has an R-squared value, since this is the only variable that has a measured score (calculated by the sum of the product of MVs with the own weight) and a predicted score (calculated by the sum of the product between the LVs). In a PLS-PM model, weights and path coefficients are determined through an iterative process, termed the path algorithm [

83], with the purpose to maximize the R-squared value.

Several PLS-PM models were built and analyzed in this study. In order to exemplify how these models can be interpreted, a PLS-PM model is demonstrated in

Figure 4, where the data were gathered for the drainage sections within a distance of 4 km and IPtI

_{N} values measured during the winter of 2017.

Each LV was formed by one or more MVs. For the present case study, an LV “Land Use” was created and composed of 4 MVs, namely “Diversity”, “Forest”, “Agriculture”, and “Artificial”. The LV “Contaminant Emissions” was composed of three MVs pertaining to different types of contaminant flows: “Point Source”, “Livestock”, and “Forest and Agriculture”. “Ecological Integrity” was formed by a single MV, which is the IPtI

_{N}. This LV accumulated the effects of the other LVs that were pressures in surface waters, which is why “Contaminant Emissions” and “Land Use” were connected to “Ecological Integrity”. The PLS-PM model was divided into two sub-models, inner and outer. The equations of the measured scores of each LV were calculated according to Equations (2)–(4) for “Land Use”, “Contaminant Emissions”, and “Ecological Integrity”, respectively, and are the equations that composed the outer models. The inner model was composed of relations between latent variables, which, in this case, is solely expressed by Equation (5).

To interpret the example model, the weights and path coefficients should be analyzed simultaneously, which can be viewed in Equation (6), where the combination of Equations (2)–(5) is made. For example, the MV “Diversity” has a positive weight (0.208), so it increases the LV “Land Use”, while the same applies to “Agriculture” (0.561). Conversely, “Forest” and “Artificial” have negative weights, namely −0.569 and −0.033, and therefore decrease “Land Use”. But since the path coefficient of “Land Use” in “Ecological Integrity” is negative (−1.085), “Diversity” and “Agriculture” are variables that decrease “Ecological Integrity” because the product of the weight and path coefficient is negative: −0.226 and −0.560, respectively. On the other hand, “Forest” and “Artificial Areas” increase “Ecological Integrity”, since the product between the path coefficient and weight is positive, respectively 0.617 and 0.036. Equation (6) expresses the total effect of each pressure in “Ecological Integrity”. The results of this study were based on the analysis of the product between the path coefficients and weights (termed pcw) for the studied 26 models.

#### 3.3. Results of All PLS-PM Models

As shown in

Table 2, the IPtI

_{N} values were collected during two seasons, winter and summer. For this reason, the 26 PLS-PM models were divided into two groups, winter (2017) and summer (2017), and traced as two dot arrays colored as blue and red, respectively, in

Figure 5. For each model in the respective group, the pressure values were used as input data, gathered from the 13 drainage sections and the IPtI

_{N} values collected in winter or summer.

Figure 5 portrays the results of the PLS-PM models. Each graphic describes the pcw of a measured variable in all models (

y axis). The

x axis represents the logarithm of the buffer distance for the respective model. For the distances of 100, 250, 500, 1000, 2000, 3000, 4000, 5000, 7000, 10,000, 15,000, 20,000, and 56,000 meters, the log

_{10} scores were 2, 2.4, 2.7, 3, 3.3, 3.5, 3.6, 3.7, 3.8, 4, 4.2, 4.3, and 4.7, respectively. The purpose of the plots was to illustrate the effects of the pressures in “Ecological Integrity” (“IPtI

_{N}”).

The effect of “Artificial” was independent of the season since the variations with distance were practically identical for both winter and summer. For distances shorter than 10 km, the effect was positive, but for longer distances the effect became negative. The strongest positive effects were detected for a distance of 100 m in summer (pcw = 0.386) and for 1000 m in winter (pcw = 0.310). The strongest negative effects were detected for the maximum distance (56 km) (i.e., for the entire drainage areas) either in winter (pcw= −0.247) or summer (pcw= −0.201).

For “Agriculture” it was seen that for both winter and summer periods, the effect was practically identical, but the summer line was below the winter line for a majority of buffer distances (only between 3 km and 5 km is the red line above the blue). The effect of agriculture was practically null for a distance of 100 m in summer (pcw = 0.008). For the same distance, it was positive during winter (pcw = 0.276) and practically null for a distance of 250 m (pcw = 0.01). For longer distances, the effect was negative, which indicated that agriculture decreased water quality. Peak values were found for distances of 4 km in winter (pcw = −0.560) and 10 km in summer (pcw = −0.648), but for distances larger than 10 km, the changes were minimal. The results lead to the conclusion that, for the Ave River basin, agriculture is a threat to water quality, while the impact seems to be stronger during the summer period.

Peak values of “Diversity” were detected for the minimum distance, 100 m, either in winter (pcw = −0.856) or in summer (pcw = −0.456). The effect was always negative for winter periods, and stronger in this season for almost all distances, except between 4 and 5 km. For the summer period, the effect was close to zero, but still negative; only distances of 500 m, 10, and 15 km were close to zero, since the pcw values were −0.024, 0.059, and 0.026, respectively. For the longest distance (56 km), the effect was practically the same for winter (pcw = −0.290) and summer (pcw = −0.298). The results provide evidence that the impact of “Diversity” is a threat in winter.

The effect of “Forest” was essentially positive. For the winter period, the effect was positive for all distances, and always higher than in summer, where the effect was negative but close to zero for the distances 500 m (pcw = −0.080) and 56 km (pcw = −0.049). Peak values occurred on the shortest scale, 100 m. In winter the pcw was 0.853, and in summer it was 0.419. As the buffer distances increased, the effects changed irregularly (drop and rise), but from an overall view, values were always between 0.853 and 0.374 during winter. It can be said that globally (for winter and summer), “Forest” favors water quality.

The variable that had less effects along all distances and seasons was “Point Source”. For this variable and the models that comprehended distances between 100 to 500 m, the attributed weight was 0, since for short distances there were no discharge points. Even so, compared to all the other variables, it had less impact because the pcw values were contained in a short range that varied from −0.198 to 0.226. For winter, negative values were found in 1 km (pcw = −0.198) and 7 km (pcw = −0.006), while in summer the negative values were observed for distances longer than 7 km. This indicates that the effect of effluent discharges only decreased IPtI_{N} values during the summer period when long distances were analyzed, but with minimal impact.

The discharges of nitrogen from diffuse pressures were represented by the variables “Forest and Agriculture” and “Livestock”. When both graphics were compared, it was seen that there was an inverse relationship between these two variables for all models in both seasons. When the effect “Forest and Agriculture” increased, “Livestock” decreases. For the summer period, the effect of “Livestock” was always negative, while the effect of “Forest and Agriculture” was always positive. These effects were stronger over shorter distances, since maximum values for “Forest and Agriculture” and minimum values for “Livestock” appeared over the short distances. But as the distance increased, both effects approached zero. The variations of both variables were minimal for short distances (≤1 km), positive for “Forest and Agriculture”, and negative for “Livestock”. At the distances 2, 3, and 4 km, the effect became positive for “Livestock” and negative for “Forest and Agriculture”, with a peak at 4 km. For distances longer than 4 km, the effect tended to zero.

The analysis of the pcw variations for the 7 measured variables for all the models is crucial to comprehend the cause–effect relationship changes as function of season and distance. But the analysis of the R-squared values (

Figure 6) reveals the models’ capacity to explain IPtI

_{N} variations.

The calculated R-squared values of summer varied less than the winter counterparts. The range of values varied from 0.75 (1 km) to 0.91 (56 km) in summer, while in winter they ranged from 0.58 (500 m) to 0.93 (7 km). For the winter period, for distances comprehended between 250 m and 3 km, the model’s explicability was below 0.75, but the in winter models that comprehended distances between 4 to 20 km, the values were higher than in the corresponding summer models.

In order to assure that the models had no multicollinearity, it was ensured that all the VIF values were below 5 (please see

Supplementary Material). The significance of weights and path coefficients was accessed through bootstrapping. By approaching the traditional threshold for statistical significance,

p values larger than 0.05 were achieved for the weights (please see

Supplementary Material). On the other hand, it was verified that the path coefficients of “Land Use” were significant, characterized by

p values lower than 0.05 for long distances (

Figure 7).

The significance of land uses seemed to follow a sigmoid pattern. For the summer period, statistical significance (p < 0.05) was achieved for distances larger than 5 km, but for the winter period, significance was achieved for distances larger than 3 km. When it comes to “Contaminant emissions” none of the models achieved statistical significance. For the summer period, p values increased with distance. For winter, p values seemed not to change for distances below 2 km, increased to 0.124 at 3 km, and then dropped consistently until a distance of 15 km. For 20 and 56 km, there was a notable loss of statistical significance.