Landscape factors related to performance of in-situ sensors for prediction of total phosphorus : a statistical analysis of data from 194 streams in Sweden

Abstract: Accurate monitoring data of phosphorus concentrations in the water will benefit the evaluation of measures taken to reduce the problems of eutrophication. Today, the quantification of phosphorus loads from diffuse sources involves large uncertainties. Traditional grab sampling could underestimate the loads exported from diffuse sources due to the flow dependence of particulate phosphorus (PP) concentrations. Technical innovations allow continuous monitoring of certain parameters in the water and by that generate data of high temporal resolution. These parameters have the potential to act as surrogate measurements for total phosphorus (TP). One parameter that is often argued to have strong correlations with total phosphorus is turbidity, which might reflect the PP fraction in the water. However, poorer correlations have also been reported. Possibly due to land-use or soil type within the catchments. By finding landscape factors related to good/poor performance of in-situ sensors, in the prediction of TP, will benefit the strategical planning of monitoring programs. In this study, data from available sensor parameters were used to create regression models for TP at 194 monitoring stations in Sweden. Turbidity, water temperature and total organic carbon were most frequently included in the significant regression models. The correlation coefficient from the most significant TP regression model at each station was compiled into a response dataset for multivariate analysis. The correlation coefficients were then predicted with landscape data from the catchments of the monitoring stations with multivariate analysis. A separate dataset of landscape data within 100 m buffer zone was also used for comparison. The results indicated weak associations between the landscape factors and the performance of the TP prediction models. The landscape factor of most significance relation to high correlation coefficients was forest on mire, whereas water and open wetlands along the buffer zone were related to lower correlation coefficient values. No preferences were found between significant catchment factors and sensor parameters included in the regression model at the stations. Data of the whole catchment explained larger variations in the TP regression model’s strength than the buffer zone data, possibly due to low data resolution in the buffer zone.