Performance of landscape composition metrics for predicting water quality in headwater catchments

Abstract

Land use is a predominant threat to the ecological integrity of streams and rivers. Understanding land use-water quality interactions is essential for the development and prioritization of management strategies and, thus, the improvement of water quality. Weighting schemes for land use have recently been employed as methods to advance the predictive power of empirical models, however, their performance has seldom been explored for various water quality parameters. In this work, multiple landscape composition metrics were applied within headwater catchments of Central Europe to investigate how weighting land use with certain combinations of spatial and topographic variables, while implementing alternate distance measures and functions, can influence predictions of water quality. The predictive ability of metrics was evaluated for eleven water quality parameters using linear regression. Results indicate that stream proximity, measured with Euclidean distance, in combination with slope or log-transformed flow accumulation were dominant factors affecting the concentrations of pH, total phosphorus, nitrite and orthophosphate phosphorus, whereas the unweighted land use composition was the most effective predictor of calcium, electrical conductivity, nitrates and total suspended solids. Therefore, both metrics are recommended when examining land use-water quality relationships in small, submontane catchments and should be applied according to individual water quality parameter.

Land Cover Effects on Selected Nutrient Compounds in Small Lowland Agricultural Catchments

“…The performance of land cover metrics in water quality prediction, calculated for different spatial scales with even additional distance of flow accumulation scaling, was previously broadly discussed [42,45,92]. However, the presented results are not clear and unequivocal.…”

“…Some of the previous studies also indicated that even with the same land use percentage, landscape configuration, measured with the patch density, edge density, and mean shape index, plays an important role in organic matter and nutrient runoff from catchments [39]. However, this can be more important in larger catchments, where their area is suggested to have a significant effect on metrics performance [42].…”

“…As indicated, land cover metrics during summer and autumn seasons could be useless. Nevertheless, this fact points to the need to search for new predictors, which could explain nutrients variability during growing season, and such additional variables could include macrophytes density, as well as soil properties metrics [42]. Moreover, from scaling (weighting) methods presented in the literature [47,98,99], the use of buffer zones and/or Euclidean distance scaling seems to be the optimal solution for modelling purposes in lowland landscapes.…”

“…Along with the development of GIS software, the increasing availability of high-resolution spatial data made it possible to precisely quantify various types of landscape properties, such as slopes, land use, and land cover types, using different types of metrics [16,29,38]. In such a way, a number of studies linked the concentration of nutrient compounds with landscape properties and these relationships were studied at different spatial scales, from entire catchment areas [39] to buffer zones along the watercourses [40,41], sometimes with additional distance or flow accumulation scaling [42]. Most of the work in this area, however, concerned catchments of over a hundred square kilometers, including upland or highland relief and steeper slopes of the terrain, where high hydrological connectivity and intensive erosion result in increased ion migration [43][44][45].…”

A Comparison of Linear and Non-Linear Machine Learning Techniques (PCA and SOM) for Characterizing Urban Nutrient Runoff

“…In addition to the intrinsic random nature of urban runoff quality, most of the works mentioned above have also demonstrated that many multifaceted interactions among these factors have occurred in multi-dimensions. In such conditions, conventional univariate (e.g., ANOVA) and multivariate (e.g., linear regression) statistical techniques have been widely adopted to investigate the correlation among water quality characteristics and influential factors [3,14]. However, the outcomes were often affected by several sources of statistical bias, unless all the analysis requirements (e.g., model selection criteria, parametric assumptions, and interaction terms) were rigidly tackled or fixed [9,15].…”