Large-scale prediction of tropical stream water quality using Rough Sets Theory

Abstract Assessing water-quality in streams is traditionally measured at the local scale and in general it is spatially restricted. To scale-up water-condition assessment, there is a need to use new tools that enable prediction of large-scale changes in water-quality by expanding the analysis to landscape-levels and bioclimatic predictors. In addition, the traditional use of linear models in biomonitoring can be inappropriate in detecting complex relationships, such as changing patterns of aquatic community structure and complex environmental gradients. In this context, Artificial Intelligence (AI) techniques such as Rough Sets Theory (RST) can be particularly useful for dealing with vague, imprecise, inconsistent and uncertain knowledge involving biotic and abiotic data to enable the classification and prediction of changes in stream water. Here, we applied RST to estimate the water-quality in streams of the Brazilian Atlantic Forest by analyzing connections between landscape and climate data with the inclusion of up to 15 families of aquatic insect groups from the orders Ephemeroptera, Plecoptera and Trichoptera (usually known as EPT taxa). First, we developed different decision sets which were the arrangements of the response variable (EPT index) and the predictor classifications. Then, we applied the best decision sets to monitor the condition of stream water in the Atlantic Forest on a large-scale. Our results showed the best decision rules were 61% accurate. Depending on the initial stream condition, this approach on a large-scale led to variable accuracy. By combining the development of different decision sets, the application of the best one on a large-scale, and the use of open-access data (landscape and climate predictors), our study approach demonstrated the potential applicability to evaluate streams in an objective with low-cost manner. This method can complement the environmental assessment of streams based only on local variables. Our framework also creates new perspectives in the analysis of water-quality to generate scenarios of changes in streams based on landscape measurements to optimize monitoring networks.