Estimativa de larguras de vegetação para zonas ripárias através de redes neurais : o caso da remoção do nitrogênio

Abstract

Water is an indispensable resource for all populations of the biosphere, as well as, a key abiotic components of ecosystems. There has been increase in degradation in its quality and there is miss of it due to several factors, such as large anthropogenic pollution and inadequate planning and management of Watersheds. In this context, we can notice that one of the attributes of green ecosystems is precisely the protection of rivers. Thus, the determination of the desing vegetation that makes up the riparian buffer zone, which allow the full use of its functions, is important for the conservation of this resource. Even though, the conservation of areas of Per-manent Preservation only occurs by the law, it is necessary that these standards have scientific basis. This study is divided into two parts (a review and a practical methodology): the first part, called Article I, consists of a review of estimates of the efficient widths for different functions of riparian vegetation, as well as it presents a brief overview of the used methodologies. We reviewed 37 scientific articles that focused on the influence of the width of the riparian vege-tation in the performance of functions and / or environmental services. As a result, we could categorize the functionality of these areas into six groups: a) sediment filtering, b) nutrients filtering, c) abiotic factors, d) diffuse pollution, e) species conservation, and f) runoff. Then, we described the minimum and maximum widths estimated as optimal for each of the functional groups. A large variability in riparian widths has been found. The nutrient removal 3.8 to 280m presented the greatest variation in the width. The other categories ranged within these values. In the second part of this study, also called Article II, we developed a methodology to estimate the width of the riparian vegetation, which was based on an Artificial Neural Network Ensemble. The following parameters were used to input data on the networks: hydraulic conductivity of soil, vegetation index, nitrate concentration and filtering efficiency of soil nitrate load. Then, the data were used to train, validate, test and select the best performing networks to form the Ar-tificial Neural Network Ensemble. With the study of the Digital Elevation Model of the Ligeiro Watershed River, which is divided into 165 smaller sub-basins, the Artificial Neural Network Ensemble calculated the satisfactory width of riparian buffer vegetation that is needed in each one of the sub-basins to remove 90% of the existing soil nitrate load. It proved to overcome the performance of individual neural networks greatly reducing the variance of output errors on the individual networks.