Uso de redes neurais artificiais e AHP na redução da subjetividade na seleção de áreas com potencial para coleta de água da chuva

Abstract

Extreme events related to rainfall, as floods and drought, although natural, can have undesirable effects in both urban and rural environments. Therefore, rainwater harvesting techniques, known as Rainwater Harvesting (RWH), are seen as a cost-effective alternative to controlling runoff. However, in order to achieve greater efficiency, it is necessary to make a careful evaluation of the watershed in question to identify the areas with the greatest potential for implementation of these structures. As this analysis depends on several factors, such as slope, land use and coverage, drainage network density, soil hydrologic characteristics and precipitation, multicriteria decision analysis with support to Geographic Information Systems (GIS) has great potential for this end. Within this methodology Hierarchical Process Analysis (AHP), is a technique widely used to weigh the criteria. Nonetheless, this technique has been questioned due to its inability to detect and minimize the interference caused by the subjectivity of the specialists in the weighting of the criteria. However, the present work proposes the use of Artificial Neural Networks (ANNs) as a way of minimizing this subjectivity in the weighting of criteria for the selection of areas with potential for RWH. For this, these techniques are applied to the Uberabinha River Watershed, located in the mesoregion of the Triângulo Mineiro and Alto Paranaíba, in the state of Minas Gerais, Brazil. It was possible to sse that the Consistency Ratio (CR) was not enough to evaluate the AHP subjectivity. It was also noted that the neural nets revealed to be effective to combine the weight assigned by the specialists and the hidden layers numbers and settings didn’t interfere in the weight combination. Regarding the activation function, it was noted that the logistic function generated distinct output than the other function. At least, it was noted that the average employed as training target was the main factor to interfere in the nets results.