Modelagem hidrossedimentológica em unidades experimentais no semiárido paraibano utilizando redes neurais.

Abstract

The management of water resources in river basins in the Brazilian semi-arid region requires monitoring at different spatial and temporal scales, which aren’t always possible to carry out. Artificial intelligence has shown promise in environmental modeling, especially when hydrological data is used. The objective of this work was to compare Artificial Neural Networks (ANN) and simple and multiple linear regression to estimate estimate runoff and sediment yield in experimental units in the state of Paraíba. Data obtained from 147 rain events with consequent flooding were used, in micro-basins and experimental plots of the Experimental Basin of São João do Cariri-PB, whose áreas varies from 100m² to 16.300m² and slope of 3,4% to 7,1%, to model parametization. The input variables were precipitation, period without rain, area, slope and vegetation cover, while the output variables were sediment yield and runoff. The analyzes carried out using correlations, self-organizing maps (SOM), simple linear regression (RLS), multiple linear regression (RLM) and Multilayer Perceptron Neural Network (MLP), allowed comparing the performance of the models through mean squared error error (RMSE) and correlation coefficient (R²) values. For RLM the (RMSE) was 453.39 and (R²) 0.63 for Sediment Yield and 3.79 and 0.67 for runoff. For the MLP network, (RSME) ranged from 301.32 to 274.55 for Sediment Yield and 2.58 to 2.45 for runoff. The (R²) varied from 0.84 to 0.86 for Sediment Yield and 3.79 and 0.67 for runoff. Susceptibility to erosion was classified as very high in microbasins and moderate in the experimental plots. The results indicated that precipitation is a determining factor in the Generation of runoff and sediment yield. Pearson's correlation analysis identifying a positive and significant correlation between rain and runoff and sediment yield, ranging from moderate to strong. Simple linear regression was not able to provide satisfactory answers for predict runoff and sediment yield, and should be not used for hydrosedimentological regionalization. Multiple linear regression suprassed simple linear regression in accuracy, suggesting that the interaction between multiple variables. The SOM network analysis revealed complex patterns and interactions between variables, highlighting the importance of slope, area and vegetation cover on runoff generation and sediment yield. It was evidente that runoff and sediment yield are strongly influenced by the interaction between precipitation, vegetation cover and slope, indicating that models based on unsupervised neural networks are attractive alternatives for áreas analyzed. When there are complex cause and effect relationships between variables in hydrosedimentological processes, it is necessary to use non-linear and multivariate models, such as artificial neural networks. The MLP network has shown flexibility and the ability to capture complex non-linear relationships, with promising performance for test data, with simple configurations that balance accuracy and generalizability. The complexity of hydrosedimentological processes and the need for multidisciplinary approaches to their analysis are highlighted. Although multiple linear regression offers an interpretation of the influence several variables, MLP networks present superior results for modeling the complexity inherent in these systems, even with underestimation or overestimation sediment yield and runoff.