Turbidity Permanence: Sensitivity-Driven Input Reduction in ANN Models for Stormwater TSS and DOC

Código

V-END_II-EBHE0062

Autores

Sabrina Bittencourt Medeiros Drummond, MARIA CRISTINA BORBA BRAGA, Stephan Fuchs, Sergio Michelotto Braga

Tema

19 - Inteligência artificial e big data na gestão de desastres

Resumo

Sensitivity analysis is increasingly applied to artificial neural network (ANN) models of urban diffuse pollution, yet comparative studies indicate no single technique reliably identifies dominant predictors across hydrological settings. This study evaluated five complementary sensitivity methods: Sobol, Morris, SHAP, one-at-a-time (OAT), and Garson, applied to ANNs trained to predict total suspended solids (TSS) and dissolved organic carbon (DOC) concentrations during stormwater events. Model performance was assessed using a weighted composite score that combines seven metrics representing efficiency, error magnitude, and bias. Across all five methods and for both water quality parameters, turbidity was consistently retained as the sole significant input, despite the well-documented nonlinearity and site specificity of the TSS-turbidity relationship. This convergence is interpreted as evidence that turbidity functions less as an isolated proxy variable and more as an aggregate optical signal reflecting the simultaneous mobilization of mineral particles, organic debris, and colloidal matter during storm events, consistent with reported co-variation between turbidity and DOC of up to 300-fold and 3.5-fold, respectively. The results support high-frequency turbidity monitoring as a parsimonious, physically grounded basis for real-time estimation of particulate and dissolved organic pollutant loads, while underscoring the need for site-specific calibration given the absence of a universal TSS-turbidity-DOC transfer function. These findings reinforce, through an independent model-based approach, an emerging consensus on the informational value of continuous turbidity monitoring.