Aplicação de técnicas de aprendizado de máquina e imagens multiespectrais para a estimativa da concentração de sólidos suspensos totais no Lago Igapó, na cidade de Londrina (PR)

Abstract

Urban rainwater drainage represents one of the most significant challenges in the environmental area around the world. One of the undesired and immediate effects of the surface runoff of rainwater is the transport of sediments and all sorts of pollutants into water bodies. Total Suspended Solids (TSS) represent one of the critical parameters of most significant interest in environmental monitoring and the implementation of Nature Based Solutions. Thus, this work proposes to evaluate an alternative to in situ measurements through the prediction of TSS concentration in an automated way, using Machine Learning (ML) techniques combined with high spatial resolution multispectral images captured by Unmanned Aerial Vehicle (UAV) from Lake Igapó, in the city of Londrina (PR). The measurement campaign considered 40 collection points, with some parameters analyzed in situ with the aid of a multiparametric probe (Hydrogenionic potential (pH), Electrical Conductivity (EC), Temperature, OD Percentage, OD Concentration, Salinity; Pressure, and Total Dissolved Solids (TDS)) and TSS concentrations measured in the laboratory. Simultaneously with collecting the water samples, multispectral images were captured in the Green, Red, Red Edge, and Near Infrared bands with the aid of a UAV. The ML techniques used were Random Forest (RF), Linear Regression (LR), Ridge Regression, Lasso Regression, and ElasticNet Regression, which were applied under different conditions (the original number of samples, reduced number by removing outliers, and increased number by interpolation). The ML models' predictive capacity was evaluated using the Mean Square Error (EQM) and the R square (R²) determination coefficient. The measured parameters were compared with the quality standards established in the Resolution of the National Environmental Council (CONAMA) 357/2005, for class 1 fresh water, thus, it was possible to affirm that the values found in this research, satisfy the requirements established in this resolution. The AM models that presented the best prediction capacity were: FA with R² of 62.30% and EQM of 5.48; RL with R² of 16.53% and EQM of 18.62; Ridge, Lasso and Elasticnet regression with R² of 16.37% and EQM of 18.69. All the best results were found when using interpolated SST values. This methodology is a promising tool with a great chance of replacing in situ measurements in the future. Expanding the number of sampling points, considering campaigns carried out under different sediment loads, can significantly improve the prediction capacity of the models tested.