Uso do sensoriamento remoto para avaliação da dinâmica espaço-temporal de parâmetros de qualidade da água de um reservatório urbano

Abstract

The assessment of patterns and trends in the water quality of reservoirs and lakes is routinely hindered by the lack of regular historical monitoring data. Conventional monitoring techniques are limited by spatial and temporal representativeness, accessibility challenges, and financial constraints. In recent years, remote sensing has expanded the spatial and temporal representativeness of water quality monitoring through the reconstruction of historical series and monitoring of the surface extent of water bodies, becoming an add-on to in situ monitoring. This study, using satellite images obtained from their sensors from Landsat-8 (OLI) and Sentinel-2 (MSI), corrected by an atmospheric correction algorithm (SIAC), aimed to estimate the concentrations of total suspended matter, chlorophyll-a, and turbidity, as well as Secchi disk depth measurements for an urban reservoir, known as Lagoa da Pampulha, located in Belo Horizonte-MG, Brazil. Satellite-sensor data was combined with historical water quality monitoring data collected by government agencies between 2013 and 2022. The data was then modeled through the adjustment of existing literature models and models proposed by this study. To increase modeling robustness, univariate and multivariate linear and nonlinear regression methods were investigated, with the implementation of cross-validation method combined with Monte Carlo simulation (10,000 repetitions). Model performance was evaluated, in calibration, using the coefficient of determination (R2), and, in validation, by the root mean square error (RMSE) and nRMSE, which is represent RMSE normalized by the observed sample mean. The chosen models from Landsat-8 were: TSM (n=45) using Random Forest model with R² of 0.89, RMSE of 7.8 (mg/L) and nRMSE of 28%; Chla (n=49) using polynomial regression model with R² of 0.77, RMSE of 91.8 (μg/L), and nRMSE of 95%; SDD (n=23) using Random Forest model with R² of 0.62, RMSE of 0.1 (m) and NRMSE of 30%; and Turbidity (n=16) using linear regression model with R² of 0.8, RMSE of 26.2 (NTU) and NRMSE of 52%. The chosen models from Sentinel-2 were: TSM (n=53) using Random Forest model with R² of 0.72 RMSE of 8.5 (mg/L), and nRMSE of 42%; Chla (n=54) using linear regression model with R² of 0.38, RMSE of 51.6 (μg/L), and nRMSE of 92%; SDD (n=33) using Random Forest with R² of 0.89, RMSE of 0.1 (m) and nRMSE of 15%; and Turbidity (n=9) using multiple linear regression with R² of 0.97, RMSE of 3.4 (NTU) and nRMSE of 13%. The selected models were applied to cloud-free images from Landsat-8 and Sentinel-2, resulting in a unique multidimensional array of water quality model parameters with dimensions (time=155, y=71, and x=95). Through cluster analysis,five homogeneous regions were identified in the reservoir, and there was a significant difference (p-value<5%) in water quality when comparing the rainy and dry seasons, periods related to the chemical treatment of the Lagoon, and wetter or drier years, as indicated by the standardized precipitation index. An improving trend in water quality was observed for the modeled period (2013-2022) with a decrease in TSM, Chla, and turbidity parameters, and an increase in SDD, inferred through the slope coefficient derived from Theil-Sen statistical test combined with seasonal Mann-Kendall test. The expansion of data quantity in the temporal and spatial scale, along with the analyses supported by quantitative information, provide guidelines for future studies and entities responsible for reservoir management.