MACHINE LEARNING AND REMOTE SENSING APPROACH FOR MONITORING WATER QUALITY IN CURUAI LAKE (PA)

Código

XXVI-SBRH1068

Autores

Lohan Barbosa Baía, Júlio César Pimenta dos Santos, Bruno Rech, Jahdy Moreno-Oliveira, Daniel Andrade Maciel, Jackison Mateus Lopes Barros, Hanna Luisa Lima Alves, CAROLINA COSTA RAMOS, Marcio Sousa Silva, Evlyn Márcia Leão de Moraes Novo, Cláudio Clemente Faria Barbosa, Pedro Walfir Martins e Souza Filho

Tema

STE110 - Sensoriamento remoto da água: avanços técnicos-científicos e aplicações na nova era de disponibilidade de dados

Resumo

Understanding the spatiotemporal patterns of optically active constituents in Amazonian floodplains remains a significant scientific challenge, particularly due to the region?s environmental complexity and limited accessibility. In this study, we developed and validated a Random Forest (RF) model to estimate total suspended solids (TSS) in Curuai Lake, Pará, Brazil, using in situ observations and satellite-based spectral reflectance data. A dataset of 483 field samples collected across distinct Amazonian aquatic systems was integrated with atmospherically corrected OLI/Landsat-8 imagery. Spectral reflectance was simulated from radiometric field measurements, and 70% of the samples were used to calibrate the RF model based on selected bands and spectral indices. The model showed strong performance in training (R = 0.94, Bias = 1.17 mg·L?¹), but moderate correlation with in situ validation (R = 0.40), reflecting limitations associated with spatial heterogeneity and reduced match-up samples. The TSS concentration map revealed values ranging from 8 to 100 mg.L?¹, with higher concentrations observed in the western portion of the lake, likely influenced by bathymetry and wind-induced resuspension. Despite increased errors inheterogeneous datasets, the RF model demonstrated coherent spatial trends, aligning with hydrodynamic processes and offering valuable insights for regional-scale monitoring. This study highlights the relevance of combining remote sensing and machine learning to assess water quality in floodplain environments, supporting early detection of anthropogenic impacts and guiding resource management in the Amazon basin.