Machine learning models applied to TSS estimation in a reservoir using multispectral sensor onboard to RPA

Water quality monitoring is fundamental for the maintenance and conservation of water resources. However, the conventional monitoring method, which uses point sampling, does not adequately represent the spatial variability of its parameters, besides being costly. As a result, remote sensing techniques with the use of orbital sensors have been applied to map some water quality parameters, mainly the Optically Active Components (OACs), such as the Total Suspended Solids (TSS), Chlorophyll-a (Chl-a), and Colored Dissolved Organic Matter (CDOM). However, the monitoring of small reservoirs with orbital sensors always presents limitations, such as spectral and spatial problems or time resolution issues, combined with the fact that the sensors are subject to atmospheric disturbances and clouds. Seeking to overcome these issues, the present work aimed to develop a system for monitoring the concentration of Total Suspended Solids (TSS) in reservoirs, based on multispectral sensors onboard a Remotely Piloted Aircraft (RPA). A MicaSense RedEdge multispectral sensor was used, which has five bands in the blue, green, red-edge, and infrared (NIR) spectrum bands. Four campaigns were carried out in two water reservoirs within the Atlantic Forest biome at different seasons of the year. The data were submitted to five machine learning algorithms: the Random Forest (RF), Support Vector Machine Radial Sigma (SVM-RS), Enhanced Adaptive Regression Through Hinges (EARTH), Multiple Linear Regression (MLR), and Cubist algorithms. The results demonstrated that the SVM-RS model was the best adapted to the data patterns (r2 = 0.87), while the RF, EARTH, MLR, and Cubist models presented limitations that hinder their use in the estimation of TSS in reservoirs. This study demonstrated that the use of a multispectral sensor onboard an RPA has high accuracy in estimating TSS in small reservoirs and that machine learning algorithms can generalize the TSS data well. •Continuous monitoring of reservoir with low cost•The infrared spectral band was the determining covariate for the SVM model.•SVM algorithms were the most accurate algorithm for TSS estimates.•Machine learning algorithms estimate with SST accuracy in reservoirs.