Development of the best retrieval models of non-optically active parameters for an artificial shallow lake by random forest algorithm

This research was carried out to acquire the optimized retrieval algorithm by the random forest method and based on remote sensing data to monitor total nitrogen and phosphorus parameters as key drivers for eutrophication in water bodies. For this purpose, the water quality monitoring data of Chitgar Lake in Tehran were used, which is an artificial shallow lake with recreational and urban scenery usage. The Landsat 8 OLI/TIRS and Sentinel 2 MSI satellite images were extracted after matching the date of field observation data and satellite images from 2014 to 2021. Data were divided into calibration and validation datasets. After performing pre-processing processes on satellite images, important bands were recognized using the random forest method. After that, appropriate band compositions and algorithms were selected and regression models were fitted and validated. The optimum model based on Sentinel-2 data was able to estimate total nitrogen concentration with Adj.R2 = 0.82, RMSE = 0.24 mg. L−1 and NRMSE = 15% as well as total phosphorus concentration with Adj.R2 = 0.6, RMSE = 0.012 mg. L−1 and NRMSE = 9% accompanied by the power of 80% in the study area. Consequently, the optimal retrieving algorithm was chosen in the synergy of satellite data and the random forest. Moreover, the predictive model has a promising capability to estimate the concentration of the objective parameters in Chitgar Lake with acceptable accuracy.