Estimating chlorophyll-a concentrations in tropical reservoirs from band-ratio machine learning models

Abstract

Remote sensing technology has been a popular tool with high potential to measure chlorophyll-a concentration in inland waters. This study proposed the band ratio algorithm to estimate the chlorophyll-a concentration in inland water using a machine learning approach. The blue-green, NIR-red band ratios, and the three-band models were sensitive to chlorophyll-a concentration in various water bodies. Various band ratio variables were used to obtain the optimum model for chlorophyll-a estimation in multiple tropical reservoirs. The spectral reflectance in the tropical reservoirs was derived from Sentinel-2 Level 2A. The corresponded reflectance of band ratios and in-situ chlorophyll-a data (N = 109) were applied to identify the correlation and thus justified its potential as appropriate inputs in the machine learning model. Result shows that the NIR-red band ratios are strongly correlated with chlorophyll-a concentration. These spectral band ratios are appropriate inputs for the machine learning model. The effective multiple band ratios are determined from correlation (r > 0.5 between ratios and concentrations) and collinearity test (variance inflation factor, VIF<3). Result confirms that the multiple band ratio algorithm represents a superior performance in increasing the accuracy of each machine learning model. Overall, the random forest provides a robust and reliable estimation using the three-band NIR-red, and blue-green band ratios (RMSE = 2.963 μg/L, r2 = 0.955).