Neural Network Modeling of Lake Surface Chlorophyll and Sediment Content from Landsat Tm Imagery

Concentrations of chlorophyll and suspended sediment are two important optically active parameters of inland water quality. In the open ocean, these two parameters can be effectively quantified by empirical algorithms relating remote sensor radiances to surface concentrations. In inland waters, however, the task becomes difficult due to the presence of suspended sediment and dissolved organic matters in high concentrations, often varying independently of each other and overwhelming the signature of chlorophyll. Thus, the transfer function becomes non-linear in nature. Moreover, broad band sensors have to be used in inland waters as the present aquatic satellite sensors lack adequate spatial resolution for monitoring in these waters. In the process, conventional algorithms fail to estimate the water quality parameters effectively. Neural networks has been regarded as a relatively simpler tool to implement with proven success in modeling various nonlinear geophysical transfer functions. In this study, back-propagation neural network is used to model the transfer function between chlorophyll concentration and suspended solid, and sensor-received radiances at the first four bands of LandsatTM. Study area is lake Kasumigaura of Japan, a shallow eutrophic lake with heavy sedimentation. Neural network with only one hidden layer could model both the water quality parameters better than conventional regression techniques from LandsatTM imagery. Root Mean Square Errors(RMSE) in estimating chlorophyll-a were 1.53μg/l(R: 0.93) and 4.39μg/l(R: 0.31) for neural networks and regression respectively. In estimating suspended sediments, RMSE for regression was 1.47mg/l(R:0.92) while for neural network the same was 2.14mg/l(R:0.85). Neural networkderived map of chlorophyll-a shows that, the lake is eutrophic even in the low productivity season.