Ocean Colour Remote Sensing Of Case 2 Waters Using An Optimised Neural Network
Loading...
Date
2016-03
Authors
Anwar, Saumi Syahreza
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
This study focused on the development of the new algorithm for retrieving ocean colour products of Case 2 water types using the neural network (NN) model and multiple types of remotely sensed data as inputs. The NN model architecture and training parameters were optimised, with inputs being selected based correlation analysis (CA) and principal component analysis (PCA). In Kelantan coastal waters, the use of in situ reflectance spectra and simulated satellite data for estimation of two water clarity parameters namely turbidity (TURB) and Secchi disk depth (SDD) have been studied. The simulated Landsat TM and AVNIR-2 data were tested against in situ reflectance spectra measurements using ASD Spectroradiometer. The results show that the simulated Landsat TM and AVNIR-2 data enables the interpretation of TURB and SDD. In Penang coastal area, the use of single and multitemporal remote sensing data for estimation of Cs and Cchl has been studied. Multidate in-situ water sample measurements concurrent with Landsat TM and AVNIR-2 satellite data were obtained in selected locations from February 1999 to March 2011. The irradiance reflectances of Landsat TM and AVNIR-2 derived by ATCOR-2/3 software from the water sampling sites were extracted and examined with numerous algorithms. Although significant correlation was detected between reflectance values of Cs and Cchl when using the multivariate, and optical model, however the application of NN appears to produce superior performance in modelling transfer function in this study. The results show that the estimation accuracy for characteristics of two ocean colour products using neural network is much better than those empirical and optical algorithms. The results also indicated that NN based on CA and PCA can improve the estimation of these characteristics. Using five independent variables (TM1, TM2, TM4, TM5 and TM7) as inputs, PCA-NN model from multitemporal Landsat TM data has shown slightly improved the retrieval performance of Cs and Cchl (R2 = 0.93 and R2 = 0.92 for RMSE = 22.38 mg/L and RMSE = 1.29 g/L) in comparison with the NN and CA-NN model. On contrary, the CA-NN model with four independent variables (AV1, AV3, AV4 and AV3/AV2) of multitemporal AVNIR-2 data has shown better prediction performance for Cs and Cchl with (R2 = 0.93 and R2 = 0.92 for RMSE = 22.38 mg/L and RMSE = 1.29 g/L) in comparison with the NN and PCA-NN model. Although the PCA-NN using the multitemporal AVNIR-2 data has a slightly reduce generalization power, however the optimization of the PCA-NN model has demonstrated better prediction performance in comparison with the NN model. Therefore, it may be possible to develop ocean colour algorithms in which CA and PCA are used as methods for selecting input variables to a NN for remote sensing observation.
Description
Keywords
Ocean colour