Remote sensing technique for mangrove studies: Tropical mangrove species classification with convolutional neural network and tropical mangrove tree biomass modelling with terrestrial laser scanner
Abstract:
This research can be split into two topics. The first topic is the use of hi-resolution satellite images Quickbird in combination with convolutional neural network for classifying tropical mangroves at the species levels. The chosen network architecture is compressed fully convolution network (CFCN) that is based on the popular FCN-8s reference network. It was found that the compressed network performs as good as the reference network during the training session. This claim is supported by the six testing parameters. Additionally, the classification capability of the proposed network was tested against the popular maximum likelihood classifier (MLC). The proposed method gained 90.36% of total classification accuracy while the MLC method got only 66.02%. The outcome indicates that the proposed method can be used for classifying tropical mangrove at the operational level. The second topic proposes the non-destructive method, a terrestrial laser scanner (TLS), to collect the tree structure and generate the allometric tree model. The result was compared to the four reference models using the RMSE. The largest error (i.e., 40% and 35%) was found when comparing the presented model to the two generic allometric models. On the other hand, the outcomes of the species-specific models were closer to the outcome of this study (i.e., the RMSE errors are less than 20%). The discrepancy between the proposed TLS model and the generic mangrove model recommended that a species-specific model is demanded more accurate results. It is supposed that the methodology presented in this study may be utilized as standard procedures for producing the A. marina allometric model in other areas.