High resolution multispectral image and machine learning algorithms for saltmarsh classification and biomass modelling

Saltmarsh is one of the important vegetation of wetlands. Due to a range of pressures, it has been declared as an Endangered Ecological Community (EEC) in Australia. Therefore, monitoring and mapping of the distribution of saltmarshes species are important to wetland management, conservation and distribution. This rigorous task requires intensive fieldwork and collection of ancillary data that are time-consuming. Remote Sensing offers a practical and economic means of plant sciences classification and biomass modelling. However, selecting suitable remote sensing systems and their data are important for mapping saltmarshes. Hyperspectral remote sensing can be used to monitor this endangered community. However, there are some crucial limitations of hyperspectral data that have been found in the current study area and discussed in the introduction. To overcome these limitations, Worldview-2 with its higher spatial (1.84 m) resolution is seen as a trade-off between the advantages of multispectral resolution satellite data and hyperspectral data. In this thesis, imagery acquired by three different sensors were used to compare the performance of machine learning classification methods and biomass regression models. Maximum Likelihood Classifier (MLC) and two advanced algorithms, Random Forest (RF) and Support Vector Machine (SVM) were used for classification. These two algorithms were also tested to develop a biomass model for Sporobolus virginicus species using multispectral Worldview-2 data. Reflectance and NDVI based vegetation indices derived from 8 bands of Worldview-2 multispectral data were used for four experiments to develop the biomass model. Proportional reduction of sample size (100 % to 33%) was focused to test the potentiality of both algorithms. RF showed significant changes in overall accuracy when the sample size was reduced from 100 % to 33%. Conversely, there were no significant changes in the accuracy for SVM when the sample size equally dropped from 100 % to 33%. When biomass model for RF (R2 =0.72, RMSE = 0.166 kg/m2) and SVR (R2 = 0.66, RMSE = 0.200 kg/m2) were compared, there was a significant (p =< 0.0001) difference was observed. Further research is crucial to explore the sensitivity of this method with the spatial autocorrelation of the training samples for random forest application in saltmarsh monitoring.