Data Extraction from River Area using Very High-Resolution Multispectral UAS imagery and Classification by Artificial Intelligence

Unmanned Aerial System (UAS) imagery has enabled very high-resolution multispectral image acquisition. Detection of wet areas and classification of land cover based on these images using the Machine Learning (ML) algorithm named Random Forest (RF) is our main purpose in this paper. Very high-resolution UAS images have been used as inputs for a machine learner in different scenarios. Starting from Random Forest Classifier, 3 different datasets consisting of RGB only, Multispectral only, RGB plus Multispectral for classification of the area have been used in three different test areas in two time epochs. Therefore, another objective of this study is to investigate performance of datasets (number of included bands and related wavelength) in the classification and wet area detection, to probe whether RGB (visible) light provide better results for our goal or multispectral data (Red-Edge and NIR) outperform in this analysis, or whether combining visible and Multispectral data is a superior alternative or not. In our case study for the sake of simplicity of comparison between implemented methods of classification and strength of datasets, only three classes have been analyzed including Vegetation, Water and Ground. Furthermore, in another time epoch in summertime, capability of spectral, vegetation indices, elevation, and texture features in the classification of land cover and detection of the wet riparian area in the case study are assessed. There are many existing methods for the classification of land cover based on UAS images, but very high-resolution centimeter-level data are of main importance in this analysis. Outstanding results have been produced in both epochs considering three extremely accurate performance analysers. Additionally, in this research, the most decisive and effective features have been discovered to compromise accuracy and the number of effectual features. In the last part of the project, a literature review on the available deep learning solutions based on Convolutional Neural Networks have been performed and possibility of training the network based on our dataset have been analysed.