2024 CSDMS meeting-069
From CSDMS
(if you haven't already)
Log in (or create account for non-CSDMS members)
Forgot username? Search or email:CSDMSweb@colorado.edu
Browse abstracts
A Comprehensive Machine Learning and GIS Approach for Dune Vegetation Mapping Using UAV-Based Imagery at an Undeveloped Dune System in Long Branch, New Jersey
Aman Kumar Phogat,
(He/Him),Montclair State University Montclair New Jersey, United States. Phogata1@montclair.edu
Shane Daiek, Environmental Science and Management PhD Program, Montclair State University New Jersey, United States. daieks1@mail.montclair.edu
Shane Nichols-O’Neill, Earth and Environmental Studies, Montclair State University New Jersey, United States. nicholsones1@mail.montclair.edu
Jorge Lorenzo Trueba, Environmental Science and Management PhD Program, Montclair State University New Jersey, United States. lorenzotruej@mail.montclair.edu
Quantifying the spatial distribution of vegetation on coastal dunes is critical for understanding their morphological evolution and ecological functioning. However, effective large-scale mapping of dune vegetation presents a challenge for coastal managers at the resolution suitable for detailed analysis of these berm-dune systems. This study presents an integrated machine learning and geographic information system (GIS) methodology to accurately classify vegetation, sand, and shadows from high-resolution unmanned aerial vehicle (UAV) imagery of an undeveloped dune system in Long Branch, New Jersey. Using a Zenmuse X5S 3-band multispectral sensor and a Zenmuse X5 RGB sensor, we collected data from a single survey in October, 2023. We then combined the band data from this flight in order to develop additional model inputs of normalized difference vegetation index (NDVI) and green NDVI (gNDVI). Thirteen model runs were performed and compared using 3 to 8 inputs including spectral bands of Red, Green, Blue, and near-infrared (NIR). The spectral indices of NDVI and gNDVI, were developed using Red, Green, and NIR band data. In our research, the random forest machine learning algorithm was employed, with optimal hyperparameters determined through empirical testing. Classifiers with 350 decision trees were found to yield highest accuracy on the training data across different spectral indices combinations. The trained random forest models were then applied to map vegetation distribution on an unseen "training grid" area. Results show that model runs containing the NDVI and gNDVI yielded higher accuracy in classification than model runs only containing a combination of R, G, B, and NIR bands. Incorporating additional Red and Green bands from our multispectral sensor produced less accurate model classification as these additional inputs lead to overfitting and increased misclassification of vegetation and shadows. Incorporating NDVI and gNDVI indices significantly improved classification performance compared to model runs that only used spectral band information. The classified vegetation, sand, and shadow maps were integrated into ArcPro GIS software for visualization and validation against other established methods of manual class identification and map digitization. Rigorous accuracy assessments confirmed the robustness of the machine learning approach that can be used to quickly and accurately identify vegetation distribution in a dune complex.
