A deep learning-based approach for mapping tall shrubs in Arctic tundra
Darko Radakovic, Montclair State University, radakovicd1@montclair.edu (Presenter)
Mark James Chopping, Montclair State University, choppingm@mail.montclair.edu
The project utilized commercial high spatial resolution satellite imagery to assess changes and improve the mapping of shrub succession in numerous locations across the Arctic tundra regions of Northern Alaska. QuickBird imagery (~ 0.6 m) from around 2005 served as the “early period” and WorldView-2 (~0.4 m) and WorldView-3 (~0.3 m) imagery from 2015 to 2022, served as “late period” image for diverse cloud-free, summer tundra landscapes. We used 242 2-by-2 km study areas with machine learning (ML) techniques. Each area was subdivided in 50- and 100 m from pansharpened QuickBird and WorldView imagery to be processed for annotations of shrubs by the Yolov8 model for object detection of individual shrubs and by the Detectron2 model for segmentation of dense tall shrub areas and polygonal ground in the satellite images.
Mapping shrub patches with segmentation in QuickBird and Worldview images produces lower accuracies compared to object detection. Object detection on the other hand had a low recall, due to the few detections when compared with the Toolik Lake Vegetation Community Map. However, most of the predicted labels were correct when compared to the training labels. F1-scores were low for all detections which could indicate that the detected classes are imbalanced. Imprecision can be reduced by utilizing Convolution Neural Network (CNN) filters to reduce false positives. Our current research is focused on generating time series data for our study areas to incorporate additional imagery and to quantify uncertainty using series of all available imagery across the period.
The use of large datasets, obtained from high-resolution remote sensing can reveal patterns not observable with previous methodologies. Combined with ML, this is a novel means of approaching a complex, thawing system experiencing important hydrologic and carbon cycle responses to climate change; and a better understanding of this behavior provides insights into the likely future environmental changes as warming accelerates.
Poster: Poster_Radakovic_3-22_89_35.pdf
Associated Project(s):
Poster Location ID: 3-22
Presentation Type: Poster
Session: Poster Session 3
Session Date: Thu (May 11) 3:00-5:00 PM
CCE Program: TE