A Deep Learning-Based Approach for Mapping Tall Shrubs in Arctic Tundra from High-Resolution Satellite Data
Darko Radakovic, Montclair State University, radakovicd1@montclair.edu (Presenter)
Mark James Chopping, Montclair State University, choppingm@mail.montclair.edu
This study evaluated deep-learning models Convolutional Neural Network (CNN), ResNet50, VGG19, U-Net, and Vision Transformer (ViT) for detecting changes in tall shrub cover across Alaska’s North Slope using very-high-resolution QuickBird (QB) and WorldView (WV) satellite imagery spanning 2002 to 2020. Models were trained on 4,100 image tiles (200-by-200 m) using various spectral inputs including Pansharpened 4-band multispectral satellite imagery and panchromatic (P1BS) at top-of-atmosphere (TOA) radiance.
Model validation against the Toolik Lake Vegetation Community Map used as field-reference indicated robust accuracy across all architectures, with ResNet50 and VGG19 achieving highest accuracies (~86%) and balanced F1 scores (57–59%). CNN and U-Net models also performed moderate (81–85% accuracy), with CNN exhibiting effective hierarchical feature extraction. Supplementary inputs from stacked imagery, such as the Normalized Difference Vegetation Index (NDVI) and Principal Component Analysis (PCA) scenes, did not substantially improve segmentation accuracy, suggesting that linear transformed TOA radiance imagery products do not further improve model performance.
Binary segmentation predictions revealed subtle variations among models, particularly in transitional vegetation zones. Continuous tall shrub cover predictions showed that model performance was negatively correlated with shrub density (r = -0.85 for accuracy; r = -0.54 for F1), implying greater classification challenges in denser shrub areas. Additionally, one-way ANOVA tests indicated significant, but modest sensor-driven differences (QB02, WV02, WV03) in shrub cover estimates.
Temporal trend analyses demonstrated a statistically significant increase in tall shrub cover over time for all models, with the ResNet50 models showing consistently highest positive slopes (?0.86% per year, p<0.001) while the CNN, VGG19, U-Net and ViT models showed more modest growth rates (0.08%, 0.29%, 0.25% and 0.04% per year respectively, p<0.001). Binary segmentation thresholding methods influenced estimated shrub expansion rates, highlighting that dynamic percentile-based thresholds (p70–p80%) produced conservative and reliable (tall) shrub delineations compared to global thresholds.
Calibration against field-reference indicated strong predictive capability for ResNet50 and VGG19 models (R²=0.78–0.81), with relatively low errors and biases. Comparisons with NASA’s G-LiHT-derived Canopy Height Model (CHM) further validated model robustness, revealing moderate correlations (R²=0.18–0.27) between model-predicted shrub fractions and CHM-derived canopy heights.
Cross-sensor comparisons with Planet imagery, utilizing the same architectures (VGG19, ResNet50), yielded broadly consistent spatial shrub patterns, though Planet-derived predictions exhibited higher interannual variability due to lower resolution and calibration differences. Despite these sensor-specific variations, both Planet and QB/WV datasets demonstrated consistent temporal trends in shrub expansion.
This study highlights the efficacy and reliability of deep-learning methodologies for mapping and monitoring tall shrub cover across heterogeneous Arctic landscapes. Our findings underscore the importance of sensor calibration, appropriate spectral inputs, and robust segmentation strategies in accurately detecting shrub dynamics. Continuous expansion of tall shrub cover aligns with documented Arctic warming trends, emphasizing the potential for deep-learning-based approaches to advance our understanding of permafrost dynamics and vegetation-climate feedbacks in rapidly changing northern ecosystems.
Poster: Poster_Radakovic_38_52_45.pdf
Associated Project(s):
Poster Location ID: 38
Presentation Type: Poster
Session: Vegetation Dynamics and Distribution
Session Date: Monday (5/12) 4:35-5:30 PM