Sandeep Mukherjee

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-81

May 10, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-81.pdf

Classifying trees from GPS-registered aerial imagery poses several challenges: data are lowsignal and noisy and there is a long-tailed, fine-grained class distribution. In this setting, supervised classification heads trained on DINO features under-perform simple fully-supervised classifiers. On the other hand, noise-resistant feature extraction excels but can be overly class-indicative. We find that the Graph Attention Transformer (GAT), is well-specified to model geo-spatial correlations present in aerial imagery, but tail class sensitivity suffers when trained on overly class-indicative features (section 6.2). We present Softening Head Imbalances with Effective Learning and Debiasing for Graph Neural Networks (SHIELDGNN), a classification method that uses temperature-softened teacher prediction penalties and test-time debiasing for graph-aware predictions, surpassing baselines by up to 7% accuracy and 3% average recall. In this thesis, we also provide a justification for why SHIELD-GNN is able to prevent tail class smoothing with empirical evidence. Finally, we provide early results for ongoing work in tree segmentation.

---

BibTeX citation:

@techreport{Mukherjee:EECS-2024-81,
    Author= {Mukherjee, Sandeep},
    Editor= {Goldberg, Ken and Efros, Alexei (Alyosha)},
    Title= {Automated Tree Censusing from Aerial Imagery with Noisy Supervision},
    Year= {2024},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-81.html},
    Number= {UCB/EECS-2024-81},
    Abstract= {Classifying trees from GPS-registered aerial imagery poses several challenges: data are lowsignal
and noisy and there is a long-tailed, fine-grained class distribution. In this setting, supervised
classification heads trained on DINO features under-perform simple fully-supervised
classifiers. On the other hand, noise-resistant feature extraction excels but can be overly
class-indicative. We find that the Graph Attention Transformer (GAT), is well-specified
to model geo-spatial correlations present in aerial imagery, but tail class sensitivity suffers
when trained on overly class-indicative features (section 6.2). We present Softening Head
Imbalances with Effective Learning and Debiasing for Graph Neural Networks (SHIELDGNN),
a classification method that uses temperature-softened teacher prediction penalties
and test-time debiasing for graph-aware predictions, surpassing baselines by up to 7% accuracy
and 3% average recall.
In this thesis, we also provide a justification for why SHIELD-GNN is able to prevent tail
class smoothing with empirical evidence. Finally, we provide early results for ongoing work
in tree segmentation.},
}

EndNote citation:

%0 Report
%A Mukherjee, Sandeep 
%E Goldberg, Ken 
%E Efros, Alexei (Alyosha) 
%T Automated Tree Censusing from Aerial Imagery with Noisy Supervision
%I EECS Department, University of California, Berkeley
%D 2024
%8 May 10
%@ UCB/EECS-2024-81
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-81.html
%F Mukherjee:EECS-2024-81