Detecting and Tracking Communal Bird Roosts in Weather Radar Data

Zezhou Cheng, Saadia Gabriel, Pankaj Bhambhani, Daniel Sheldon, Subhransu Maji, Andrew Laughlin, David Winkler

Introduction

For decades, the US NEXRAD weather radar network has unintentionally captured a biological goldmine: the morning "exodus" of millions of birds, such as Tree Swallows, as they leave their communal roosts. These events appear on radar as distinctive, expanding rings. While this data offers a unique opportunity to understand bird migration and population trends at a continental scale, the sheer volume—spanning 20 years across 143 stations—makes manual analysis impossible.

Communal roost exodus and radar geometry showing the appearance of roost rings over time.

The Challenge of Labeling Noise

A significant technical hurdle in automating this process is labeling noise. Existing "ground truth" labels come from various researchers with differing annotation styles. For instance, one annotator might draw a tight bounding box around the center of a roost, while another might trace the outer edges of the expanding ring. Standard machine learning models often struggle with such inconsistencies, as they are penalized during training for correct detections that do not perfectly match a specific annotator's idiosyncratic style.

Proposed Methodology

The researchers developed a multi-stage AI pipeline to address these challenges:

  1. Detection: A Faster R-CNN architecture was modified to ingest three radar channels (reflectivity at two elevations and radial velocity) rendered as RGB-like images.
  2. Handling Label Bias: The core innovation is treating the "true" roost location as a latent variable. Using a Variational Expectation-Maximization (EM) algorithm, the system jointly learns the bird detector and a model of each individual annotator's labeling style (specifically, their bounding box scaling).
  3. Tracking: The system employs a "detect-then-track" approach. It uses a greedy heuristic to link detections across time frames, followed by a Kalman smoother to model the growth and movement of the roost.
  4. Post-processing: To filter out non-biological "look-alikes" like rain or wind farms, the pipeline integrates auxiliary data from the U.S. Wind Turbine Database and utilizes dual-polarization radar products to distinguish biological matter from water.
Visual representation of the tracking system linking detections across radar scans.

Datasets and Evaluation

The system was trained and tested on an extensive dataset:

  • NEXRAD Archive: Covering the contiguous US over two decades.
  • Manual Annotations: 88,972 radar scans containing 63,691 labeled roosts from seven eastern US stations.
  • Case Study: 419,000 additional scans from 86 stations used to demonstrate "in the wild" performance.

Key Results and Insights

The study revealed that accounting for user-specific bias significantly improves detection accuracy. On the "KDOX" station, which featured highly inconsistent labels, the Mean Average Precision (MAP) jumped from 9.1% to 50.2% after applying the EM framework. In a blind test of the top 521 tracks, the system achieved 93.5% precision.

Beyond technical metrics, the system enabled significant biological discovery. It successfully mapped the fall migration of Tree Swallows across the Eastern US, highlighting their reliance on coastal wetlands and Mississippi Valley croplands. The analysis also calculated the average airspeed of dispersing swallows at approximately 6.61 m/s.

Limitations and Future Directions

While highly effective, the current evaluation metric requires knowing which user provided a label. Additionally, distinguishing between different species that form similar radar rings (e.g., robins vs. blackbirds) remains a challenge. Future work aims to address more complex biases, such as annotators missing small roosts or misidentifying specific weather patterns.

Summary updated

Abstract

The US weather radar archive holds detailed information about biological phenomena in the atmosphere over the last 20 years. Communally roosting birds congregate in large numbers at nighttime roosting locations, and their morning exodus from the roost is often visible as a distinctive pattern in radar images. This paper describes a machine learning system to detect and track roost signatures in weather radar data. A significant challenge is that labels were collected opportunistically from previous research studies and there are systematic differences in labeling style. We contribute a latent-variable model and EM algorithm to learn a detection model together with models of labeling styles for individual annotators. By properly accounting for these variations we learn a significantly more accurate detector. The resulting system detects previously unknown roosting locations and provides comprehensive spatio-temporal data about roosts across the US. This data will provide biologists important information about the poorly understood phenomena of broad-scale habitat use and movements of communally roosting birds during the non-breeding season.
BibTeX
@inproceedings{cheng2020detecting,
  author = {Zezhou Cheng and Saadia Gabriel and Pankaj Bhambhani and Daniel Sheldon and Subhransu Maji and Andrew Laughlin and David Winkler},
  title = {{Detecting and Tracking Communal Bird Roosts in Weather Radar Data}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i01.5373},
}
Figure