Validating a Bird Detection Machine-Vision Model
Session Number
CMPS 05
Advisor(s)
Dr. Yuki Hamada, Paul Tarpey, Adam Szymanski, Argonne National Lab
Discipline
Computer Science
Start Date
17-4-2025 2:30 PM
End Date
17-4-2025 2:45 PM
Abstract
Recent development of large-scale solar farms could pose an environmental threat to birds, calling for accurate monitoring to ensure that photovoltaic solar energy development does not come at the cost of wildlife conservation. Recently, Argonne National Laboratory has produced a novel machine-vision model to track bird activity, but its performance for bird prediction requires further evaluation.
Here, we aim to compare bird detections from manual video review alongside the model’s output, noting possible discrepancies and the underlying factors that could have caused them. The results show that bird detection agreement between human interpretation and model prediction is higher when fewer birds are present yet increased when the model detected no birds or when human counts exceeded seven birds. With two sets of data collected on differing days, the disagreement percentage was calculated at 2.70% on the May 8 dataset and 17.24% for the May 14 dataset, notable inconsistencies in the model’s performance. These discrepancies were mostly observed in the early morning and late afternoon, suggesting that lighting conditions may impact the model’s accuracy. These findings in this study will help in improving the current bird monitoring technology for solar farms.
Validating a Bird Detection Machine-Vision Model
Recent development of large-scale solar farms could pose an environmental threat to birds, calling for accurate monitoring to ensure that photovoltaic solar energy development does not come at the cost of wildlife conservation. Recently, Argonne National Laboratory has produced a novel machine-vision model to track bird activity, but its performance for bird prediction requires further evaluation.
Here, we aim to compare bird detections from manual video review alongside the model’s output, noting possible discrepancies and the underlying factors that could have caused them. The results show that bird detection agreement between human interpretation and model prediction is higher when fewer birds are present yet increased when the model detected no birds or when human counts exceeded seven birds. With two sets of data collected on differing days, the disagreement percentage was calculated at 2.70% on the May 8 dataset and 17.24% for the May 14 dataset, notable inconsistencies in the model’s performance. These discrepancies were mostly observed in the early morning and late afternoon, suggesting that lighting conditions may impact the model’s accuracy. These findings in this study will help in improving the current bird monitoring technology for solar farms.