Validation of Japanese Collision Risk Models Using Post-Construction Carcass Survey Data

Yasuo SHIMADA, Hiroyuki Matsuda

Collision risk models (CRMs) are widely used to predict bird collisions with wind turbines; however, few studies have evaluated their performance using post-construction monitoring data. In this study, we assessed the validity of two collision risk models commonly used in Japan, the Ministry of the Environment model (CRMmoe) and the Yui and Shimada (2013) model (CRMyui), by comparing model predictions with carcass survey results from two wind power facilities. A total of 14 validation cases were recorded during post-construction monitoring. Assuming that the cumulative number of detected carcasses follows a Poisson distribution, we estimated 95% confidence intervals for cumulative collision numbers using three interval-estimation methods: the log-transformed Wald method, the Garwood method, and the likelihood-ratio method. Predicted collision numbers from the CRMs were compared with these intervals after accounting for effective survey coverage. The proportions of model predictions that fell within the estimated 95% confidence intervals were 62.5% (5/8 cases) for CRMmoe and 71.4% (10/14 cases) for CRMyui, both substantially lower than the nominal coverage probability of 95%. All predictions falling outside the confidence intervals corresponded to underestimation of collision numbers inferred from carcass survey data. A generalized linear model incorporating survey coverage as an offset further indicated that observed carcass counts were approximately 5.4 times higher than CRM predictions under the reference condition, suggesting a systematic tendency toward underestimation.
The avoidance rate, one of the few CRM parameters not estimated directly from site-specific observations, was evaluated using its original definition based on the ratio of observed to predicted collisions assuming no avoidance. Although avoidance rate was considered a plausible source of the discrepancy, the estimated scaling factor may also reflect uncertainty in other CRM components and the cumulative propagation of uncertainty throughout the model. Our results demonstrate the limitations of evaluating CRM performance using point estimates alone and highlight the importance of interval-based validation, long-term post-construction monitoring, and explicit treatment of uncertainty in collision-risk assessments for wind power facilities.