The model provides highly accurate predictions by integrating diverse information, including past encounter records, environmental factors, population data, and weather conditions

Shin Nakamoto and Associate Professor Yusuke Fukazawa from the Graduate Program of Applied Data Science have developed a model to predict bear encounters in Akita Prefecture. This study integrated various factors, including past encounter records, land cover, population distribution, weather, proximity to roads, elevation, and the abundance of beech nuts (a key food source for bears), to create a time-series prediction of bear encounters.

After addressing the data imbalance during the training process, the developed prediction model achieved a high accuracy rate of 63.7%, with a precision rate of 63.5% and a recall rate of 63.6%. Furthermore, analysis of important features revealed that past encounter conditions, land-cover types (such as artificial structures, rice paddies, and bamboo forests), the distribution of the elderly population, and elevation were particularly significant factors in predicting bear encounters.

These findings are expected to help identify areas at high risk for bear encounters in advance, allowing for the effective distribution of warning information and implementation of countermeasures. The results of this research were published online in the international academic journal “International Journal of Data Science and Analytics” on July 22, 2025.

Using this model, the probability of bear encounters has been predicted based on publicly available data from each municipality, and the results are presented as the “Bear Encounter AI Prediction Map.” This map was created using the model developed in the laboratory of Associate Professor Yusuke Fukazawa, who led this research. Please note that this map is intended solely as a tool for raising awareness and caution. Users are advised to consider this information in conjunction with other sources and up-to-date local information.

In recent years, the number of human casualties resulting from bear encounters in urban areas has been increasing nationwide. Specifically, in Akita Prefecture, bear encounters surged from an annual average of 800 to > 3,900 in 2023. This increase has been attributed to a combination of environmental factors, including a poor harvest of beech nuts—a key food source for bears—along with depopulation and the expansion of abandoned farmland in mountainous and hilly regions.

To help prevent human injuries from unexpected bear encounters, it is essential to predict the risk of bear encounters in advance and to implement warnings and appropriate countermeasures for residents.

In this study, we designed various features to enhance the accuracy of our bear encounter prediction model. To consider the temporal trends in bear behavior, we included past encounter data along with their dates and times. We also incorporated land-cover data that reflects population structures and vegetation in the surrounding environment, as well as meteorological data capturing weather conditions, temperature, humidity, and precipitation on the day of the encounter and the day prior.

Additionally, elevation data and information regarding beech nut abundance or scarcity were included, as these factors influence bear activity. Recognizing that human activity affects bear encounters and reports, we integrated age-specific population distribution data and road information. This comprehensive approach allows the model to learn the interactions between various temporal, environmental, and social factors that influence bear encounters.

The prediction model uses the presence or absence of daily encounters across 1-km mesh grids throughout Akita Prefecture as the objective variable. One challenge in developing the prediction model was the class imbalance of the objective variable in the training data, which hindered the learning of features. To address this issue, we employed a combination of similarity-based undersampling and random undersampling to adjust the number of data points for each class. Ultimately, the training data for the 2021–2022 fiscal year comprised 1,736 bear encounters and 2,078 cases of no encounters (totaling 3,814 instances). In contrast, the test data for the 2023 fiscal year consisted of 3,981 encounters and 4,772 cases of no encounters (totaling 8,753 instances).

Using the ExtraTrees model, the prediction achieved an accuracy rate of 63.7%, a precision rate of 63.5%, and a recall rate of 63.6%. This performance surpassed that of conventional simple spatial and temporal rules as well as k-nearest neighbor models. Additionally, SHAP analysis was conducted to identify the most important features in the prediction model. This analysis revealed that past bear encounter patterns in the vicinity, land cover (including artificial structures, rice paddies, and bamboo forests), population distribution (particularly the number of elderly residents), and elevation were the key predictive factors. Furthermore, a Leave-One-Location-Out evaluation confirmed stable prediction performance even in unknown locations, demonstrating regional generalization capability (accuracy 0.601, precision 0.568, and recall 0.563).

This research enabled advanced bear encounter predictions by integrating not only past encounter patterns but also relevant temporal, environmental, and social information, while effectively addressing the issue of imbalanced data.

In the future, we plan to enhance the predictions of the model by integrating real-time weather data and updated bear encounter information. Additionally, we aim to apply this approach to other regions and assess its effectiveness in various environments. Through these efforts, we expect to contribute to nationwide strategies that address bear encounters.

For inquiries about the research described above, please contact:

Media Contact: