Knowledge of crustal stress fields is essential for understanding
tectonics and earthquake generation. One approach for estimating the
crustal stress field is based on the focal mechanisms of earthquakes.
This study investigated the focal mechanisms of approximately 110,000
microearthquakes in the area of the Japanese islands that occurred at a
depth shallower than 20 km, based on the first-motion polarities picked
by a simple neural network model. The model was first trained using a
data set of mainly moderate to large earthquakes throughout Japan.
Following on, the model was re-trained using a data set of
microearthquakes in two regions of Japan. The threshold of the
confidence score from the neural network model was chosen to maximize
the overall quality of the focal mechanism solutions. The P- and T-axes
of the numerous focal mechanism solutions provided more detailed
distributions of the crustal stress field. For example, in the Chugoku
region, small differences were observed in the trend of P-axes azimuths
between the northern and southern areas, spatially corresponding to
geodetic observations. The results of this study are useful for
revealing the crustal stress field, and, as such, for assessing past and
current tectonic activities and potential future earthquake generation.