Application of Domain Adaptation Approach for Weather Data Mining

The fast increase in the availability of weather data from various sensors and weather stations allows weather data mining to achieve much higher accuracy over time, serving for important economic and socioeconomic purposes. However, the availability and sparsity of weather data differs drastically for geologically separated locations and there exists wide across domain differences for different sources, resulting in various accuracy in predicting the weather for target locations with different weather patterns. This paper applies domain adaptation approach for weather classification, where a system is trained from one source domain but deployed on another target domain. This methodology outperforms other two alternative methods, showing lower misclassification rate than using only target domain or naïve combination of both target and source domain ignoring cross-domain differences. This work provides a framework for future weather data mining and encourages the domain adaptation approach in other applications in data mining with wide cross-domain differences in general.