Low-cost-sensor correction with strong generalization ability necessitates data-driven techniques to be practical. This study aims to use an ultrasonic anemometer (UA) as a reference sensor to correct multiple cup anemometers (CAs). Paired UA and CA measure the training data, while other CAs collect the test data. Yet, this introduces a distribution discrepancy in the training and test data from different CAs. As wind conditions are below the starting velocity, CA response mainly depends on its mechanical properties. Generally, preparing new training data using the same sensor or bias alignment can alleviate this problem. As an explanatory and efficient data preprocessing, we developed cross-sensor domain adaptation (CSDA) by signal decomposition to reconstruct measurements from two CAs. CSDA contributes to drawing close the distribution of two CA responses by erasing the short-term fluctuations of their measures at low wind speeds. Further, a clustering-based artificial neural network (CANN) is adopted as a transfer-learning model to predict the samples from CA to UA, facing the changing wind conditions. CANN accuracy is assessed at three locations. Comparing the CANN in the single-sensor and cross-sensor usage cases, CSDA-CANN achieved the highest accuracy. Our method can be applied to wind comfort, wind energy, urban sustainability, etc.