Data assimilation algorithms rely on a basic assumption of an unbiased observation error. However, the presence of inconsistent measurements with nontrivial biases or inseparable baselines is unavoidable in practice. Assimilation analysis might diverge from reality since the data assimilation itself cannot distinguish whether the differences between model simulations and observations are due to the biased observations or model deficiencies. Unfortunately, modeling of observation biases or baselines which show strong spatiotemporal variability is a challenging task. In this study, we report how data-driven machine learning can be used to perform observation bias correction for data assimilation through a real application, which is the dust emission inversion using PM10 observations. PM10 observations are considered unbiased; however, a bias correction is necessary if they are used as a proxy for dust during dust storms since they actually represent a sum of dust particles and non-dust aerosols. Two observation bias correction methods have been designed in order to use PM10 measurements as proxy for the dust storm loads under severe dust conditions. The first one is the conventional chemistry transport model (CTM) that simulates life cycles of non-dust aerosols. The other one is the machine-learning model that describes the relations between the regular PM10 and other air quality measurements. The latter is trained by learning using 2 years of historical samples. The machine-learning-based non-dust model is shown to be in better agreement with observations compared to the CTM. The dust emission inversion tests have been performed, through assimilating either the raw measurements or the bias-corrected dust observations using either the CTM or machine-learning model. The emission field, surface dust concentration, and forecast skill are evaluated. The worst case is when we directly assimilate the original observations. The forecasts driven by the a posteriori emission in this case even result in larger errors than the reference prediction. This shows the necessities of bias correction in data assimilation. The best results are obtained when using the machine-learning model for bias correction, with the existing measurements used more precisely and the resulting forecasts close to reality.