Traditional climate data, such as air temperature and precipitation, have been widely used in various models for crop yield prediction. One of the major challenges is that most of these climate data were derived from either reanalysis products with relatively coarser spatial resolution or from in situ measurements with limited representativeness, which would inevitably reveal significant mismatches regarding spatial scale with other synchronously used vegetation and soil parameters at high resolution (e.g., ~1 km). To this end, satellite-derived land surface temperature (LST) and soil moisture (SM) at a high spatial resolution of 1 km were used as proxies of air temperature and precipitation to evaluate the feasibility of predicting maize yield in three major regions (northeast, northwest, and north China). Specifically, each region includes three provinces. Three widely used machine learning models, namely, the gradient boosting decision tree, extreme gradient boosted tree, and random forest (RF), were considered to avoid the contingency of a single model. In this study, the three models were trained at two spatial scales: 1) region by region and 2) entire maize planting area. Results indicated that using satellite-based LST and SM instead of traditional climate data of air temperature and precipitation can obtain a significantly improved maize yield prediction with the average root mean square error decreased from 862 to 827 kg/ha when the models were trained region by region and from 894 to 840 kg/ha when the models were trained over the entire maize planting area.