Satellite imagery has been widely used to map urbanization processes. To address the urgent need for urban landscape mapping that goes beyond urban footprint analysis, the local climate zone (LCZ) scheme has been increasingly used to reveal the urban forms and functions important to urban heat islands and micro-climates across the globe. As with most supervised classification strategies, proper application of training data is critical for the success of LCZ classification models. However, the collection and application of LCZ training areas brings with it two challenges that may affect mapping success. First, because digitizing training areas is a timeconsuming task, there is a broad effort in the LCZ mapping community to create a crowdsourced data collection among different experts. However, this strategy likely leads to inconsistencies in labels that could weaken models. Second, the LCZ labeling process typically involves the delineation of large zones from which multiple training samples are drawn, but those samples are likely spatially autocorrelated and lead to overly optimistic estimates of model accuracy. Although both effects - inconsistent labeling and spatial autocorrelation - are theoretically possible, it is unknown whether they substantially affect accuracy. We investigated both issues, specifically asking: (i) how do the discrepancies of LCZ labeling by different experts impact broad-scale LCZ mapping? (ii) to what extent does spatial correlation affect model prediction power? We used two classifiers (Random Forests and ResNets) to map eight metropolitan areas in the US into LCZs, comparing training areas drawn by different or consistent interpreters, and data splitting strategy using rules that allow or reduce spatial autocorrelation. We found large discrepancies among results built from crowdsourced training areas digitized by different experts; improving the consistency of labels can lead to substantial improvements in LCZ classification accuracy. Second, we found that spatial autocorrelation can boost the apparent accuracy of the classifier by 16% to 21%, leading to erroneous interpretation of mapping results. The two effects interplay as well: spatial auto correlation in the raw data can lead to an underestimation of the model's predictive error when modeling with crowdsourced training areas of high inconsistency. Due to the uncertainty in the labeling process and spatial autocorrelation in derived training data, broad-scale LCZ mapping results should be interpreted with caution.