Historical data on vegetation fires is necessary for evaluating risks and predict future vulnerability under climatic and socio-economic change. Satellite time-series are the most effective tools to map past burned areas in mountainous regions characterized by high environmental heterogeneity and lack of base data. We evaluated existing global datasets in our study region, northwest Yunnan, China, and found that they are not suitable for accurately quantify the burning activity in these complex landscapes. Small fires, frequent seasonal clouds, rugged topography, fast vegetation recovery, patchy landcover, and no training data for training and tuning classification algorithms, were identified as the major factors limiting remote sensing applications in these areas. Based on these findings, we developed an automated burned area extraction routine which attempts to overcome these specific challenges. We created a 30 years (1987-2018) fire dataset and performed an extensive accuracy assessment. Omission and commission errors were both in the order of 20%, representing a great improvement compared to existing products.