An automated burned area extraction routine that attempts to overcome the particular difficulties of remote sensing applications in complex landscapes is presented and tested in the mountainous region of northwest Yunnan, China. In particular, the lack of burned samples to use for training and testing, the rugged relief, the small size of fires and the constant presence of clouds during the rainy season heavily affecting the number of usable scenes within a year are addressed. The algorithm flows through five phases: creation of standardised difference vegetation indices time series; automatic extraction of multiclass training areas using adaptive z-score thresholds; Random Forests classification; Seeded Region Growing; and spatiotemporal clustering to form polygons representing fire events. A final database spanning the period 1987–2018 was created. Accuracy assessment of location and number of fire polygons using a stratified random sampling design showed satisfactory results with reduced omission and commission errors compared with global fire products in the same region (20 and 22% respectively). Mapping accuracy of single burned areas showed higher omission (27%) but reduced commission (13%) errors. This methodology takes a step forward towards the inclusion of regions characterised by small fires that are often poorly represented in impact assessments at the global scale.