In　recent　years,　with　the　rapid　development　of　economy　and　urbanization,　a　large　number　of　parks　have　been　built　all　over　the　world.　The　number　of　fire　accidents　occurring　in　the　parks　is　also　increasing,　resulting　in　tremendous　casualties　and　property　damage.　These　parks　usually　cover　multiple　sub-scenes,　such　as　houses,　vehicles　and　greenery,　and　each　sub-scene　has　its　own　unique　fire　accident　features.　Existing　fire　detection　methods　cannot　meet　the　diversity　of　sub-scenes　in　the　park.　To　improve　the　detection　capability　of　each　sub-scene　model,　we　propose　a　cloud　edge　collaboration　framework　based　on　transfer　learning　for　training　fire　detection　models　with　multiple　sub-scenes　in　a　park.　The　framework　has　two　phases:　initialization　phase　and　update　phase,　using　the　rich　data　and　powerful　computing　power　in　the　cloud　to　assist　the　edge　nodes　with　limited　resources　to　train　the　models.　First,　the　cloud　integrates　the　fire　data　of　all　edge　nodes　for　model　training　to　obtain　all　fire　features,　and　then　each　edge　node　uses　its　own　data　set　to　fine-tune　the　cloud-trained　model　after　freezing　some　parameters.　In　this　paper,　we　construct　a　fire　image　data　set　with　multiple　sub-scenes　in　a　park　and　use　the　YOLOv5s　model　for　experiments,　and　the　experimental　results　show　that　the　framework　can　improve　the　mAP　of　each　edge　node　model.　This　paper　provides　a　novel　approach　for　fire　detection,　which　is　of　positive　significance　for　fire　prevention　and　control.　©　2023　SPIE.