[M-AG33] Satellite Land Physical Processes Monitoring at Medium/High/Very High Resolution

[MAG33-01] Globally applicable high resolution burned area mapping by deep learning PlanetScope image pairs

★Invited Papers

*David Roy¹, Haiyan Huang ¹, Luigi Boschetti ², Hugo DeLemos ¹, Hankui Zhang ³ (1.Center for Global Change and Earth Observations, Michigan State University, East Lansing, MI 48824, USA, 2.Department of Forest Rangeland and Fire Sciences, University of Idaho, Moscow ID 83843, USA , 3.Geospatial Sciences Center of Excellence, Department of Geography and Geospatial Sciences, South Dakota State University, Brookings, SD 57007, USA)

Keywords:Fire, Burned area, PlanetScope, Deep Learning , Validation

Satellite data have a long provenance for fire monitoring and have been used to map systematically the area burned at medium and coarse spatial resolution. The commercial sector is providing satellite imagery with extraordinary new capabilities. Notably, the PlanetScope constellation acquires multispectral 3 m images on a global near daily basis that enables mapping of small and spatially fragmented burns that cannot be detected at coarser resolution. In our previously published research, we trained a conventional Unet deep learning model by transfer learning with 30 m Landsat-8 burned area annotation data and refined the model with a small set of PlanetScope interpreted 3 m image pairs. The model was applied to PlanetScope image pairs across Africa acquired one day apart, generating 3 m burned area maps with low (~12%) commission and omission errors. In this study, we present a new model and results derived using (i) globally distributed PlanetScope image pairs selected at locations informed by the NASA MODIS fire products and filtered using solar geometry and cloud-cover criteria, (ii) PlanetScope image pairs acquired from 1 to 16 days apart to better capture the variable temporal persistence of burned areas on the landscape and image acquisitions gaps, (iii) the Swin-Unet deep learning model that incorporates an advanced attention mechanism, (iv) active learning to derive 3 m burned/unburned annotation data set by iteratively retraining the model, manually refining the classified results, and then retraining the model with the refined results and new image pairs. The final Swin-Unet took more than one month to train on 4 NVIDIA Tesla V100S GPUs with >1.2 million 256 × 256 3 m pixel patches extracted from 1,124 annotated image pairs. A global biome-based validation on 486 annotated image pairs not used in the model training, covering 66,445 km², indicates very high accuracy of the 3 m burned area maps (3.4% commission and 4.7% omission errors). This research demonstrates the feasibility of wildfire monitoring at scales never undertaken before, enabling potentially transformative research on the role of small fires in the global carbon cycle, as well as provision of accurate and timely datasets to support fire suppression and post-fire operations, including in the wildland urban interface where disastrous fires are increasingly occurring.