Forest foliage fuel load estimation from multi-sensor spatiotemporal features

Foliage fuel is the most flammable component in crown fires. Spatiotemporal dynamics of foliage fuel load (FFL) are important for fire managers to assess fire risk. Here, we integrated optical data from the Landsat 8 Operational Land Imager (OLI) with synthetic aperture radar (SAR) data from Sentinel-1 to estimate FFL. We first reconstructed seamless time series from the Landsat 8 and Sentinel-1 imagery by accounting for unequal time intervals between image observations and outliers. We then extracted temporal features that are proxies of the intra- and inter-annual dynamics from these time series. In addition, we derived spatial features from the imagery that quantify spatial context and therefore used varying window sizes. The random forest regression was implemented to assess the importance of the spatiotemporal features, reduce errors, and derive robust FFL estimates. The satellite estimates were validated against 96 field measurements from Pinus yunnanensis forests in the Liangshan Yi Autonomous Prefecture, Sichuan Province, China. Both the spatiotemporal features of SAR and optical data importantly contributed to FFL estimation. When only optical data was used, the model achieved a R² of 0.75 (relative Root Mean Squared Error (rRMSE) = 25.3 %), while when only SAR data was used the R² was 0.76 (rRMSE = 25.6 %). However, when optical and SAR data were combined, the R² increased to 0.81 (rRMSE = 23.2 %). We also found that temporal features were more important predictors of FFL than features that captured spatial context. We demonstrated our FFL mapping method by a case study in the Chinese Sichuan Province, in relation to the occurrence of a fire. Our method needs additional validation over different tree species and forest types, yet has potential for mapping forest fuel loads and fire risk.