Accurate estimation of potential fire intensity (PFI) can improve wildfire management. The PFI can be simulated by fire spread models, but with immeasurable uncertainties. There are also some difficulties in estimating PFI with multi-source drivers, since the fire spread is limited by fire suppression. This study aimed to estimate the probability density of PFI over southwestern China, using time-series fuel and weather data as well as topographic data. The Quantile Regression and Bi-directional Long Short-Term Memory were selected to establish the prediction model of PFI. The results showed that the QR-BiLSTM performed best at the 90% confidence level. The modal PFI values extracted from the estimated probability density were closer to the observed values. This study suggests the potential of probability density estimation of PFI with artificial intelligence, for which improves wildfire risk assessment.

This study presents a radiative transfer process and machine learning combined method to estimate the crown foliage fuel load (FFL), an important factor influencing the characteristics of crown fires. To this end, the GeoSail, SAIL and PROSPECT radiative transfer models (RTMs) were firstly coupled together to simulate the near realistic scenario of a two-layered forest structure. For the backward inversion process, the coupled RTMs were linked to the machine learning models of multi-layer perceptron (MLP) to train this model and then derive the FFL estimates from the Landsat products. The performance of retrieved FFL was validated and compared with the traditional look-up table (LUT) method. Results showed that the MLP performed better than LUT, revealing the reasonable skill of the combination of RTMs and machine learning modeling in deriving the FFL from remote sensing data and putting the insight into wildfire danger assessment from space.

Forest biomass is a complex parameter that is a function of multiple forest structure parameters, such as forest height, DBH, and woody density. At present, forest height is most likely to be estimated over the globe from spaceborne remote sensing, particularly with the upcoming BIOMASS mission. This study inverted forest height using P-band repeat-pass multi-baseline PolInSAR data and evaluated the robustness across three different forest sites. Afterward, the PolInSAR height was converted into forest AGB. Results show that the method produced a more accurate forest height in tropic forests than that in boreal forests, but there is little bias for forest height estimates across three sites, with$\mathrm{R}^{2}$of 0.86 and RMSE of 3.1m in contrast to LiDAR forest height. One power equation was found able to convert PolInSAR forest height into forest AGB across three sites, producing$\mathrm{R}^{2}$of 0.85 and RMSE of 51.8 tons/ha using LOO cross-validation, which is at a similar level to the AGB accuracy estimated using LiDAR-derived forest height$(\mathrm{R}^{2}$of 0.87 and RMSE of 47.9 tons/ha).

Forest fuel load plays an important role in fire ignition, spread and intensity. Accurate spatial distribution information of fuel load is vital for fire managers to make decisions. However, most of the existing studies using spectral or texture information of optical data, which is seriously affected by atmospheric conditions with limited penetration. SAR with all-day and all-weather work characteristics provides a favorable opportunity to estimate fuel loads in Southwest China with cloudy and foggy. Moreover, SAR can penetrate through leaves to branches and stems, which reflects the forest vertical structure well. However, there was little research applying SAR data to fuel load estimation. In this study, we focus on forest above ground live fuel load estimation (biomass fractions), including stem fuel load (SFL), branch fuel load (BFL) and foliage fuel load (FFL). We explored the potential of dual polarimetric data, Sentinel-1 for coniferous forest fuel loads estimation in the Southwest of Sichuan, China. To understand scattering mechanisms at C-band in Pinus yunnanensis forest, the Michigan Microwave Canopy Scattering (MIMICS) radiative transfer model was used. And Multiple Linear Regression (MLR) method was used to estimate fuel loads. Results show that VH and VV polarizations were both sensitive to three types of fuel load. Combined with the simulation of MIMICS, we found VH polarization was more sensitive to FFL while VV was more sensitive to SFL. Additionally, Sentinel-1 performs well in all three types of fuel load estimation (FFL: R2= 0.52, RMSE=1.43 Tons/ha; BFL: R2=0.58, RMSE=1.88 Tons/ha; SFL: R2=0.57, RMSE=2.97 Tons/ha), indicating that Sentinel-1 data has great potential in FFL estimation and fire prevention.

Forest Dead Fuel Load (FDFL) is vital for fire prevention and suppression since it affects the surface fire ignition and intensity directly. The accurate spatial distribution information of FDFL can provide decision support for fire managers. Remote sensing (RS) technology is a unique way to estimate FDFL on a large scale. However, most researches focus on the application of LiDAR data which is expensive to analyze fuel dynamics on spatiotemporal scale. Little attention has been devoted to freely accessible datasets such as Sentinel-1 and Sentinel-2 with global coverage. This study not only combined these two data but also the site conditions (i.e., elevation, slope and aspect) to estimate the FDFL in the southwest of Sichuan, China. The machine learning method, Random Forest Regression (RFR) was selected to manage the multiple and nonlinear relationships between RS data and FDFL. Results show that 1h and all FDFL (the sum of 1h, 10h, 100h and litter) can be more indicated by RS data (1h:$\mathrm{R}^{2}=0.57,\ \text{RMSE}=0.18$Tons/ha; All:$\mathrm{R}^{2}=0.59,\ \text{RMSE}=1.81$Tons/ha). The representation ability of RS data for 10h, 100h and litter is relatively weaker (10h:$\mathrm{R}^{2}=0.41,\ \text{RMSE}=0.47$Tons/ha; 100h:$\mathrm{R}^{2}=0.40,\ \ \text{RMSE}=1.07$Tons/ha; litter:$\mathrm{R}^{2}=0.29,\quad \text{RMSE}=1.66$Tons/ha). Hence, this study demonstrated the potential of multi-source RS data for FDFL estimation.