Investigating Live Fuel Moisture Content Estimation in Fire-Prone Shrubland from Remote Sensing Using Empirical Modelling and RTM Simulations

Abstract

Previous research has demonstrated that remote sensing can provide spectral information related to vegetation moisture variations essential for estimating live fuel moisture content (LFMC), but accuracy and timeliness still present challenges to using this information operationally. Consequently, many regional administrations are investing important resources in field campaigns for LFMC monitoring, often focusing on indicator species to reduce sampling time and costs. This paper compares different remote sensing approaches to provide LFMC prediction of Cistus ladanifer, a fire-prone shrub species commonly found in Mediterranean areas and used by fire management services as an indicator species for wildfire risk assessment. Spectral indices (SI) were derived from satellite imagery of different spectral, spatial, and temporal resolution, including Sentinel-2 and two different reflectance products of the Moderate Resolution Imaging Spectrometer (MODIS); MCD43A4 and MOD09GA. The SI were used to calibrate empirical models for LFMC estimation using on ground field LFMC measurements from a monospecific shrubland area located in Madrid (Spain). The empirical models were fitted with different statistical methods: simple (LR) and multiple linear regression (MLR), non-linear regression (NLR), and general additive models with splines (GAMs). MCD43A4 images were also used to estimate LFMC from the inversion of radiative transfer models (RTM). Empirical model predictions and RTM simulations of LFMC were validated and compared using an independent sample of LFMC values observed in the field. Empirical models derived from MODIS products and Sentinel-2 data showed R2 between estimated and observed LFMC from 0.72 to 0.75 and mean absolute errors ranging from 11% to 13%. GAMs outperformed regression methods in model calibration, but NLR had better results in model validation. LFMC derived from RTM simulations had a weaker correlation with field data (R2 = 0.49) than the best empirical model fitted with MCD43A4 images (R2 = 0.75). R2 between observations and LFMC derived from RTM ranged from 0.56 to 0.85 when the validation was performed for each year independently. However, these values were still lower than the equivalent statistics using the empirical models (R2 from 0.65 to 0.94) and the mean absolute errors per year for RTM were still high (ranging from 25% to 38%) compared to the empirical model (ranging 7% to 15%). Our results showed that spectral information derived from Sentinel-2 and different MODIS products provide valuable information for LFMC estimation in C. ladanifer shrubland. However, both empirical and RTM approaches tended to overestimate the lowest LFMC values, and therefore further work is needed to improve predictions, especially below the critical LFMC threshold used by fire management services to indicate higher flammability (<80%). Although lower extreme LFMC values are still difficult to estimate, the proposed empirical models may be useful to identify when the critical threshold for high fire risk has been reached with reasonable accuracy. This study demonstrates that remote sensing data is a promising source of information to derive reliable and cost-effective LFMC estimation models that can be used in operational wildfire risk systems.