Estimation of live fuel moisture content of shrubland using MODIS and Sentinel-2 images

Abstract

Live fuel moisture content (LFMC) is a critical parameter affecting vegetation flammability and fire behaviour. Reliable and updated estimations of LFMC are needed by fire managers for operational wildfire risk assessment. However, detailed and constant monitoring of LFMC in the field is costly and time-consuming. Remote sensing technologies are an important source of geospatial data that can provide spectral information related to LFMC at different temporal and spatial resolution. In this study, we used a database of LFMC monitoring sampled during 2016 and 2017 (n=81) in a monospecific Cistus ladanifer L. shrubland in Madrid region (Central Spain). C. ladanifer is a representative shrub species commonly found in Mediterranean fire-prone areas, and has been already identified as an indicator species for wildfire risk assessment by different regional fire services. A set of spectral indices (SI) derived from MODIS images (MOD09GA) were calculated at 500 m resolution and compared with field data. We also used Sentinel-2 images for SI retrieval at 20 m resolution with the aim of addressing the scale problem between field sampling site and the low spatial resolution of MODIS data. The same SI were calculated adapting formulations to Sentinel-2 spectral resolution. The timelag between images and field sampling date was limited to a maximum of 2 days for operational purposes. Multiple linear regression analysis was used to assess the potential of SI for LFMC estimation, comparing results for both type of images. Most of the SI tested showed a significant correlation with LFMC data derived from MODIS (n=62) and Sentinel-2 (n=35). For MODIS, the best indices were EVI, VARI, and VIGREEN (R2=0.82, MAE=12%), followed by NDVI and SAVI (R2=0.76, MAE=14%). For Sentinel-2, the best indices were VARI (R2=0.72, MAE=13%), EVI (R2=0.71, MAE=13%), VIGREEN (R2=0.67, MAE=14%) and NDVI (R2=0.62, MAE=14%). In both cases, a significant multivariate model was found including NDVI and VARI, with a slight increase in prediction accuracy compared to simple regression models (R2=0.85 with MAE=11% for MODIS, and R2=0.76, MAE=12% for Sentinel-2). Our findings indicate that MODIS and Sentinel-2 images provide similar results for the SI tested, and that both satellites can be used for near real-time estimation of LFMC in C. ladanifer shrubland. The proposed models can be used to improve monitoring of the variability of LFMC during the year, as well as helping the integration of remote sensing data on wildfire danger rating systems.