Investigating the Role of Spatial Resolution and Vegetation Structure in Live Fuel Moisture Content Estimation from Microwave Remote Sensing

Live fuel moisture content (LFMC) - the mass of water per unit dry biomass in vegetation - is a key factor in determining fire risk. Because LFMC varies with plant species and physiological traits, estimates from optical remote sensing and meteorological indices are often insufficient, making LFMC difficult to capture at large spatial scales. Microwave backscatter, due to its sensitivity to water, provides a complementary source of information to optical remote sensing for estimating LFMC. However, microwave backscatter data from remote sensing contains speckle noise, which can be accounted for by retrieving the data at a coarser spatial resolution.

This thesis aims to understand if accounting for plant heterogeneity with a finer resolution can outweigh the benefits of speckle noise reduction from a coarser resolution when estimating LFMC. This thesis also aims to understand if accounting for vegetation structure traits like canopy cover and canopy bulk density can lead to improved estimations of LFMC, since these traits are representative of plant physiology.

In this thesis, I built off of Rao et al (2020)’s recurrent neural network for estimating LFMC across the Western United States, which produced 15-day LFMC maps for the region at a 250 meter resolution. The training data from the National Fuel Moisture Database was rescaled so that the model from Rao et al. (2020) could be trained and validated at six different common spatial resolutions. Of the six resolutions tested, 250 m leads to the most optimal trade off between a lower root mean squared error and sufficient speckle noise reduction.

Using vegetation data from the California Forest Observatory, I then updated the model inputs with canopy structure variables. Although initial results suggested that training data that was more representative of the validation data would lead to improved estimates, adding the representative CFO data led to a drop in performance, indicating the need for an improved vegetation structure dataset.