The Mediterranean region has experienced increasing amounts of forest fires, in increasingly larger magnitudes, in the last two decades. Whereas traditional wildfire risk products are of kilometer scales, there is a need for high-resolution risk maps to support at-risk communities. With Earth-observing spectral remote sensing, it is possible to characterize vegetation conditions in the forest and estimate a fire risk based on the retrieval of vegetation traits. While vegetation trait models are developed usually for closed-canopy conditions, measured pixels are usually of a mixture of endmember surfaces. This work quantifies the uncertainty due to this discrepancy and provides a new model called the Dimensionality Reduction Emphasizing Analysis of Mixtures for Spectroscopy (DREAMS). The DREAMS model is estimating endmember reflectance signatures directly from at-sensor radiances, leveraging variations in geometry and texture conditions of the unique surfaces. We then estimate vegetation conditions based on the endmember signatures and calculate an improved fire risk index. We show the reduction in uncertainty and improved prediction performance when using the DREAMS model and evaluate it operationally over a selected study site within an Israeli forest.