Abstract: Flame spectra contain useful information about combustion and hence the spectral features of flame radicals may be used to identify different biomass fuels. A technique for biomass fuel identification was proposed based on the spectral features of flame radicals, feature engineering and improved support vector machine. The spectral intensity signals of biomass flames and flame radicals（OH*（310.85 nm）, CN*（390.00 nm）, CH*（430.57 nm） and C₂*（515.23 nm, 545.59 nm）） were acquired using a spectrometer. Feature engineering was built, which can accurately reflect the characteristics of sample category, through feature extraction, feature selection based on Filter and feature learning based on dictionary learning. The support vector machine was used to build the identification model, where radial basis kernel parameter γ and error penalty factor C are optimized using an improved grid search algorithm. Experimental results from a laboratory-scale combustion rig show the effectiveness of the proposed method for the identification of biomass fuel.