Timely and accurate knowledge of species-level biomass is essential for forest managers to sustain forest resources and respond to various forest disturbance regimes. In this study, maps of species-level biomass in Chinese boreal forests were generated by integrating Moderate Resolution Imaging Spectroradiometer (MODIS) images with forest inventory data using k nearest neighbor (kNN) methods and evaluated at different scales. The performance of 630 kNN models based on different distance metrics, k values, and temporal MODIS predictor variables were compared. Random Forest (RF) showed the best performance among the six distance metrics: RF, Euclidean distance, Mahalanobis distance, most similar neighbor in canonical correlation space, most similar neighbor computed using projection pursuit, and gradient nearest neighbor. No appreciable improvement was observed using multi-month MODIS data compared with using single-month MODIS data. At the pixel scale, species-level biomass for larch and white birch had relatively good accuracy (root mean square deviation < 62.1%), while the other species had poorer accuracy. The accuracy of most species except for willow and spruce was improved up to the ecoregion scale. The maps of species-level biomass captured the effects of disturbances including fire and harvest and can provide useful information for broad-scale forest monitoring over time.