Fire is a major driver of forest landscape change in boreal forests. Burn severity is one of the main indexes for measuring the damage degree of fire on forest ecosystems. Spatial patterns of burn severity affect numerous ecological processes (e.g., species composition, seed dispersal, and vegetation restoration). Explaining spatial patterns of burn severity is conducive to reveal the formation mechanism of forest landscape patterns after fire, which is of great significance for predicting spatial patterns of burn severity in the future and formulating scientific fire management strategies. Based on Landsat TM/ETM remote sensing images, we mapped the burn severity of 36 fires that occurred between 2000 and 2016 in Huzhong forest region of the Great Xing'an Mountains by calculating the post-fire Normalized Burn Ratio index (NBR) and classified the fires into unburned, low, moderate and high severity classes. For each fire, we calculated five landscape metrics to quantitatively describe spatial patterns of burn severity at the class level using the FRAGSTATS program. The landscape pattern metrics were percentage of landscape (PLAND), area-weighted mean patch size (AREA_AM), area-weighted mean fractal dimension index (FRAC_AM), perimeter-area ratio (PARA_AM), and patch density (PD). Using Random Forest models, we analyzed the relative importance and marginal effects of weather, topography, and vegetation variables on determining spatial patterns of burn severity. The results showed that:1) compared with unburned, low-, and moderate-severity patches, the high-severity patches were more larger and simpler in shape. 2) Elevation played an important role in shaping spatial patterns of burn severity, followed by aspect, slope, vegetation coverage, relative humidity, and temperature. 3) With the increase in elevation, the marginal effect curve of area-weighted mean patch area and area-weighted mean patch fractal dimension showed an obvious increasing trend, whereas area-weighted perimeter-area ratio and patch density exhibited a decreasing trend. In addition to area-weighted mean patch area, all of them were affected by pre-fire vegetation coverage. When pre-fire vegetation coverage ranged fom 0.2 to 0.3, the proportion of high-severity patches in the landscape were the largest. In general, the high-severity patches differed significantly from unburned, low- and moderate-severity patches for five spatial pattern metrics. Topography and vegetation were more important in shaping the spatial pattern of high-severity patches than climate. Therefore, it would be urgent to implement forest fuel treatment in high-severity areas. It is necessary to allocate different forest patches reasonably from the landscape level, then to reduce the risk of high-severity forest large fires.