Face recognition in extreme situations is still challenging to researchers. While several algorithms have shown great recognition results in ideal conditions, accuracy decreases when recognition tasks present a high illumination variation. In this paper, we propose to add two components to the recognition system to make the surf descriptor efficient in such extreme situations. First, we learn a discriminant image filter that maximizes the discrimination of surf. Second, the obtained discriminant surf (d-surf) is further strengthened by using multispectral images instead of broad band images. DSURF and multispectral d-surf (MD-SURF) were evaluated against two face databases: the feret database, which served as a benchmark, and the iris-m3 multispectral face database, which presented sun lighted faces. Our algorithms have been evaluated against three state-of-the-art algorithms that are MBLBP, HGPP and LGBPHS. The results validated the superiority of D-SURF over the traditional surf descriptor, while MD-SURF performed best out of all studied algorithms.