We　consider　the　two-stage　submodular　maximization　problem　which　has　been　studied　extensively　in　machine　learning.　In　this　problem,　we　are　given　a　ground　set　V　of　n　articles　and　m　categories,　and　the　goal　is　to　select　a　subset　S　subset　of　V　of　articles　that　best　represents　the　different　categories　to　maximize　the　total　similarity　F(S)　=　Sigma(m)(j=1)　max(T　is　an　element　of　I(S))　f(j)(T),　where　each　f(j)　is　a　nonnegative　monotone　submodular　functions　measuring　the　similarity　of　each　article　in　category　j.　We　consider　the　case　where　S　satisfies　the　knapsack　constraint　and　I(S)　is　a　k-matroid　constraint　and　present　a　1/2(k+1)　(1　-　e(-(k+1)))-approximation　algorithm　for　this　problem.　We　also　extend　the　k-matroid　constraint　to　k-exchange　system　constraint　and　give　corresponding　approximation　ratio.