Protein　carbonylation　is　a　typical　oxidative　stress-induced　post-translational　modification　(PTM),　which　has　an　important　role　in　proteasome　activity　and　cellular　dysfunctions.　Mass　spectrometry　is　one　of　the　most　common　ways　to　analyze　protein　carbonyl　level　and　identify　exact　modification　sites.　It　has　been　found　that　four　types　of　amino　acid　residues,　including　lysine　(K),　arginine　(R),　threonine　(T)　and　proline　(P),　are　prone　to　carbonylation.　However,　these　experimental　approaches　are　not　suitable　for　batch　processing　of　proteins,　and　the　predictive　power　of　existing　relevant　bioinformatical　tools　are　still　weak.　In　the　paper,　an　improved　method　with　a　new　max-significance　and　min-redundancy　(MSAMR)　feature　selection　criterion　was　proposed　to　predict　human　carbonylation　sites.　This　method　can　achieve　total　accuracies　of　87.43%,　86.83%,　86.56%　and　88.10%　for　K,　R,　T　and　P　carbonylation　site　predictions　respectively　using　10-fold　cross-validation.　Different　kinds　of　features　involved　in　the　method,　especially　a　new　kind　of　customized　amino　acid　composition　(CAAC)　features,　were　analyzed,　and　the　performance　of　MSAMR　feature　selection　criterion　was　discussed.　Furthermore,　a　software　tool　CarSPred2.0　was　released　to　serve　for　human　carbonylation　site　prediction.All　datasets　and　the　software　can　be　available　at　https://github.com/lhqxinghun/bioinformatics/tree/master/CarSPred-2.0/.　©　2018　IEEE.