Spectral　reconstruction　is　an　important　task　in　the　field　of　computer　vision　with　growing　attraction　and　scope　of　applications.　This　task　aims　to　generate　hyperspectral　images　(HSIs)　with　the　same　spatial　resolution　by　improving　the　spectral　resolution　from　RGB　images.　With　the　successful　application　of　convolutional　neural　network,　some　researchers　introduce　this　idea　into　the　task　of　spectral　reconstruction.　In　this　paper,　we　propose　a　novel　method　with　a　cluster-based　deep　residual　convolutional　neural　network　(DRCNN)　for　spectral　reconstruction　from　RGB　images.　To　make　full　use　of　the　local　feature　on　the　residual　branches,　we　propose　a　hierarchical　feature　fusion　(HFF)　module　to　propagate　the　useful　information　of　preceding　blocks　to　the　end　of　the　module.　In　the　training　phase,　we　use　RGB　images　that　contain　pixels　belonging　to　the　same　cluster　to　train　the　specific　DRCNN.　In　the　testing　phase,　we　apply　the　trained　model　to　reconstruct　the　corresponding　HSIs.　As　a　postprocess,　element-wise　addition　is　utilized　to　generate　the　complete　HSI.　Extensive　experiments　show　that　the　proposed　method,　compared　with　the　state-of-the-art　spectral　reconstruction　approaches,　achieves　superior　reconstruction　performance　in　three　hyperspectral　datasets:　CAVE,　Harvard,　and　ICVL.　©　2021　Elsevier　B.V.