Empirical　mode　decomposition　(EMD)　is　a　time-frequency　analysis　method　for　non-stationary　and　nonlinear　signal.　The　EMD　decomposes　signal　into　a　collection　of　intrinsic　mode　functions　(IMFs),　which　can　highlight　the　local　characteristics　of　the　original　signal　and　discriminate　the　signal　from　the　noise.　However　the　method　has　a　problem　of　end　effect.　The　problem　makes　the　IMFs　polluted　in　the　sifting　process　of　EMD,　and　some　feature　information　of　the　original　signal　is　lost　after　reconstruction.　In　order　to　mitigate　this　problem,　an　extreme　point　expansion　method　based　on　kernel　extreme　learning　machine　(KELM)　is　proposed　in　this　paper.　The　proposed　method　is　utilized　to　predict　several　extreme　points　at　both　ends　of　the　original　signal,　improve　the　accuracy　of　sifting　process,　increase　the　effectiveness　of　the　reconstructed　signal.　In　case　of　compared　with　the　other　traditional　methods,　firstly,　the　proposed　method　is　verified　by　the　simulate　signal　decomposition,　simulation　results　demonstrate　the　method　can　restrain　end　effect　effectively　and　improve　the　decomposition　results　of　EMD;　secondly,　the　proposed　method　is　applied　to　denoising　wheat　canopy　reflectance　spectrum　and　the　result　shows　that　the　method　can　offer　better　signal-to-noise　ratio　(SNR)　and　root　mean　square　error　(RMSE),　remove　noise　from　the　original　signal　accurately　and　effectively.　©　2015　Technical　Committee　on　Control　Theory,　Chinese　Association　of　Automation.