The invention relates to the technical field of extreme learning machine classification algorithms, in particular to an extreme learning machine classification algorithm based on an improved crow search algorithm. The method includes the following steps that an ELM network model is built, an ICSA algorithm is adopted, input weights and threshold values generated randomly by the ELM model are optimized, global and local search performance is balanced by introducing an AP value dynamic gradient function, a Levy flight search method is introduced to avoid blind search, a multi-individual variable factor weighted learning method is introduced to guarantee population diversity, and the global and local search performance is improved. And an adjacent-generation dimension crossing method is introduced to enhance the quality of the optimal individual food storage position, so that a local optimal value is prevented from being obtained, and an accurate prediction result is realized. According to the method, a series of defects caused by randomly generating input weights and threshold values are overcome, the ELM model classification precision is improved, and when ELM model parameters are optimized, based on a traditional CSA algorithm, global and local search performance is balanced by introducing an AP value dynamic gradient function, and blind search is avoided by introducing a Levy flight search method.