This paper proposes an improved farmland fertility (IFF) algorithm to increase the convergence rate and precision of the farmland fertility algorithm. A search mode that combines subspace and full space is proposed. The two modes are automatically converted in accordance with the current learning level of the population. Such hybrid search balances the algorithm exploration and exploitation capabilities. The global memory capacity with a fixed size is processed self-adaptively to make its size adaptively change with the iterative process. A mechanism of soil fusion based on neighbor memory learning is also proposed to expand the search range of the current population and ensure population diversity. On the basis of the cosine similarity between the population and the current optimal individual, the area to which the individual belongs is periodically redivided, and the computing resources are reasonably allocated. The results of experiments on the CEC2013 test suite indicate that IFF has evident advantages over seven excellent algorithms in terms of convergence rate and precision.