Wood moisture content is an important technical specification in the wood drying process. Considering the strong coupling, large lag non-linear features of the wood drying process and the problem of low precision of wood moisture content detection, we proposed a soft sensor method using least squares support vector machines (LS-SVM) to learn time series data of a non-linear system, and built a soft sensor model of the controlled object. We also used the particle swarm optimization (PSO) algorithm in the moving horizon optimization of the penalty factor and the kernel function parameter of LS-SVM to improve the prediction precision of the soft sensor model. Taking the inner temperature and humidity of a wood drying kiln as the sample data, the wood moisture content at a specific point can be detected with the model based on LS-SVM optimized by PSO, which is denoted by PSO-LSSVM. The simulation reveals that the PSO-LSSVM has a high prediction precision and strong generalization ability, and can fulfill the actual measurement demand of a wood drying control system.