An online hidden feature extraction algorithm is proposed for unknown and unstructured agricultural environments based on a supervised keel locally linear embedding (SKLLE) algorithm.Firstly,an online obtaining method for scene training samples is given to obtain original feature data.Secondly,Bayesian estimation of the a posteriori probability of a cluster center is performed.Thirdly,nonlinear keel mapping function construction is employed to map the original feature data to hyper-high-dimensional keel space.Fourthly,the automatic determination of hidden feature dimensions is performed using a local manifold leaing algorithm.Then,a low-level manifold computation in hidden space is completed.Finally,long-range scene perception is realized using a 1-NN classifier.Experiments are conducted to show the effectiveness and the influence of parameter selection for the proposed algorithm.The keel principal component analysis (KPCA),locally linear embedding (LLE),and supervised locally linear embedding (SLLE) methods are compared under the same experimental unstructured agricultural environment scene.Test results show that the proposed algorithm is more suitable for unstructured agricultural environments than other existing methods,and that the computational load is significantly reduced.