Using a Hidden Markov Model for Improving the Spatial-Temporal Consistency of Time Series Land Cover Classification

Time series land cover maps play a key role in monitoring the dynamic change of land use. To obtain classification maps with better spatial-temporal consistency and classification accuracy, this study used an algorithm that incorporated information from spatial and temporal neighboring observations in a hidden Markov model (HMM) to improve the time series land cover maps initially produced by a support vector machine (SVM). To investigate the effects of different initial distributions and transition probability matrices on the classification of the HMM, we designed different experimental schemes with different input elements to verify this algorithm with Landsat and HJ satellite images. In addition, we introduced spatial weights into the HMM to make effective use of spatial information. The experimental results showed that the HMM considered that spatial weights could eliminate the vast majority of illogical land cover transition that may occur in previous pixel-wise classification, and that this model had obvious advantages in spatial-temporal consistency and classification accuracy over some existing classification models.