Abstract: Abstract: Modeling of farmers' willingness to consolidation of abandoned homesteads plays an essential role in the prediction of regional land consolidation potential. Previous prediction models on land consolidation potential still have some limitations in the simulation of farmers' consolidation willingness and the spatial explicit prediction. The complex system modeling and machine learning provide effective tools for the behaviors simulation of land-use stakeholders. Land consolidation potential depends mainly on the farmers' willingness to consolidation, as well as the policies and land use planning. It is difficult to obtain enough negative training samples from the non-reclaimed area where farmers are opposed to the consolidation. There is a balance on the training samples, meaning that most training samples are positive. One-class classification approach has provided a good solution for the classification of imbalanced samples, due to only positive samples is selected to complete the training of classifiers. Hence, one-class classification can be used to solve the negative samples in the modeling of farmers' willingness to land consolidation. Therefore, an one class support vector machine (OCSVM) was selected to simulate the decision-making behaviors of the farmers. The OCSVM has been widely used as a type of one-class classification in image recognition and abnormal detection. A geographic information system (GIS) was used to build the model, in order to predict the land consolidation potential in a spatially explicit way. Furthermore, high-resolution remote sensing images were used to identify the abandoned homesteads in the study region. Pingtang was selected as the study area to evaluate the accuracy of model, where a mountainous and poverty town located in western Guangdong province, China. 4 449 positive samples were obtained, where the farmers would like to confer from the historical land consolidation project data in the study area. Another 141 negative samples were randomly selected from the non-reclaimed areas to evaluate the accuracy of model. Thus, a total of 4 590 unlabeled samples were obtained to train the model. The experimental results showed that the overall accuracy of model reached 96.36%, the prediction accuracy of positive sample was 96.88%, and the accuracy of negative samples was 80.14%, indicating that the performance of model was reliable for the potential prediction of land consolidation. The model was used to predict the land consolidation potential in the whole study area. The total area of abandoned homesteads identified by high-resolution remote sensing images was about 103.96 hm2, whereas, the predicted potential obtained by the model was about 94.74 hm2. However, there were many small spots in the study area that were too fragmented to be reclaimed. According to the land consolidation of Pingtang, the abandoned homesteads that can be reclaimed was only 36.02 hm2, accounting for 34.65% of the total areas. Consequently, terrain factors were also essential to affect the consolidation potential in mountainous and hilly areas. The model can be expected to better support the decision-making of land use planning, regional land remediation planning, and site selection in land remediation project.