Abstract: Abstract: The basic prices of agricultural land appraisal (ALA) has the extremely vital significance on establishing a complete set of agricultural land price system and smoothly carry out land assets, such as reasonable allocation of land consolidation rural land management. Existing models such as artificial neural networks, support vector machines, multivariable regression cannot build the complex function relationship fully between the affecting factors and land prices, and the above-mentioned models with shallow structure have no ability to handle a high-dimensional sample set for land appraisal. So deep learning method was firstly introduced into ALA, and a novel method for ALA based on deep belief networks (DBN) was proposed. A group of 19-dimensional original features reflecting status of land location and quality were employed as inputs, and the land prices were used as outputs of DBN model. The parameters of DBN model were firstly initialized by unsupervised learning method with no-label samples, and then fine-tuned by supervised learning method with labeled samples. The land price of each assessment unit can be calculated by using the well-trained DBN deep neural network with the input of feature vector, and the level of assessment unit can be determined by taking advantage of total value frequency distribution histogram, then the benchmark land price for each level can be calculated via the method of area weighted technique. Take the city of Puning in Guangdong province as a case study, feasibility and validity of the model was validated. The results of the present study indicate that: 1) With respect to the artificial neural networks and support vector machines models, the DBN model get better assessment accuracy with a slight increase of 3.61% and 3.12% because of it is able to take the advantage of feature extraction of deep structure, and can enhance its generation ability by a large amount of no-label land price samples; 2) The simulating error of DBN model for land price appraisal is 16.43% when the number of training samples is only 300, which is less than the artificial neural networks and support vector machines models with a least reduction of 6.76%, DBN model gets high assessment accuracy with a small amount of training samples resorting to its unsupervised learning framework, and the assessment accuracy increases with the number of no-label samples; 3) The running time and memory usage of DBN for single training sample is higher than the artificial neural networks and support vector machines models, and the running time of support vector machines model is 193 and 187 s for DBN model when the assessment accuracy reaches about 95%, both of which are equal because of the DBN model need less training samples than the SVM model; 4) The cultivated land assessment units of Puning city are divided into 5 levels based on the their land price results which are calculated based on the DBN model, and the benchmark land price of Puning city ranges from 21.34 to 26.23 yuan/m2 which are calculated based on the area weighted technique. The above experiment results indicate that the assessment accuracy of DBN model is significantly better than the models with shallow structure, and the spatial distribution pattern of benchmark land price and quality level of cultivated land for Puning city are consistent with each other. It is concluded that the method is feasible and effective in measurement and calculation on agricultural base land price.