Abstract: Grain security is a complex, resource-intensive problem being addressed by governments, international organizations, and scientific community. Ensuring adequate grain supply is vital to the survival of humanity, and its key lies in improving total agricultural land productivity. The level of the comprehensive productive capacity of agriculture is directly related to the effective supply of grains. In 2011, China conducted a nationwide agricultural productivity survey in order to improve its overall capacity, protect farm quality, and accomplish the National Food Security Strategy. The core components of agricultural land productivity calculations are the theoretical and accessible yields. The theoretical yield calculation was traditionally performed by establishing samples' linear regression models between natural level index and theoretical yield. The accessible yield, meanwhile, was performed by establishing samples' linear regression models between usage level index and accessible yield. Then the agricultural land productivity can be calculated by the yield multiplying with the total area of classification units. However, these models suffer from significant calculative limitations. A novel approach based on the Genetic Algorithm and Support Vector Machine (GA-SVM) is proposed herein. The theoretical yield can be calculated by establishing a GA-SVM model between the theoretical yield and quality scores of gradation factors. The accessible yield can be calculated by establishing a GA-SVM model between the accessible yield and the product of usage factor and quality scores of gradation factors. The theoretical and accessible yield prediction models are built using Matlab software, taking full advantage of the GA and SVM toolbox, to test the rationality and accuracy of the new GA-SVM model. Jiexi County in Guangdong Province is provided as an example to verify the theoretical and accessible yields based on the GA-SVM models. The GA-SVM results as well as the traditional linear regression method results are then compared with the actual yields. The analysis reveals several interesting conclusions. First, GA-SVM requires less computational time to provide more accurate results. The linear regression models turn to cumbersome process and need many calculations. Some human errors occurred. However, the GA-SVM models acquire results by running programs. This method can save time and effort, also can avoid human errors. Second, the paper compared the results of GA-SVM model and the actual values, also compared the results of linear regression method and the actual values. The study found that the GA-SVM yield predictions are much closer to the actual values than those provided by the linear regression method. In addition, GA-SVM model is more suitable in forecasting a single sample value, which is more accurate than the linear regression method. Therefore, the GA-SVM model can forecast the theoretical and accessible yield better. Since the GA-SVM model is faster and more accurate, it should be used as a new method to calculate agricultural land productivity. Based on the theoretical and accessible yields calculated by GA-SVM model, the theoretical and accessible productivities can be enumerated with the theoretical and accessible yield multiplying with the total area of classification units, while the accuracy of agricultural land productivity can be guaranteed.