Abstract: Abstract: The characteristics and interactions of arable land quality components determine the external manifestation of arable land quality. It was the vital significance to objectively determine natural quality of arable land for arable land classification and grading. In this paper, advantage and disadvantage of the existing methods used in the arable land natural quality evaluation was analyzed, SOFM (self organizing feature map) neural network method was proposed to evaluate arable land natural quality basing on spatial database in Jiutai city of Jilin province. The nine evaluation indicators including surface soil texture, profile configuration, content of soil organic matter, soil pH value, barrier layer depth, soil salinity, effective soil depth, drainage condition and slope, were chosen and the corresponding database layer was established in the method. At the same time, attribute values were input and data were normalized. By training step for 1000, the system automatically generated 13 categories. Based on temperature potential productivity index and crop yield ratio of Jiutai city, evaluation index of arable land natural quality were calculated based on GIS. According the size of arable land natural quality indicators, the arable land natural quality of Jiutai city was divided into 3 grades. The ratios of Ⅰ, Ⅱ and Ⅲ grade of arable land nature quality which was classified with proposed method in this paper to the total arable land area in Jiutai city were 42.13%, 30.40%, 27.47%, respectively. The natural quality evaluation results of arable land based on SOFM neural network were compared with that with farmland natural quality grading of update results in Jiutai city. The comparison results indicated that in the nature quality grade Ⅰwith proposed method in this paper, the ratio of fifth, sixth and seventh grade which was classified with farmland natural quality grading of update results was 33.91%, 4.25%, 3.98%, respectively; in the nature quality grade Ⅱ with proposed method in this paper, the ratio of fifth, sixth and seventh grade which was classified with farmland natural quality grading of update results was 4.58%, 25.38%, 0.45%, respectively; in the nature quality grade Ⅲ with proposed method in this paper, the ratio of fifth, sixth and seventh grade which was classified with farmland natural quality grading of update results was 0.85%, 6.48%, 20.13%, respectively Quality grade Ⅰ, Ⅱ and Ⅲ, classified with SOFM neural network corresponded to fifth, sixth and seventh grade of national nature grades classified with farmland natural quality grading of update results, figure spot coincidence rate and area overlap rate were 80.78% and 79.42%, respectively. The seasons for the discrepancies in the results of the two kinds evaluation method were that the slope factor was considered in proposed method, at the same time, uniform weights in Jiutai city, Jilin province was used in arable land natural quality grading of update results and qualitative description of the two kinds of evaluation methods for some indicators were quantified through information weighting value method, due to different quantization methods, different assignments. In this method, by self-organizing neural network combining with geographic information systems based on MATLAB, effectively influencing factors of arable land quality that related to soil and soil environment were integrated and quantification and space for spatial and non-spatial datum was realized. Especially, weights were subjectively determined, which made evaluation results more objective and true in propose method. And this method had more advantages in evaluating for highly nonlinear relationship between the arable land quality and its affecting parameters. Overall, the study method provides a new idea for arable land quality evaluation and broadens depth and breadth of arable land quality evaluation, and enhances the results credibility of natural quality evaluation of arable land.