Abstract: Spatial patterns and temporal variation of arable land resources can greatly contribute to regional protection and sustainable utilization, particularly for food security and social stability in the Qinghai-Tibet Plateau. A spatial sampling with fewer points is required to comprehensively characterize the overall features of regional arable land. Long-term observation is also of great significance in evaluating the quality of arable land resources. In this study, a data-driven spatial sampling was presented to determine the indicators of arable land quality in the Tu Autonomous Country of Huzhu, Qinghai Province, China. The grid units of the sample population were set as 1 km of arable land, with a total of 790 points. 21 indicators were extracted from three dimensions of topographic features, soil properties, and tillage technical conditions. An indicator system was then constructed for the sampling of arable land at the point scale, such as the slope, soil bulk density, organ carbon content, and agriculture mechanization level. The performance of spatial sampling was quantified by the natural quality, technical level, and arable land productivity index, which were collected from the pilot project of the Ministry of Natural Resources of China. Local spatial heterogeneity was represented to simulate the accuracy of the overall quality of arable land. Multiple indicators were also calculated from five dimensions of topographic features, soil properties, tillage technical conditions, environmental conditions and biological characteristics, in order to evaluate the quality and productivity of arable land. RSM (random sampling), SPCOSA (spatial coverage sampling and random sampling), CLHS (conditioned Latin hypercube sampling), XY_CLHS (CLHS with x and y coordinates as covariates), and SPCOSA_CLHS (spatial coverage sampling and random sampling-conditioned Latin hypercube sampling) were compared from the perspectives of information entropy, Kullback–Leibler divergence, similarity distance, the spatial heterogeneity and distribution uniformity of samples in the overall quality of arable land. A survey of observation points was carried out on the indicators of arable land quality in the county-level areas. Entropy-based tests were also performed on each sampling using the nine groups with the multigroup sample sizes: 10, 20, 30, 40, 50, 75, 100, 150, and 200. A suitable number of points were explored in the study area. The results show that the SPCOSA_CLHS model integrated the SPCOSA into the CLHS model, and then represented the spatial heterogeneity with the lower spatial constraints. A better performance was achieved to express the overall index attribute and spatial heterogeneity of arable land quality. When the number of sampling points was between 100 and 200, five sampling models shared similar applicability for the attribute in the overall quality indicators of arable land. Once the sample size dropped below 100, there were the greatest differences among sampling models, where the SPCOSA and RSM offset most. A better performance of SPCOSA_ CLHS was also obtained in the information entropy, KL divergence, and similarity distance, in terms of convergence and stability. When the number of sample points was 40-50, a similar spatial sampling was observed as the sample size of 100-200. Therefore, SPCOSA_CLHS can be expected to describe the spatial heterogeneity of arable land quality indicators, the spatial uniformity of sample point distribution, and the simulated accuracy of the overall quality. This finding can provide strong support to the survey and monitoring of arable land quality in the Qinghai-Tibet Plateau. In turn, the evolution of arable land quality can also be used to explore the sustainable use pathways of arable land.