Variables and a Validation Data Analysis to Improve the Prehistoric Cultivated Land Predictive Precision of Yulin, Northern Shaanxi, China

The distribution of cultivated land in prehistoric times was primarily influenced by natural conditions and population density. This article presents a case study on modern cultivated land simulation to analyze the potential impact of variable selection and validation data accuracy on model precision. Additionally, methods were explored to enhance the accuracy of prehistoric cultivated land simulation. Seven natural variables and one settlement density variable were selected to simulate the distribution of cultivated land based on a Binary Logistic Regression model. The simulated results were then compared with real land use data from 1985, which are commonly used as validation data for prehistoric farmland reconstruction. The findings revealed that all eight selected parameters could explain the distribution of cultivated land in the research area, with annual precipitation being the most influential factor. The initial prediction accuracy was relatively low at 65.8%, with a Kappa coefficient of 0.316. Several factors were identified as affecting the prediction accuracy. Firstly, the scale effect diminished the impact of the slope and elevation on cultivated land distribution, and errors were introduced in the method used to calculate the distance from residential areas. Secondly, the loess hilly area in the southeastern part of the research area overpredicted cultivated land due to insufficient data on actual residential land demand. Lastly, strong human activity since the 1950s has altered the natural distribution of cultivated land, resulting in poor consistency ratings. To address these issues, a batch modification method was employed to correct the 1985 data. The validation of the prediction model using the corrected data demonstrated a significant improvement in accuracy. Therefore, it is recommended to use the revised 1985 land use data for verifying prehistoric cultivated land simulation in the region. However, further research is required to mitigate the impact of the first two errors.