Research on the index system of cultivated land quality grading based on random forest algorithm

Objective  To analyze the difference of cultivated land quality in the study region, optimize the use and layout of cultivated land, and provide a reference for cultivated land protection.

Method  Taking the cultivated land in Gonghe County, Dulan County and Wulan County in Qinghai Province as the research object, the influencing factors of cultivated land quality were collected based on the history and existing literature, and the random forest algorithm and correlation analysis were used to screen the grading indicators and confirm the weight. We calculated the grading index and divided the levels by weighted sum method to get the grading result. We compared the results with the grading results of commonly used Delphi method.

Result  The value of variable importance I obtained by random forest algorithm ranged from 0.03 to 11.94. Correlation analysis showed that the correlation between most influencing factors was not significant, eight of which were significant correlation. The 14 rating indicators under four dimensions were astringed from 30 influencing factors. The main factors influencing the quality of cultivated land in the study area were ecosystem vulnerability, mean precipitation of growing season and annual solar radiation amount, with the weights of 0.11, 0.10 and 0.09, respectively.

Conclusion  Compared with Delphi method, the random forest algorithm has better stability and smaller level of index variation interval, which is more conducive to construct a comparable sequence of cultivated land levels at provincial spatial scale.