Analysis and prediction of occurrence of non-croplands by machine learning methods in Ayabe City, Kyoto

Abstract

Recently, the increasing in the rate of abandoned cultivated land has resulted in a marked decrease of farmlands. Farmlands have various functions and play an important role in our lives. Therefore, it is necessary to promote conservation of farmlands. Moreover, although non-croplands don’t fall under the purview of abandoned cultivated land in the statistical survey “Census of Agriculture and Forestry”, non-croplands to an extent have become part of abandoned cultivated land or are most likely to be so in the near future. In this study, we analyzed the occurrence of non-croplands in Ayabe City, Kyoto, Japan. The purpose of this study was to analyze the current state and factors influencing the occurrence of non-croplands, and construct a high accuracy model for predicting the occurrence of non-croplands. For analysis of the factors mentioned above, we constructed a random forest prediction model and calculated the variable importance score (random forest is one of the machine learning algorithms). Besides, we verified the accuracy of the model by using cross-validation. The analysis results demonstrated that non-croplands have increased year upon year in Ayabe City. Also, it was suggested that the type of farmer, distance to densely inhabited district (DID), and farmland leasing were significant and important factors influencing occurrence of non-croplands. Further, as a result of cross-validation, the average accuracy of the random forest model, which consisted of 14 explanatory variables, was 97.4%. It was demonstrated that our model could predict the occurrence of non-croplands with high accuracy.