Application of land-use inventory data and random forest models for estimating population densities in rural areas [隨機森林法及國土利用調查圖資於山村人口密度預測之應用]

In regions frequently affected by natural disasters, risk assessment and identification of vulnerable areas are critical to the management and security of the population. However, it is difficult for traditional choropleth maps to meet the requirements of fine-scale and actual population estimates. This study developed 2 random forest models named the Town-Village model and Forest model to explore correlations between population densities and land-use patterns in Taiwan for the development of dasymetric mapping approaches. A simple linear regression of observations versus predictions was performed, and the mean squared error between observations and predictions was calculated to evaluate the performance of the 2 models. In addition, prediction error rates were calculated to evaluate the prediction accuracy of the models ranging from inhabitantsparse areas to densely populated cities. Model evaluations revealed that both the Town-Village and Forest models had sufficient predictive abilities with mean absolute error rates of < 10% in areas with population densities ranging 51~25,000 people km-2. The models were applied to estimate population densities for all 261 villages in Nantou County, central Taiwan. Results revealed that in Nantou County, the Forest model exhibited a lower prediction error in inhabitant-sparse areas (with densities of < 100 people km-2) and could precisely estimate population densities under error rates of 0.34~7.31% in villages with an actual density of fewer than 5000 people km-2. The Forest model could be a more-applicable and better-suited approach for socioeconomic and policy initiative studies in mountain villages. ? 2021 Taiwan Forestry Research Institute. All rights reserved.