Recent, PM2.5 generated by various causes is spreading over a large area. As a result, the technique of estimating the concentration of PM2.5 has become important. Many studies have been conducted to estimate the PM2.5 concentrations. Recently research using learning-based models has been actively conducted. These learning-based models have exceeded the accuracy of existing statistical models. Existing research has a limitation that only data of the monitoring site can be used. In this paper, we propose a method to improve the accuracy of PM2.5 concentration estimation using the information of the surrounding area around the PM2.5 monitoring site. To do this, we study the difference in performance between the 3-dimensional input data and the existing vector data. The algorithm used in this experiment is CNN which can efficiently use the correlation of spatial information. Experimental result show that using large spatial data gives high accuracy of PM2.5 concentration estimation. ⓒ 2017 DGIST