Human impressions of landscapes are considered to be influenced not only by the Green View Index (GVI), but also by the composition of greenery. While GVI can be automatically calculated using AI, green composition has been classified manually. In this study, we developed a model to automatically classify green composition using deep learning with two types of CNNs, VGG16 and ResNet50. The results showed that the model based on VGG16 achieved the highest classification accuracy. Additionally, to clarify the effects of GVI and green composition on human impressions, we conducted a visual evaluation experiment with human participants. The results showed that images with a GVI of 18% or more and greenery positioned in the center or on both sides of the image received higher evaluations. These findings provide fundamental insights for developing a green landscape evaluation system capable of estimating impressions based on image features.