The color characteristics and their variation rules of landscape plants are used as scientific basis for the seasonal analysis and design of landscape plants. However, with the disadvantages of heavy workload and low accuracy, the traditional quantitative analysis results are easily affected by subjective factors. In order to solve the problems above, based on deep learning network framework of UNet++, an improved image segmentation model was proposed, adding a new encoder embedded with an attention module and dilated convolution to the UNet++ network so as to enhance the capture of detail plant information. After extracting the color feature components of the segmented plant images, the Relief algorithm was used to filter these 12 color features. The effectiveness of the improved model was verified on the established landscape plant dataset. The experimental results show that the accuracy of the segmentation results is 97.8%, and the ‘a’ channel feature in the Lab color space can be used as the most discriminative color index to measure the seasonal changes of landscape plants. The improved model and the feature selection method can provide technical support for the research of the seasonal changes of landscape plants and the acquisition of the characteristics of agricultural crops.