Illumination invariant segmentation of vegetation for time series wheat images based on decision tree model

Effective and efficient segmentation of vegetation from digital plant images is an actively studied topic in crop phenotyping. Many of the formerly proposed methods showed good performance in the extraction under controlled light conditions but it is still hard to properly extract only vegetation from RGB images taken under natural light condition where the images can contain shadowed and lighted parts with specularly reflected parts of plants. In this paper, we propose a robust method to extract vegetation from the plant images taken under natural light conditions using wheat images. The method is based on a machine learning process, decision tree and image noise reduction filters. We adopted the CART algorithm to create a decision tree in the training process and examined its performance using test images, comparing it with the performances of other methods such as ExG, ExG-ExR and Modified ExG which are widely used recently. The results showed that the accuracy of the vegetation extraction by the proposed method was significantly better than that of the other methods particularly for the images which include strongly shadowed and specularly reflected parts. The proposed method also has an advantage that the same model can be applied to different images without requiring a threshold adjustment for each image.