Green Crop Image Segmentation Based on Superpixel Blocks and Decision Tree

Abstract

When segmenting green crops, we usually use green indexes, such as Excess Green Index (ExG), Combined Indices 2 (COM2), Modified Excess Green Index (MExG) etc., which are regarded as efficient methods. However, they can’t extract green crop exactly under complex environmental conditions. Particularly, they can’t segment green crops from complex soil backgrounds, such as with high light or deep shadow areas in crop leaves. To address current deficiencies in green crop segmentation, this paper introduces a new crop segmentation method to extract more compete green crops. Firstly, a pre-processing procedure divides the crop images into superpixel blocks by using Simple Linear Iterative Clustering (SLIC) algorithm, then these superpixel blocks are classified into three class by using Classification And Regression Tree (CART) decision tree based on a seven-dimensional(7-D) features vector constructed in this paper: only crop blocks (OC-block), only background blocks (OB-blocks) and CBE-blocks within which crop and background both coexist. Finally, CBE-blocks are processed by making good use of the advantage of ExG to exact green crops in them. Experimental results show that the algorithm proposed in this paper can segment accurately the green crops from the soil backgrounds, even under relative complex field conditions with \( Accuracy \) of 98.44%, \( Precision \) of 91.75%, Recall of 91.78%, FPR of 0.55% and F1_score of 89.31%.