A novel labor-free method for isolating crop leaf pixels from RGB imagery: Generating labels via a topological strategy

Crop leaf pixel isolation is a prerequisite step for conducting high-throughput crop phenotype surveys at a large scale. Automated isolation of leaf pixels from RGB imagery has been a challenging task, and the mechanism is not clearly understood. This study proposed a labor-free and automated method for isolating crop leaf pixels from RGB imagery, in which the strategy of labeling pure leaf pixels and the random forests (RF) algorithm were fused. First, a novel strategy for automatically labeling partial pure pixels of the leaf and background from RGB imagery was created. That strategy was depicted in the form of topology-based features, which originate from the CIVE-ExG space. Then, a machine learning method of random forest was developed for isolating leaf pixels based on labeled pure pixels. Subsequently, the leaf foreground images were refined using the adaptive active contours without edges (ACE) algorithm. Finally, the performance of the proposed method was evaluated with the reference leaf foreground image in indoor, natural, and public datasets and further compared with the benchmark leaf pixel extraction approaches. Simultaneously, the mechanism of leaf pixel isolation and partial pure pixel labeling was revealed. The proposed method is capable of isolating crop leaf pixels in various landscapes, with the mean of precision, recall, Dice, Manhattan, and Jaccard up to 0.9799, 0.9177, 0.9416, 0.9416, and 0.9007 over different datasets, respectively. The performance is comparable to or exceeds that of the latest methods. In addition, compared with the benchmark Otsu, local adaptive threshold segmentation (ATS), and marked watershed methods, the proposed method showed improvements in selected evaluation indicators, ranging from 9% to 310% higher. In conclusion, the proposed method holds great potential for extracting crop leaf pixels from multisource RGB imagery captured by various platforms, such as cameras, smartphones, or unmanned aerial vehicle platforms.