As the primary photosynthetic organ of plants, the leaf is essential for almost all crops. Extracting pure leaf pixels is a prerequisite for estimating leaf physiological parameters or plant disease by using remote sensing images accurately. Therefore, identifying crop leaf pixels accurately, efficiently, and automatically from images are significant for the research of plant phenomics. Unfortunately, previous methods usually were developed from the view of computer vision with the process of having insight into leaf spectral characters abandoned, which is harmful to extract leaf pixels from hyperspectral or multispectral images, so the existing method is poor in these images. An automated method of leaf pixels extraction for hyperspectral or multispectral images was proposed by exploring the internal mechanism of crop leaf pixels identification. After spectral feature compression and conversion for measured hyperspectral and simulated multispectral images, this method performed local adaptive threshold segmentation (ATS) and Canny edge detector (Canny), respectively, so that the advantages of the selected two algorithms were integrated. Following all of this, a novel gradient-based breakpoint connection algorithm was applied. Eventually, an automated crop leaf pixels identification method was developed. The results demonstrated that EVI was superior in enhancing the spectral signatures of the crop leaves. Additionally, NDVI also can strengthen the leaves features, but this ability was slightly worse than EVI. Furthermore, it was highlighted that the ability of vegetation indexes derived from red edges bands was limited in enhancing leaf spectral features. The proposed method can identify crop leaf pixels effectively, with all accuracy evaluation parameters up to above 0.94. Compared with OTSU, Marker-watershed, and other typical methods, the accuracy was remarkably improved with increases in all evaluation parameters by 29% to 98%, which was similar to the performance of a fully convolutional network (FCN) or random forests (RF) algorithms. However, when ignoring the time-consuming labeling collection activities of FCN and RF, the leaf identification efficiency was improved by about 73% than them. By the combination of feature conversion, ATS, and Canny detector, the crop leaf pixels can be identified accurately, efficiently, and automatically from hyperspectral or multispectral images, without labor-extensive and time-consuming labeling collection activities. The input images were arbitrary for the method so that it’s potential for estimating leaf physiological parameters or taking the place of the manual labeling activities in deep learning or supervision algorithms.