Hemisphere photos are now widely applied to provide information about solar radiation dynamics, canopy structure and their contribution to biophysical processes, plant productivity and ecosystem properties. The present study aims to improve the original 'edge detection' method for binary classification between sky and canopy, which works not well for closed canopies. We supposed such inaccuracy probably is due to the influence of sky pixels on their neighbor canopy pixels. Here, we introduced a new term 'neighbor distance', defined as the distance between pixels participated in the calculation of contrast at the edges between classified canopy and sky, into the 'edge detection' method. We showed that choosing a suitable neighbor distance for a photo with a specific gap fraction can significantly improve the accuracy of the original 'edge detection' method. We developed an ND-IS (Neighbor Distance-Iteration Selection) method that can automatically determine the threshold values of hemisphere photos with high accuracy and reproductivity. It combines the modified 'edge detection' method and an iterative selection method, with the aid of an empirical power function for the relationship between neighbor distance and manually verified gap fraction. This procedure worked well throughout a broad range of gap fractions (0.019-0.945) with different canopy compositions and structures, in five forest biomes along a broad gradient of latitude and longitude across Eastern China. Our results highlight the necessity of integrating neighbor distance into the original 'edge detection' algorithm. The advantages and limitations of the method, and the application of the method in the field were also discussed.