Abstract: The task of grading oil palm bunches by ripeness poses a number of significant challenges for computer vision. The small difference in hue between ripe and unripe bunches means that colour-based models are susceptible to errors when presented with images shot in novel lighting conditions. In this paper, we investigate the effectiveness and performance characteristics of coloured edge maps when used as an input feature to a Convolutional Neural Network (CNN) by comparing the Laplacian of Gaussian, Sobel, Prewitt, and Kirsch edge extraction techniques. We show that under normal lighting conditions, coloured edge maps are able to match the performance of fully-coloured images. More notably, they significantly outperform fully-coloured images when variance to lighting is applied. When images are darkened or brightened, classification accuracy for fully-coloured images drops by 19.89% vs only 4.97% on average for the coloured edge map methods tested. This is of major benefit in commercial applications, where images are often captured by a multitude of devices under different lighting conditions, leading to potentially unreliable performance when fully-coloured images are used.