Study and comparison of color models for automatic image analysis in irrigation management applications

Abstract

Image processing and computer vision are increasingly being used in water management applications in agriculture. Images can provide valuable information on the percentage of ground cover, which is essential in determining crop irrigation needs. Techniques based on color analysis allow classifying accurately and efficiently soil/plant regions in the images. Many color spaces have been proposed, among them: RGB, rgb, XYZ, L*a*b*, L*u*v*, HSV, HLS, YCrCb, YUV, I1I2I3 and TSL. Different possibilities to model the probability distribution of a given color class appear for each space; one of the most widespread non-parametric methods is modeling using histograms. This presents various alternatives in order to represent a color class: the number of channels, which channels to use, and the size of histograms. Using a wide and varied set of images of lettuce crops (Lactuca sativa)—previously classified manually in soil and plant pixels—a comprehensive analysis and comparison of the proposed color models has been conducted for the soil/plant classification problem. The experimental results demonstrate the superiority of models that separate luminance from chrominance. In particular, L*a*b* provides the best results with a* channel, producing a 99.2% of correct classification. Further processing stages improve this performance up to 99.5% accuracy, taking less than 1/3 of a second per image in a normal laptop. These results can be applied to reduce water consumption by optimizing the accuracy and efficiency of automatic image analysis of crops.