The extraction method of cotton coverage was studied based on the difference of vegetation and non-vegetation pixels of RGB images from hand-held camera and unmanned aerial vehicle (UAV) remote sensing in different color spaces. Under different lighting conditions, totally 29 high-resolution (0.4mm) RGB images of cotton in seedling and bud stage were obtained by hand-held digital camera. The recognition abilities of cotton in Lab (a), RGB (2G-R-B) and HIS (H) color spaces were compared and analyzed. Two threshold classification threshold getting methods, dynamic threshold and fixed threshold were used to extract cotton coverage. The dynamic thresholds were determined by the intersection of the Gaussian distributions of vegetation and non-vegetation pixels. The fixed thresholds were set as the mean values of dynamic thresholds in the three color spaces, respectively. The results showed that vegetation and nonvegetation pixels obeyed Gaussian distribution in a, 2G-R-B, and H color spaces, which could be fitted by using nonlinear least-squares algorithm. The distribution range of dynamic classification thresholds was relatively concentrated, and their mean values of -3.78, 0.06 and 0.13 could be set as fixed classification thresholds. Compared with 2G-R-B and H, the a color space had the best ability to identify green vegetation and was more suitable for extracting cotton vegetation coverage. Compared with dynamic threshold, the extraction accuracy based on fixed threshold was better and the average extraction error was 0.0094. It can also accurately extract fractional vegetation cover (FVC) from UAV images captured under different light conditions (sunny and cloudy) with different soil moistures. After preliminary tests and analysis, it was believed that based on the differences of vegetation and non-vegetation pixels in Lab (a) color space, combining with a fixed classification threshold of -3.78,cotton coverage in seedling and bud stage could be accurately extracted under different light conditions.