Abstract: With the increasing maturity of digital imaging technology and the increasing popularity of high resolution camera equipment, the advantages of high resolution and low cost have prompted the use of digital imaging technology to conduct more qualitative and quantitative descriptions of phenotypic traits for plant appearance. The RGB model is the most commonly used color representation for digital images. In order to explore the feasibility of using color gradation distribution parameters of the RGB model in soybean yield prediction, and to verify the universality of the method in different fertilizer operations and varieties, two soybean varieties, Qujing and Xudou 18, were selected to design field fissure experiments with different densities and nitrogen fertilizer levels in this study. Digital cameras were carried by Unmanned Aerial Vehicle (UAV) to collect soybean canopy digital images during three important reproductive growth stages. The results showed that the cumulative distribution of canopy color gradation of soybean at the florescence, pod-setting and grain-filling stages, all conformed to the skewed distribution, and a total of 20 Color Gradation Skewness-Distribution (CGSD) parameters were obtained by skew analysis. These parameters simultaneously described the shade and distribution of the canopy leaf color. The 20 CGSD parameters were significantly different among the florescence, pod-setting and grain-filling stages. And the variation trend of color depth parameters (mean, median, and mode) was opposite to that of the distribution parameters (skewness and kurtosis). The prediction model of soybean yield by using prediction model multiple stepwise regression method was constructed based on CGSD parameters with P value of 0.012. The model had high estimation accuracy in both the modeling group and the verification groups. The prediction accuracy of the model in modeling group reached 91.30% on average; the average prediction accuracy of 18 plots in the nitrogen operation research validation group was 87.33%. Although the prediction accuracy of the validation group for different varieties was lower than that of the modeling group and the validation group for nitrogen fertilizer operation research, it was also close to 80%. In conclusion, the RGB color model based on skewness distribution provided detailed soybean canopy image information, and the canopy color information quantitatively described systematically from the degree of depth, distribution bias and uniformity. And thus the yield prediction model based on CGSD parameters had high prediction accuracy, which can be widely used to predict yield of soybean grown in different production conditions. At the same time, the use of UAV and digital cameras improves the efficiency of image acquisition, while reduces the cost of image acquisition, which is more conducive to the popularization and application of this method.