Abstract: Objectives:  The spectral characteristics of supraglacial debris are extremely similar to those of rocks, which is difficult to retrieve by semi-automatic or automatic interpretation of remote sensing. Obtaining the extent of supraglacial debris and glacial surface flow velocity can help to understand the characteristics of glacier mass balance about debris-covered glaciers.   Methods:  The Random Forest method based on feature optimizatio is used to identify the supraglacial debris of the Koxkar Glacier. To increase the distinguishability between supraglacial debris and rocks, sentinel-2 images, remote sensing indexes, terrain features and texture features are also added for image classification. Glacial flow velocity is estimated using the Coregistration of Optical Sensed Images and Correlation (COSI-Corr) method, which is considered to be one of the most effective methods for small-scale and high-precision estimation of glacial flow velocity.   Results:  Compared with other machine learning classification methods, the Random Forest method based on feature optimization can effectively avoid the misclassification between supraglacial debris and rock. The result of identification shows that the area of debris is about 24.6 km², accounting for 31.7% of the total area of Koxkar Glacier. The results of glacial flow velocity show that the maximum average annual flow velocity occurs at about 4380 m above sea level in the eastern branch and can reach 145.9 m·a^-1 in 2020. The flow velocity reduces to 0~10 m·a^-1 with the decreasing elevation. The flow velocity during the ablation period is much higher than the annual mean velocity. The differences of velocity on the glacier tongue are induced by differences of altitude, ice lake outburst, collapse of Subglacial meltwater channeland and other factors. The changes of flow velocity in the study area from 1989 to 2021 shows that the flow velocity in the upper and middle part of the glacier continues to accelerate with the increase of surface temperature. In the range from 3090 m to 3500 m above sea level located in the lower part of the glacier, the flow velocity decreases continuously due to the accumulation of debris and the continuous thinning of glacier tongue.   Conclusion:  (1) Compared with other machine learning classification methods, the accuracy of the Random Forest method based on feature optimization has been significantly improved in identification of supraglacial debris. (2) In the past 30 years, the overall flow velocity of debris-covered Koxkar Glacier is slowly increasing with obviously spatio-temporal differences, and the glacier mass balance of the glacier below 3500 m above sea level is in a continuous deficit state.