Algebra models and machine learning are commonly used remote sensing methods in traditional land cover change detection studies. However, these traditional methods often fail to achieve accurate change detection on complex landscape, especially in detecting areas where changes have occurred, where a significant number of missed detections tend to occur. To more precisely classify land cover types and detect land cover changes, this study integrated index models and machine learning, proposing a change detection algorithm that combines spectral gradient difference information with random forest classification technology. This modified spectral gradient difference (MSGD) change detection method inherits the advantage of spectral gradient difference (SGD) that com-presses noise information and the strong ability of big data analysis. More importantly, it optimizes the completeness of the SGD change information through the branching structure of the tree model, achieving the description of both change intensity and change direction attributes. The results indicate that the proposed MSGD method detects land cover change with an overall accuracy of 96.13%, an misdetection rate of 0.94% and a Kappa coefficient of 0.65. Compared with traditional methods such as change vector analysis (CVA) and the original SGD method, the MSGD model reduced the missed detection rate of land cover change detection results by more than 50% and improved the Kappa coefficient by more than 25%. Therefore, the MSGD change detection results are more accurate, and it has more potential in megacity land cover change detection application.