The large-scale agile earth observation satellite scheduling problem is a complex combinatorial optimization problem. Existing approaches typically address this by adjusting the sequence of observation missions, which can lead to a limited solution space and restricted optimization potential. To overcome these limitations, we propose a Window Decision Network. It enhances the accuracy of the problem state representation by integrating both the static and dynamic attributes of the observation process. Additionally, we developed a visible time window update algorithm to guide the construction of the solution sequence, expanding the solution space by adjusting the window sequence directly. Experimental results demonstrate that this method outperforms the state-of-the-art approaches, increasing scheduling profit by up to 23.163%.