Aiming at the research problems of large-scale landscape patterns such as provinces, taking Guangdong Province as the research object, using Sentinel-2 remote sensing images with a resolution of 10m and their corresponding land use type data, taking landscape pattern index and remote sensing images as feature sets, and taking image types based on visual features as label sets, Adam optimization algorithm is used to construct a landscape visual feature classification model based on Keras deep learning framework. The feature set and the corresponding label set data are divided into training set and test set according to 8∶2, and cross-validated. The results show that the accuracy of the model on the training set and test set is 99.57% and 98.93%, respectively. The model can effectively correlate landscape pattern index and image visual features, and has strong generalization ability. It is suitable for remote sensing image classification tasks in large-scale regional landscape pattern research and township layout planning.