Aiming at the problems of incomplete recognition of forest land boundary area and low accuracy of small forest land segmentation in remote sensing image segmentation by ordinary convolutional neural network, an improved method of DeeplabV3+ network based on transformer and attention mechanism is passively proposed. First, the transformer mechanism is introduced in the encoding stage, and the hole convolution operation in the original pooling pyramid is replaced by a transformer operation that can obtain more context information, thereby improving the network's ability to extract forest boundary information; then, the attention mechanism is introduced. Go to the decoding part of the network to improve the model's ability to detect small forests. Experiments show that the average intersection-over-union ratio (MIou) of the improved method can reach 81.83%, which is 1.25% higher than the original DeepLabV3+ network model. The method fully considers the extraction of forest edge information and the attention to small targets in satellite remote sensing image segmentation, and the improved method proposed can improve the accuracy of forest land extraction in remote sensing images.