Abstract: A lychee (also called litchi) is one of the most favorite fruits in Asian areas. An accurate extraction of the plantation area is crucial to lychee production. Traditional remote sensing has failed to accurately measure large-scale plantation areas, due to limited time and space. It is often required for the extraction of the plantation area from the high-resolution images using deep learning. However, the existing deep learning cannot fully meet the requirement to extract the plantation area of the lychee tree. Specifically, the low accuracy of the feature extraction has confined to the segmentation of the original image, due to the long-time sample training and the high complexity of the original model. In this study, an improved DeepLabV3+ model was proposed to replace the original DeepLabV3+ backbone network Xception with the MobileNetV2. A better performance was achieved to effectively alleviate the overfitting in the loss function convergence. The high accuracy of the image segmentation with the time saving was also improved to extract the features of the lychee tree; The ASPP (Atrous Spatial Pyramid Pooling) module was also improved to the DenseASPP. A dense connectivity mechanism was also introduced using ASPP; The global information of the image was more effectively utilized to recognize the complex scenes when performing the Lychee Tree Multi-scale Feature Extraction. A null convolution also varied to connect the outputs of each layer with all the previous layers. A denser feature map was then formed after introduction; The Channel Attention Mechanism (Squeeze-and-Excitation) was used to better capture the semantic information of the lychee tree during semantic segmentation. The high accuracy of the segmentation was obtained in the edge region and the details; The Strip Pooling (SP) was introduced to enhance the spatial context understanding of the model, especially in the scenes with complex objects and backgrounds. The multi-scale pooling was utilized to capture more semantic information, In turn, the intersection and integration ratio (IoU) were computed to effectively suppress the interference for high accuracy; A systematic comparison was also carried out on the improved DeepLabV3+ model with the image segmentation of the SegFormer, PSPNet, and UNet models. The results show that the mean intersection over union (MIoU), the mean pixel accuracy (mPA), and the accuracy (Ac) of the improved model reached 83.55%, 91.58%, and 91.15%, respectively. There were 8.15, 5.27, and 4.97 percentage points higher than those of the original DeepLabV3+ model, respectively. The number of parameters was 5.8M after structural optimization. The computational complexity was reduced to 22.4 GFLOPs, which was 94% and 95% lower than the original DeepLabV3+ and the PSPNet, respectively. The improved mode can be expected to extract the plantation area of the lychee tree. The finding can also provide technical support to accurately extract the plantation area distribution of the lychee tree in the trend of spatial variation.