Land Classification Research Based on the Improved Deeplabv3+ Model and the AI Earth Remote Sensing Cloud Platform

Complex geospatial object classification plays a pivotal role in resource management and planning, environmental conservation, and related fields. Deep learning has consistently emerged as one of the premier avenues for processing remote sensing imagery. Existing deep learning models, however, are beset with challenges such as data acquisition difficulties, intricate training processes, and suboptimal accuracy. In this paper, we address the aforementioned issues by proposing a method that integrates deep learning (DL) with a remote sensing cloud platform for the processing and creation of datasets. Leveraging the foundational DeepLabv3+, HRNet, PSPNet, and U-net models provided by the remote sensing cloud platform, we conduct online experimental training and analysis, thereby simplifying the dataset production and comparative experimentation. Furthermore, we have designed an improved DeepLabv3+ model to enhance the identification accuracy of deep learning models. By incorporating the Convolutional Block Attention Module (CBAM) into the base DeepLabv3+ model and replacing the Atrous Spatial Pyramid Pooling (ASPP) module with a Dense Atrous Spatial Pyramid Pooling module, we have effectively augmented the model’s precision. The experimental results show that compared with the original model, the MIOU (Mean Intersection over Union) index of this paper’s model has increased by 2.51%-2.90%, the MPA (Mean Pixel Accuracy) index has increased by 0.97%-2.54%, and the Precision index has increased by 1.03%-2.68%, proving that this paper’s method also has a significant improvement in segmentation effect.