A Fully Integrated Deep Learning Framework for Semantic Segmentation of Vegetation Classification Based on Active Learning Strategies and UAV Remote Sensing

With the development of urbanization and the increasing use of land for construction, there is a growing focus on detecting different types of surface vegetation cover. Unmanned Aerial Vehicle (UAV) remote sensing is an important method for acquiring high-resolution multispectral images of surface vegetation, and its data provides reliable support for deep learning. However, labeling training samples requires researchers to conduct field surveys, which is a time-consuming and expensive process. In order to reduce the cost of labeling, this study proposes a deep learning framework for fully fusing semantic segmentation for vegetation classification, which is built based on an active learning strategy. The framework utilizes the Gated Full Fusion module (GFF) and the attention mechanism to enhance the model’s learning ability. It dynamically extracts more important samples to be labeled by the Oracle, reducing the number of sample markers while ensuring accuracy. We conducted several experiments in areas with real urban vegetation coverage, and the results demonstrate that the framework proposed in this study surpasses the current state-of-the-art baselines. Additionally, it ensures model stability while reducing the cost of labeling a large number of samples.