Semi-supervised knowledge distillation framework for global-scale urban man-made object remote sensing mapping

Abstract

Accurate mapping of global urban man-made objects such as buildings and roads is critical for monitoring urbanization. Remote sensing imagery provides a cost-effective way of mapping these objects, but the challenge of “knowledge forgetting” arises due to urban diversity and the continuous growth of global samples. Although the existing knowledge distillation approaches can transfer knowledge from a larger teacher model to a smaller student model by distilling the knowledge learned from reliable labels, they fail to work for global-scale mapping, which lies in two aspects: low-quality labeling and fixed-size models. In this paper, we propose GUMONet, which is a semi-supervised knowledge distillation framework for global-scale urban man-made object mapping. For the first phase, a label diversity progressive learning module is introduced for generating high-quality labels in a semi-supervised manner. Label diversity is used to measure the diverse urban patterns based on spatial-semantic uncertainty, where the diversified labels clustered in object boundaries and heterogeneous areas are attributed to high spatial uncertainty and semantic uncertainty, respectively. Based on the label diversity, the model decision boundary is progressively determined from coarse to fine. Specifically, at the early stage, instances away from the decision boundary are selected to ensure the stability of the model training. As the iteration progresses, instances close to the decision boundary are associated with a higher probability of further enhancing the quality of the uncertain labels by hard sample mining. For the second phase, a size-variable knowledge distillation module is adopted to optimize the data-model matching process. This module consists of a noise teacher model that prevents overfitting by injecting noise perturbations to increase the data distribution complexity and a size-variable student model that avoids underfitting by dynamically adjusting its size with the growth of global samples. We applied GUMONet to six study areas across four continents, with data from different sensors, achieving an 18.97% improvement in intersection over union, compared with the previous methods. Our results also demonstrate a positive correlation between urban development and urban diversity, with a correlation coefficient of 0.749. As urban development progresses, urban diversity stabilizes and building transformation becomes the primary means of promoting further development.