Research and Implementation of Landslide Labeling Method Oriented to Crowdsourcing Mode

Landslide identification is an important task in geological disaster prevention and reduction. Using deep learning models to identify landslides from remote sensing images is a prevalent method. However, annotation of landslide samples requires domain knowledge in geology, which makes it a difficult task. To cope with the issue, this paper researches and implements a crowdsourcing based landslide sample data annotation method. A fusion and evaluation algorithm for annotation results is proposed. The algorithm can integrate the annotation results of multiple annotators on the same landslide into a consensus curve to solve the problem of subjective biases in annotation results owing to differences in the perception of landslides among annotators. Meanwhile, in order to match annotators of different abilities with different tasks, a task recommendation algorithm which can assess difficulty of annotation tasks based on complexity of landslide contour curves and establish user profiles based on historical task records is proposed. The experimental results demonstrate that the method can effectively improve the annotation quality of landslide data.