Small sample remote sensing scene classification method based on meta-kernel network

Abstract

The invention discloses a small sample remote sensing scene classification method based on a meta-kernel network. The method comprises the following steps: selecting a given non-overlapping remote sensing data set; for each task, embedding a given support set into a vector; a parameterized linear classifier is adopted to classify the feature vectors obtained through embedding; secondly mapping the features in the metric space to a high-dimensional space by using nonlinear mapping, and expanding the spatial resolution around the surface of the separation boundary; using a stretching loss function to constrain an intra-class and inter-class variance ratio, and forcing intra-class feature aggregation; and performing scene classification on the small sample remote sensing images by using the final optimal solution. According to the method, the limitation of a fixed distance on the model is relieved by utilizing specific task information; the features in the embedded space are secondarily mapped into a high-dimensional space, the classification boundary definition is improved, and the classification effect is enhanced; the dependency of the category boundary on sample selection is reduced, the robustness of the classification boundary is increased, and the model classification effect is improved.

Description

technical field [0001] The invention belongs to the technical field of remote sensing image recognition, and in particular relates to a small-sample remote sensing scene classification method based on a meta-kernel network. Background technique [0002] Scene classification is a fundamental and important problem in the field of remote sensing image interpretation, which aims to assign a specific semantic label to each unlabeled remote sensing image according to its semantic content. This task has a wide range of applications, including disaster detection, housing planning, environmental monitoring, land resource management, etc. Few-shot remote sensing scene classification attempts to quickly adapt the model to new scenes that do not appear in a closed training set, with only a few labeled examples for each new scene. On the one hand, it is a challenge for the model to learn good generalization features due to limited samples that are difficult to describe the distribution ...

AI Technical Summary

Problems solved by technology

At this time, the classification boundary formed by sampling a data point will often misclassify large areas, thereby affecting the overall classification accuracy.

Images