Meta-learning based on parameter transfer for few-shot classification of remote sensing scenes
Chenhui Ma, Xiaodong Mu, Peng Zhao, Xin Yan
Abstract
Meta-learning is an effective way to deal with the few-shot problem that only very few annotated data samples are available for training the model. In remote sensing scene classification, previous methods with meta-learning often apply a shallow neural network due to lacking sufficient annotated training data that can support the larger number of parameters of a deep neural network (DNN). However, it is a common phenomenon that a DNN can bring more improvement than a shallow neural network. To this end, we propose a meta-learning method based on parameter transfer. Specifically, we firstly apply parameter transfer to fix the parameters in a DNN to relax the problem of training the large number of parameters. In detail, two variants of parameter transfer are designed to adapt to two different conditions about available annotated data samples, including applying supervised learning for normal meta-training and applying unsupervised domain adaptation for even less meta-training data. Then, the model uses a simple trimming operation to make a minor overhaul on the fixed parameters, which avoids updating the large number of parameters but obtains significant improvement. We perform experiments on three common datasets to demonstrate the improved accuracies and advantages of the feature extraction capability of our method.