An Object-Based Extremely Weak Supervised Learning Paradigm: Focusing on the Introduction of Algorithm

Although deep learning-based algorithms have achieved satisfactory classification performance, they require a large amount of annotated training data. To overcome this issue, we proposed an extremely-weak supervised learning algorithm for remote sensing (RS) classification tasks, which only requires several carefully chosen ground-feature sample labels as input. To be specific, our method consists of three subnetworks working in tandem and under the object-based framework. The first one produces pseudo labels by using two unsupervised autoencoders with shared weights, the second one attempts to optimize the pseudo labels through noisy learning and post-processing, and the third one involves running inference with a powerful residual network for labeling the unseen data. To the extent of our information, we are the first to introduce in the RS classification task the autoencoder and noisy learning modules, which minimized the data acquisition and human labeling efforts. Due to the space limitations, this article focuses only on the algorithms and theory of computation. Readers are referred to the companion piece for detailed case studies.