Abstract:This paper proposes a method for unsupervised whole-image clustering of a target dataset of remote sensing scenes with no labels. The method consists of three main steps: (1) finetuning a pretrained deep neural network (DINOv2) on a labelled source remote sensing imagery dataset and using it to extract a feature vector from each image in the target dataset, (2) reducing the dimension of these deep features via manifold projection into a low-dimensional Euclidean space, and (3) clustering the embedded features using a Bayesian nonparametric technique to infer the number and membership of clusters simultaneously. The method takes advantage of heterogeneous transfer learning to cluster unseen data with different feature and label distributions. We demonstrate the performance of this approach outperforming state-of-the-art zero-shot classification methods on several remote sensing scene classification datasets.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to perform unsupervised clustering on remote - sensing images without labels. Specifically, the author proposes a method to perform unsupervised whole - image clustering on remote - sensing scene images in the target dataset through heterogeneous transfer learning. The target dataset has no labels, and many existing clustering methods require pre - specifying the number of clusters or neighborhoods, which is often difficult to achieve in practical applications, especially in applications where feature discovery is the goal. In addition, some current deep - clustering techniques need to train expensive deep neural network clustering layers for each target dataset, which increases the computational cost. To solve these problems, the author proposes a method consisting of three main steps: 1. **Fine - tuning the pre - trained deep neural network**: First, fine - tune the pre - trained deep neural network (such as DINOv2) on a labeled source remote - sensing image dataset, and then use this fine - tuned model to extract feature vectors from each image in the target dataset. 2. **Dimensionality reduction**: Embed these high - dimensional features into a low - dimensional Euclidean space through manifold projection techniques to preserve the structural relationships between features. 3. **Clustering**: Use a Bayesian non - parametric method (such as the Dirichlet process Gaussian mixture model DPGMM) to cluster the embedded features while inferring the number of clusters and membership relationships. The advantage of this method is that it does not require labeling of the target dataset and can adaptively determine the number of clusters, and is suitable for datasets with different feature and label distributions. Through experiments, the author has proven that the performance of this method on multiple remote - sensing scene classification datasets is better than that of existing zero - shot classification methods.

Deep Clustering of Remote Sensing Scenes through Heterogeneous Transfer Learning

Transfer Heterogeneous Unlabeled Data for Unsupervised Clustering.

Cross-domain Residual Deep NMF for Transfer Learning Between Different Hyperspectral Image Scenes.

Clustering-Based Representation Learning through Output Translation and Its Application to Remote--Sensing Images

Deep Learning-Based Crop Mapping in the Cloudy Season Using One-Shot Hyperspectral Satellite Imagery.

Scene Classification of High Resolution Remote Sensing Images Via Self-Paced Deep Learning.

Unsupervised remote sensing image classification with differentiable feature clustering by coupled transformer

Deep Ensemble Remote Sensing Scene Classification via Category Distribution Association

Scene Classification Based on Heterogeneous Features of Multi-Source Data

Transferring Deep Convolutional Neural Networks for the Scene Classification of High-Resolution Remote Sensing Imagery

Unsupervised Transfer Learning Using For Multi-Model Remote Sensing Data Classification

Deep Learning for Feature Extraction in Remote Sensing: A Case-Study of Aerial Scene Classification

Deep Clustering for Unsupervised Learning of Visual Features

Extracting feature fusion and co-saliency clusters using transfer learning techniques for improving remote sensing scene classification

Unsupervised Deep Feature Extraction for Remote Sensing Image Classification

Deep Updated Subspace Networks for Few-Shot Remote Sensing Scene Classification

Deep Clustering With Intra-class Distance Constraint for Hyperspectral Images

Unsupervised Transfer Learning with Self-Supervised Remedy

Land-cover classification with high-resolution remote sensing images using transferable deep models

Hierarchical Segmentation Of Remote Sensing Images By Unsupervised Deep Learning Features

Analysis of the inter-dataset representation ability of deep features for high spatial resolution remote sensing image scene classification