Land-Cover Classification with High-Resolution Remote Sensing Images Using Transferable Deep Models (1807.05713v3)

Published 16 Jul 2018 in cs.CV

Abstract: In recent years, large amount of high spatial-resolution remote sensing (HRRS) images are available for land-cover mapping. However, due to the complex information brought by the increased spatial resolution and the data disturbances caused by different conditions of image acquisition, it is often difficult to find an efficient method for achieving accurate land-cover classification with high-resolution and heterogeneous remote sensing images. In this paper, we propose a scheme to apply deep model obtained from labeled land-cover dataset to classify unlabeled HRRS images. The main idea is to rely on deep neural networks for presenting the contextual information contained in different types of land-covers and propose a pseudo-labeling and sample selection scheme for improving the transferability of deep models. More precisely, a deep Convolutional Neural Networks is first pre-trained with a well-annotated land-cover dataset, referred to as the source data. Then, given a target image with no labels, the pre-trained CNN model is utilized to classify the image in a patch-wise manner. The patches with high confidence are assigned with pseudo-labels and employed as the queries to retrieve related samples from the source data. The pseudo-labels confirmed with the retrieved results are regarded as supervised information for fine-tuning the pre-trained deep model. To obtain a pixel-wise land-cover classification with the target image, we rely on the fine-tuned CNN and develop a hybrid classification by combining patch-wise classification and hierarchical segmentation. In addition, we create a large-scale land-cover dataset containing 150 Gaofen-2 satellite images for CNN pre-training. Experiments on multi-source HRRS images show encouraging results and demonstrate the applicability of the proposed scheme to land-cover classification.

Authors (7)

Xin-Yi Tong (6 papers)
Gui-Song Xia (139 papers)
Qikai Lu (6 papers)
Huanfeng Shen (39 papers)
Shengyang Li (9 papers)
Shucheng You (2 papers)
Liangpei Zhang (113 papers)

Citations (662)

View on Semantic Scholar

Summary

An Essay on "Land-Cover Classification with High-Resolution Remote Sensing Images Using Transferable Deep Models"

The paper "Land-Cover Classification with High-Resolution Remote Sensing Images Using Transferable Deep Models" addresses the challenge of accurately classifying land-cover types in heterogeneous high-resolution remote sensing (HRRS) images. The complexity of such images, due to diverse sources and varying conditions, necessitates efficient classification schemes leveraging transferable deep models. This work proposes a robust methodology utilizing Convolutional Neural Networks (CNNs) to achieve high-accuracy classification in HRRS domains.

Methodology Overview

The central innovation of the paper lies in adapting deep learning models for land-cover classification across varied HRRS images without the requirement for labels on new datasets. The approach follows a multi-stage process:

Pre-Training on Source Data: The methodology starts with pre-training a deep CNN using a well-annotated dataset, referred to as the source data. The paper introduces the Gaofen Image Dataset (GID) to support this phase, which includes 150 high-resolution Gaofen-2 satellite images covering diverse geographic areas.
Pseudo-Labeling and Sample Selection: Unlabeled target images undergo classification using the pre-trained CNN. Patches with high prediction confidence receive pseudo-labels. A retrieval mechanism further refines the candidate set by matching these pseudo-labeled patches with similar source data, enhancing the reliability of the labeling.
Hybrid Classification: The model fine-tuned with pseudo-labeled data combines patch-wise classification and object-based voting using hierarchical segmentation. This dual approach leverages both category and boundary information to improve classification accuracy.

Experimental Results

The paper provides empirical evidence showcasing the applicability of the proposed scheme. Various multi-source HRRS images (including Gaofen-1, Jilin-1, Ziyuan-3, Sentinel-2A, and Google Earth data) were used to validate the model's transferability. The results demonstrate a significant improvement in classification accuracy, compared to traditional methods using features such as spectral, GLCM, and LBP. For instance, fine-tuning using the proposed scheme achieved an Overall Accuracy (OA) of up to 94.56% on complex datasets, surpassing conventional classifiers.

Implications and Future Directions

Practically, this research advances the capability of employing HRRS imagery for applications in land resource management and urban planning, without the usual constraints of manual labeling. Theoretically, it expands the discourse on transfer learning in remote sensing by emphasizing model adaptability via pseudo-labels and feature similarity.

Speculatively, future developments might explore enhancements in domain adaptation techniques, particularly involving multi-temporal analysis for more granular classification tasks. Integrating sophisticated feature learning techniques could further bolster the classification framework's ability to resolve inter-class ambiguities within HRRS images.

In conclusion, the paper presents a significant contribution to the field of remote sensing, offering a scalable, label-free classification solution that is both efficient and adaptable to diverse imaging conditions across multiple sources.

PDF Markdown

Related Papers

Find Related Papers