[Paper Review] WRICNet:A Weighted Rich-scale Inception Coder Network for Multi-Resolution Remote Sensing Image Change Detection

Yu Jiang, Lei Hu+2 more|arXiv (Cornell University)|Aug 18, 2021

Remote-Sensing Image Classification39 references4 citations

TL;DR

WRICNet proposes a novel deep learning architecture for multi-resolution remote sensing image change detection, integrating a weighted rich-scale inception module for shallow multi-scale feature extraction and a weighted rich-scale coder module for deep multi-scale representation. The method improves edge accuracy and reduces false/missed alarms by adaptively fusing features across scales, outperforming existing methods on multi-resolution datasets with significant gains in F1-score and IoU.

ABSTRACT

Majority models of remote sensing image changing detection can only get great effect in a specific resolution data set. With the purpose of improving change detection effectiveness of the model in the multi-resolution data set, a weighted rich-scale inception coder network (WRICNet) is proposed in this article, which can make a great fusion of shallow multi-scale features, and deep multi-scale features. The weighted rich-scale inception module of the proposed can obtain shallow multi-scale features, the weighted rich-scale coder module can obtain deep multi-scale features. The weighted scale block assigns appropriate weights to features of different scales, which can strengthen expressive ability of the edge of the changing area. The performance experiments on the multi-resolution data set demonstrate that, compared to the comparative methods, the proposed can further reduce the false alarm outside the change area, and the missed alarm in the change area, besides, the edge of the change area is more accurate. The ablation study of the proposed shows that the training strategy, and improvements of this article can improve the effectiveness of change detection.

Motivation & Objective

• Address the limitation of existing change detection models that perform poorly on multi-resolution remote sensing data.
• Improve feature representation by fusing shallow and deep multi-scale features effectively.
• Enhance detection accuracy, especially at change boundaries, by assigning adaptive weights to features of different scales.
• Reduce false alarms in non-changing regions and missed detections in actual change areas.
• Develop a robust model architecture that generalizes across diverse resolution levels in remote sensing imagery.

Proposed method

• Introduces a weighted rich-scale inception module to extract and fuse multi-scale features from early network layers.
• Employs a weighted rich-scale coder module to capture hierarchical deep features with multi-scale context.
• Applies learnable attention-like weights to scale-specific features, emphasizing informative scales for better edge localization.
• Uses skip connections to preserve spatial details and enhance feature propagation across the network.
• Integrates multi-scale features through a weighted fusion strategy that dynamically adjusts feature importance based on scale and context.
• Trains the network using a supervised loss function optimized for change detection, focusing on boundary precision and overall accuracy.

Experimental results

Research questions

• RQ1 Can a unified deep learning framework effectively handle change detection across multiple remote sensing image resolutions?
• RQ2 How does adaptive weighting of multi-scale features improve detection accuracy, especially at change boundaries?
• RQ3 To what extent does combining shallow and deep multi-scale features reduce false and missed alarms in change detection?
• RQ4 Does the proposed architecture generalize better than existing models on diverse multi-resolution datasets?
• RQ5 What is the contribution of the weighted fusion mechanism to overall performance in complex remote sensing scenarios?

Key findings

• WRICNet achieves superior performance on multi-resolution datasets, significantly reducing false alarms outside change regions.
• The model improves detection accuracy at change boundaries, resulting in more precise segmentation of changing areas.
• Quantitative results show a notable increase in F1-score and Intersection over Union (IoU) compared to state-of-the-art methods.
• Ablation studies confirm that the weighted feature fusion mechanism and training strategy are critical for performance gains.
• The network demonstrates robustness across varying image resolutions, validating its generalization capability.