Combining Contextual Information by Integrated Attention Mechanism in Convolutional Neural Networks for Digital Elevation Model Super-Resolution

High-resolution digital elevation models (HR DEMs) can provide accurate geographic information, which is widely used in flood risk evaluation, hazard mapping, and hydrological modeling. The study of DEM super-resolution (SR) algorithms has relieved the need for HR DEMs. However, most CNN backbone ne...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:IEEE transactions on geoscience and remote sensing Ročník 62; s. 1 - 16
Hlavní autori: Chen, Zhanlong, Han, Xiaoyi, Ma, Xiaochuan
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: New York IEEE 2024
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Predmet:
Algorithms
Artificial neural networks
Attention
Attention mechanism
Attention mechanisms
contextual information
Convolution
Convolutional neural networks
Deep learning
Deformation effects
dependence of long-range geographical features
digital elevation model super-resolution (DEM SR)
Digital Elevation Models
Elevation
Encoders-Decoders
Environmental risk
Feature extraction
Formability
Information processing
Machine learning
Modules
Neural networks
Ridges
Risk assessment
Rivers
Superresolution
Task analysis
ISSN:0196-2892, 1558-0644
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:High-resolution digital elevation models (HR DEMs) can provide accurate geographic information, which is widely used in flood risk evaluation, hazard mapping, and hydrological modeling. The study of DEM super-resolution (SR) algorithms has relieved the need for HR DEMs. However, most CNN backbone network methods suffer from contextual information loss, resulting in a lack of ability to model long-distance dependencies on geographic features (such as ridges, rivers, drainage lines, and valleys) when handling DEM SR tasks. To tackle this issue, we propose a combining contextual information method by integrated attention mechanism (IAM) network for DEM SR (CIASR). It models both the dependence on long-range geographical features and local contextual information by using the attention mechanism. To achieve this goal, a residual module is designed to extract the low-level geographic features effectively. Then, we further propose an IAM to capture dependencies on geographic features. Specifically, a channel-spatial attention module and a self-attention module are adaptive to extract local and global contextual information, respectively. In addition, to take full advantage of the contextual information, we use a pixelshuffle module and a deformable convolution module to integrate geographic features. Compared with the convolutional neural network based DEM super resolution (SRCNN), super resolution ResNet network (SRResNet), feature-enhanced deep learning network (FEN), super-resolution via repeated refinement (SR3), Terrain feature-aware super resolution model (TfaSR), conditional encoder-decoder generative adversarial neural networks (CEDGANs), enhanced residual feature fusion network (ERFFN), and global-information-constrained deep learning network for DEM SR (GISR) methods, the experimental results show that the CIASR method has outstanding advantages in complex and uneven surface terrains.
Bibliografia:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0196-2892
1558-0644
DOI:10.1109/TGRS.2024.3423716