25 March 2024 in IEEE Transactions on Geoscience and Remote Sensing
The completeness of road extraction is very important for road application. However, existing deep learning (DP) methods of extraction often generate fragmented results. The prime reason is that DP-based road extraction methods use square kernel convolution, which is challenging to learn long range contextual relationships of roads. The road often produce fractures in the local interference area. Besides, the quality of extraction results will be subjected to the resolution of remote sensing (RS) image. Generally, an algorithm will produce worse fragmentation when the used data differs from the resolution of the training set. To address these issues, we propose a novel road extraction framework for RS images, named the Multi-Scale and Multi-Direction Feature Fusion Network (MSMDFF-Net). This framework comprises three main components: the Multi-Directional Feature Fusion (MDFF) Initial Block, the Multi-Scale Residual (MSR) encoder, and the Multi-Directional Combined Fusion (MDCF) decoder. Firstly, according to the road’s morphological characteristics, we develop a strip convolution module with a direction parameter (SCM-D). Then, to make the extracted result more complete, four SCM-D with different directions are used to MDFF-Initial Block and MDCF-decoder. Finally, we incorporate an additional branch into the ResNet encoding module to build MSR-encoder for improving the generalization of the model on different resolution RS image. Extensive experiments on three popular datasets with different resolution (Massachusetts, DeepGlobe, and SpaceNet datasets) show that the proposed MSMDFF-Net achieves new state-of-the-art results. The code will be available at https://github.com/wycloveinfall/MSMDFF-NET.
19 March 2025 in IEEE Geoscience and Remote Sensing Letters
Contextual road features and multi-scale spatial semantic information play a vital role in road extraction from remote sensing (RS) images. However, accurately modeling these essential features with current CNN-based road extraction algorithms remains challenging, leading to fragmented roads in occluded areas. Inspired by the self-attention mechanism of Transformers in natural language processing, we propose an innovative MSMDFFNet in conjunction with Swin Transformer (SwinMSMDFFNet) for road extraction from RS images. Firstly, the Swin Transformer is embedded as an auxiliary encoder into the MSMDFFNet to incorporate necessary self-attention mechanisms. Meanwhile, a Multi-Granularity Sampling (MGS) module is introduced to enhance the computation of self-attention at multiple granularities by the Swin Transformer. This module specifically transforms the feature maps produced by the main encoder into suitable inputs for the auxiliary encoder. Furthermore, to enhance the connections between adjacent local windows in the auxiliary encoder, a Cross-Direction Fusion (CDF) module is designed for feeding the features of the auxiliary encoder back into the main encoder. Extensive experiments conducted on the DeepGlobe dataset and LSRV dataset demonstrate that our proposed SwinMSMDFFNet has significant advantages in extracting road structure, particularly in areas with long-distance occlusions. It surpasses existing methods in pixel-level metrics such as F1 score, IoU, and connectivity metric APLS. The code will be made publicly available at: https://github.com/wycloveinfall/SwinMSMDFFNet.
16 December 2024 in IEEE Transactions on Geoscience and Remote Sensing
The challenge of computing electromagnetic (EM) scattering from rough surfaces is a notably complex yet crucial issue in the field of microwave remote sensing. The Decomposition-Reconstruction Theory (DRT) has been previously introduced with the objective of reducing the computational complexity of scattering from rough surfaces. The central idea of DRT is to replace the scattering from rough surfaces with that from sinusoidal or other simpler, basic surfaces. The EM scattering from these simpler surfaces can be calculated more easily. The total scattering from the original rough surface is then reconstructed with the scattering from all the basic surfaces. This article introduces the DRT-EBCM, a novel method that combines the DRT and the Extended Boundary Condition Method (EBCM) to handle scattering from rough surfaces. The DRT-EBCM method decomposes the system integrals within the Dirichlet/Neumann matrices (system matrices) of the rough surface and represents them with the basic system integrals from complex sinusoidal surfaces (basic surfaces). This process leads to the derivation of new DRT-EBCM formulations. These new formulations not only establish a quantitative relationship between EM scattering from rough surfaces and basic surfaces, but they also significantly enhance computational efficiency due to the reduced complexity of the basic system integrals. Moreover, a new Fourier-Bessel (FB) function is introduced, which plays a crucial role in the scattering from basic surfaces. This function possesses a series of excellent properties that simplify the DRT-EBCM formulations and can potentially be applied in other scattering calculation methods. Simulation experiments were conducted to validate the proposed DRT-EBCM, confirming its correctness and validity. Furthermore, the application of DRT-EBCM in fast scattering calculation for rough surfaces was demonstrated. The results showed that DRT-EBCM significantly improves computational efficiency compared to the classical EBCM.
5 December 2025 in IEEE Transactions on Geoscience and Remote Sensing
A stable and accurate Energy-Constrained T-matrix (EC-T-matrix) method is developed to calculate electromagnetic (EM) scattering from arbitrary rough surfaces. This method is proposed to enhance the stability and convergence of existing T-matrix methods. The conventional matrix inversion process is replaced with nonlinear programming to eliminate divergence caused by ill-conditioned matrix operations. The ill-conditioned matrix is systematically decomposed into a numerically stable block and an instable block. These blocks are then assigned as the objective function and linear constraint, respectively, in the optimization process. Furthermore, the law of energy conservation (unitarity condition) is implemented as a nonlinear constraint to ensure physically consistent scattering solutions. Numerical simulations using the proposed method are compared with established approaches including Stable Extended Boundary Condition Method (SEBCM), Small Perturbation Method (SPM), Small Slope Approximation (SSA), Truncated Extended Boundary Condition Method (TEBCM) and Multi-Level Fast Multiple Method (MLFMM). The stability and convergence of the EC-T-matrix method are confirmed by experimental results across diverse rough surface profiles. Stable convergent solutions are consistently achieved even under extreme corrugations. These results demonstrate superior accuracy and computational robustness compared to existing analytical and numerical methods.
29 April 2026 in IEEE Transactions on Geoscience and Remote Sensing
Recent decades have seen significant advances in road network extraction via artificial intelligence. However, single-source approaches remain vulnerable to sensor-specific limitations, such as occlusions in optical imagery. While multi-source fusion can mitigate these gaps, sensor heterogeneity and the lack of well-registered, well-labeled multisource datasets make establishing a unified fusion strategy difficult. In this paper, we propose a generic framework for fusing road information derived from multisource remote sensing (RS) images at the decision level, enabling effective combination of outputs from diverse single-source machine learning (ML) models. The presented method consists of two stages. First, to address redundancy from repeated extractions, we propose a geometric rule-based fusion strategy, featuring a novel non-maximum suppression algorithm tailored for linear targets. Second, evidence from different sources is aggregated using the evidential reasoning (ER) rule, guided by each ML model’s performance and the confidence maps of their outputs. The final fusion result is achieved by optimizing graph labeling through an energy function that integrates the ER rule with topological constraints. The proposed framework effectively accommodates heterogeneous inputs and is validated on multi-temporal and multi-modal RS images. Experimental results demonstrate marked improvements in road network completeness and correctness over single-source extraction, while also outperforming representative decision-level fusion methods in overall quality with reduced redundancy.