Yongsheng Zhang

Yongsheng Zhang


Research Focus & Metrics

Academic Career Length

9Years (2018-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
21 February 2024 in Remote Sensing
Verified
Journal Article
Open access
PDF file
ZWZiquan Wang
Ziquan Wang
YZYongsheng Zhang
Yongsheng Zhang
ZZZhenchao Zhang
Zhenchao Zhang
ZJZhipeng Jiang
Zhipeng Jiang
YYYing Yu
Ying Yu
LLLei Li
Lei Li
Robust segmentation in adverse weather conditions is crucial for autonomous driving. However, these scenes struggle with recognition and make annotations expensive, resulting in poor performance. As a result, the Segment Anything Model (SAM) was recently proposed to finely segment the spatial structure of scenes and to provide powerful prior spatial information, thus showing great promise in resolving these problems. However, SAM cannot be applied directly for different geographic scales and non-semantic outputs. To address these issues, we propose SAM-EDA, which integrates SAM into an unsupervised domain adaptation mean-teacher segmentation framework. In this method, we use a “teacher-assistant” model to provide semantic pseudo-labels, which will fill in the holes in the fine spatial structure given by SAM and generate pseudo-labels close to the ground truth, which then guide the student model for learning. Here, the “teacher-assistant” model helps to distill knowledge. During testing, only the student model is used, thus greatly improving efficiency. We tested SAM-EDA on mainstream segmentation benchmarks in adverse weather conditions and obtained a more-robust segmentation model.

21 January 2022 in Remote Sensing
Verified
Journal Article
Open access
PDF file
FFFan Feng
Fan Feng
YZYongsheng Zhang
Yongsheng Zhang
JZJin Zhang
Jin Zhang
BLBing Liu
Bing Liu
Hyperspectral images can capture subtle differences in reflectance of features in hundreds of narrow bands, and its pixel-wise classification is the cornerstone of many applications requiring fine-grained classification results. Although three-dimensional convolutional neural networks (3D-CNN) have been extensively investigated in hyperspectral image classification tasks and have made significant breakthroughs, hyperspectral classification under small sample conditions is still challenging. In order to facilitate small sample hyperspectral classification, a novel mixed spatial-spectral features cascade fusion network (MSSFN) is proposed. First, the covariance structure of hyperspectral data is modeled and dimensionality reduction is conducted using factor analysis. Then, two 3D spatial-spectral residual modules and one 2D separable spatial residual module are used to extract mixed spatial-spectral features. A cascade fusion pattern consisting of intra-block feature fusion and inter-block feature fusion is constructed to enhance the feature extraction capability. Finally, the second-order statistical information of the fused features is mined using second-order pooling and the classification is achieved by the fully connected layer after L2 normalization. On the three public available hyperspectral datasets, Indian Pines, Houston, and University of Pavia, only 5%, 3%, and 1% of the labeled samples were used for training, the accuracy of MSSFN in this paper is 98.52%, 96.31% and 98.83%, respectively, which is far better than the contrast models and verifies the effectiveness of MSSFN in small sample hyperspectral classification tasks.

11 April 2022 in IEEE Transactions on Geoscience and Remote Sensing
Verified
Journal Article
Closed access
GWGongxue Wang
Gongxue Wang
LJLingmei Jiang
Lingmei Jiang
CXChuan Xiong
Chuan Xiong
YZYongsheng Zhang
Yongsheng Zhang
The normalized difference snow index (NDSI) plays an important role in mapping snow cover with spaceborne visible and shortwave-infrared imagery. The NDSI variation depends on illuminating-viewing geometry and snow physical properties, including equivalent grain size (EGS), snow depth (SD) and impurity concentration, as well as fractional snow cover (FSC) within a mixed pixel; however, it is still not fully understood. To quantifiably characterize the pattern of snow NDSI variation, we use a light scattering model of snow to calculate bidirectional reflectance and consequent NDSI values for a wide range of illuminating-viewing geometries, SD, and EGS values. In the model designated bicontinuous snow model using Geometric Optics theory and Radiative Transfer Equation (bicontinuous-GO/RTE), snowpack is represented by a bicontinuous microstructure, and bidirectional reflectance is simulated based on geometric optics and vector radiative transfer equation. The discrete ordinates radiative transfer (DISORT) algorithm is used to simulate the soot concentration effect on snow NDSI. A soil spectral reflectance model (SOILSPECT) is utilized with the assumption of the linear spectral mixture of snow and soil to quantify the effect of FSC on NDSI. As for discontinuous forests, an analytical hybrid geometric-optical and radiative transfer (GORT) model in conjunction with the bicontinuous-GO/RTE model and the PROpriétésSPECTrales (PROSPECT) model is used to examine the effect of canopy cover, which is related to the maximal FSC viewable to satellites. Modeling results indicate that: 1) snow NDSI is comparably low at off-nadir viewing angles, and this effect is exacerbated by the decline in solar elevation but limited by large EGS; 2) snow NDSI increases with EGS yet becomes saturated at EGS of 500 $\mu \text{m}$ ; 3) the effect of SD that is as low as 1 cm on NDSI is rarely noticeable; 4) the concentration of internally mixed soot up to 1 ppm has little reducing effect on snow NDSI (<0.05); 5) the angular effects on the NDSI thresholding method to identify binary snow cover for soil background are nonsignificantly pronounced in the cross plane (CP); 6) NDSI increases linearly with FSC for soil background at nadir as well as in the CP; 7) the SNOw-MAPping algorithm (SNOMAP) method could omit snow cover at large zenith angles for solar zenith angles $> 20^{\circ }$ in forests; and 8) forests complicate the nonlinear relationship between NDSI and canopy cover with fully snow-covered ground beneath canopies. These findings imply important uncertainty sources of binary and FSC mapping with NDSI.

25 February 2019 in Remote Sensing
Verified
Journal Article
Open access
PDF file
KLKai Li
Kai Li
YZYongsheng Zhang
Yongsheng Zhang
ZZZhenchao Zhang
Zhenchao Zhang
GLGuangling Lai
Guangling Lai
Automatic image registration for multi-sensors has always been an important task for remote sensing applications. However, registration for images with large resolution differences has not been fully considered. A coarse-to-fine registration strategy for images with large differences in resolution is presented. The strategy consists of three phases. First, the feature-base registration method is applied on the resampled sensed image and the reference image. Edge point features acquired from the edge strength map (ESM) of the images are used to pre-register two images quickly and robustly. Second, normalized mutual information-based registration is applied on the two images for more accurate transformation parameters. Third, the final transform parameters are acquired through direct registration between the original high- and low-resolution images. Ant colony optimization (ACO) for continuous domain is adopted to optimize the similarity metrics throughout the three phases. The proposed method has been tested on image pairs with different resolution ratios from different sensors, including satellite and aerial sensors. Control points (CPs) extracted from the images are used to calculate the registration accuracy of the proposed method and other state-of-the-art methods. The feature-based preregistration validation experiment shows that the proposed method effectively narrows the value range of registration parameters. The registration results indicate that the proposed method performs the best and achieves sub-pixel registration accuracy of images with resolution differences from 1 to 50 times.

4 January 2023 in Remote Sensing
Verified
Journal Article
Open access
PDF file
FFFan Feng
Fan Feng
YZYongsheng Zhang
Yongsheng Zhang
JZJin Zhang
Jin Zhang
BLBing Liu
Bing Liu
Hyperspectral images contain rich features in both spectral and spatial domains, which bring opportunities for accurate recognition of similar materials and promote various fine-grained remote sensing applications. Although deep learning models have been extensively investigated in the field of hyperspectral image classification (HSIC) tasks, classification performance is still limited under small sample conditions, and this has been a longstanding problem. The features extracted by complex network structures with large model size are redundant to some extent and prone to overfitting. This paper proposes a low-rank constrained attention-enhanced multiple feature fusion network (LAMFN). Firstly, factor analysis is used to extract very few components that can describe the original data using covariance information to perform spectral feature preprocessing. Then, a lightweight attention-enhanced 3D convolution module is used for deep feature extraction, and the position-sensitive information is supplemented using a 2D coordinate attention. The above widely varying spatial–spectral feature groups are fused through a simple composite residual structure. Finally, low-rank second-order pooling is adopted to enhance the convolutional feature selectivity and achieve classification. Extensive experiments were conducted on four representative hyperspectral datasets with different spatial–spectral characteristics, namely Indian Pines (IP), Pavia Center (PC), Houston (HU), and WHU-HongHu (WHU). The contrast methods include several advanced models proposed recently, including residual CNNs, attention-based CNNs, and transformer-based models. Using only five samples per class for training, LAMFN achieved overall accuracies of 78.15%, 97.18%, 81.35%, and 87.93% on the above datasets, which has an improvement of 0.82%, 1.12%, 1.67%, and 0.89% compared to the second-best model. The running time of LAMFN is moderate. For example, the training time of LAMFN on the WHU dataset was 29.1 s, and the contrast models ranged from 3.0 s to 341.4 s. In addition, ablation experiments and comparisons with some advanced semi-supervised learning methods further validated the effectiveness of the proposed model designs.

15 June 2021 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Verified
Journal Article
Open access
ZWZiquan Wang
Ziquan Wang
YZYongsheng Zhang
Yongsheng Zhang
YYYing Yu
Ying Yu
LZLei Zhang
Lei Zhang
JMJie Min
Jie Min
GLGuangling Lai
Guangling Lai
The lack of data samples has become a bottleneck for the performance improvement of bridge detection model training via transfer learning. This article proposes a deep learning bridge detection algorithm aided by a traditional bridge detection method to alleviate this problem. First, we explore the theoretical constraints on the integration of prior-information feature maps extracted via the traditional unsupervised method. Second, we design a prior-information auxiliary module (PAM) to receive and inject prior information. Finally, we introduce this module into different kinds of object detection networks for testing. The experimental results show that prior masks generated through multiextremal segmentation can assist object detectors in many respects. In addition, the PAM can use a few prior-information feature maps to enhance the performance of four representative object detection models on the bridge recognition dataset from the 2020 Gaofen challenge on automated high-resolution earth observation image interpretation. Our method preserves the validity of end-to-end training, making the traditional method more valuable in the artificial intelligence context.