Ni Songhao

Mr. Ni Songhao

orcid

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

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

Representative Publications

View All Publications
9 October 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Verified
Journal Article
Open access
LKLitao Kang
Litao Kang
CLChaoyue Liu
Chaoyue Liu
HFHuaitao Fan
Huaitao Fan
ZZZhimin Zhang
Zhimin Zhang
ZCZhen Chen
Zhen Chen
NSNi Songhao
Ni Songhao
In recent years, the deep learning-based synthetic aperture radar (SAR) ship detection technology has been rapidly developed, and has been widely used in fishery management, maritime surveillance, maritime rescue and other fields, but there are still some challenges. For instance, the feature maps extracted by the backbone network contain substantial redundant features, and repeatedly processing these redundant features introduces significant computational redundancy and inflates the model's parameter count. Furthermore, the blurred target edges caused by SAR imaging mechanisms also degrade detection accuracy. In order to solve these problems, this paper first proposes a backbone network using multi-receptive field partial convolution (MRFPconv) to extract features. While preserving the quality of backbone feature maps, our approach eliminates significant computational redundancy and reduces the parameter count. Subsequently, a Neck network with Bi-directional Concatenation (BiC) modules and reparameterizable modules, Reparameterizable Bi-directional Path Aggregation Network (RepBi-PAN), was adopted. This architecture reduces network parameters while incorporating shallower features, thereby enhancing the network's detection capabilities. Finally, this paper uses expert experience to design a ship contour extraction method based on image morphology. The extracted contours and a specially designed dual-branch network architecture with a channel attention mechanism are used to further enhance the network's detection capability. A lot of experiments are carried out on the SAR ship detection data sets HRSID and SSDD. MRFP-DBDet achieves a mAP value of 69.3% and an ${F_{1}}$ of 0.895 on the HRSID dataset, and a mAP value of 74.6% and an ${F_{1}}$ of 0.975 on the SSDD dataset, with a model parameter size of only 4.135M. The experimental results demonstrate that the algorithm achieves state-of-the-art detection performance among SAR ship detection methods with comparable parameter counts.


16 January 2026 in Remote Sensing
Verified
Journal Article
Open access
PDF file
NSNi Songhao
Ni Songhao
ZFzhao fuhai
zhao fuhai
MZMingjie Zheng
Mingjie Zheng
ZCZhen Chen
Zhen Chen
XLXiuqing Liu
Xiuqing Liu
Sub-meter-resolution Polarimetric Synthetic Aperture Radar (PolSAR) imagery enables precise building footprint extraction but introduces complex scattering correlated with fine spatial structures. This change renders both traditional methods, which rely on simplified scattering models, and existing deep learning approaches, which sacrifice spatial detail through multi-looking, inadequate for high-precision extraction tasks. To address this, we propose an Orthogonal Dual-Resolution Network (ODRNet) for end-to-end, precise segmentation directly from single-look complex (SLC) data. Unlike complex-valued neural networks that suffer from high computational cost and optimization difficulties, our approach decomposes complex-valued data into its orthogonal real and imaginary components, which are then concurrently fed into a Dual-Resolution Branch (DRB) with Bilateral Information Fusion (BIF) to effectively balance the trade-off between semantic and spatial details. Crucially, we introduce an auxiliary Polarization Orientation Angle (POA) regression task to enforce physical consistency between the orthogonal branches. To tackle the challenge of diverse building scales, we designed a Multi-scale Aggregation Pyramid Pooling Module (MAPPM) to enhance contextual awareness and a Pixel-attention Fusion (PAF) module to adaptively fuse dual-branch features. Furthermore, we have constructed a VHR PolSAR building footprint segmentation dataset to support related research. Experimental results demonstrate that ODRNet achieves 64.3% IoU and 78.27% F1-score on our dataset, and 73.61% IoU with 84.8% F1-score on a large-scale SLC scene, confirming the method’s significant potential and effectiveness in high-precision building extraction directly from SLC.

1 January 2026 in IEEE Geoscience and Remote Sensing Letters
Verified
Journal Article
Closed access
NSNi Songhao
Ni Songhao
MZMingjie Zheng
Mingjie Zheng
ZFzhao fuhai
zhao fuhai
XLXiuqing Liu
Xiuqing Liu

11 November 2025 in TechRxiv
Verified
Preprint
Open access
NSNi Songhao
Ni Songhao
MZMingjie Zheng
Mingjie Zheng
FZFuhai Zhao
Fuhai Zhao
XLXiuqing Liu
Xiuqing Liu
Dual-frequency (DF) Polarimetric Synthetic Aperture Radar (PolSAR) is crucial for Earth observation, offering richer information than single-frequency (SF) systems for land cover segmentation. However, SF tasks remain prevalent, and current deep learning methods typically require separate models for SF/DF tasks, hindering feature reuse across bands and broader applicability. To overcome this limitation, this letter proposes a unified network with training strategy capable of processing both SF and DF PolSAR images. Leveraging the high spatial similarity yet distinct polarimetric characteristics of DF PolSAR images, the architecture employs a shared backbone consisting of Large Kernel Selective (LKS) for joint feature extraction and Lightweight Band-specific Mixture of Experts (LBME) for capturing distinct patterns across bands. The Decoder employs Mixing Feature Transformer (MFT) blocks to refine fusion features by cross-attention and maintains robustness for SF features by explicit prompts. The proposed unified model achieves higher metrics on both SF (mIoU: 72.58% for L-band, 72.74% for S-band) and DF (mIoU: 75.19% for LS combined) tasks. The code will be made publicly available at URL