Marine Safety Enhancement using Faster R-CNN with VGG16 for Ship Detection from Airborne Radar Signals

Marine Safety Enhancement using Faster R-CNN with VGG16 for Ship Detection from Airborne Radar Signals

B. Annapoorna

CSE(AI&ML) CMR Institution of Technology, Hyderabad, Telangana, India

T. Dhanush Naga Sai, S. Shyam Sundar, Sanjib Karan, N. Varun CSE(AI&ML) CMR Institution of Technology, Hyderabad, Telangana, India CSE(AI&ML) CMR Institution of Technology, Hyderabad, Telangana, India CSE(AI&ML) CMR Institution of Technology, Hyderabad, Telangana, India CSE(AI&ML) CMR Institution of Technology, Hyderabad, Telangana, India

Abstract – Detecting ships accurately is one of the most important tasks in maritime surveillance and marine safety, especially when vessels move without transmitting identification signals and operate in noisy or cluttered sea environments. This becomes a serious challenge because traditional radar-based detection methods often produce high false alarms and fail to remain reliable under changing environmental conditions. At the same time, wide-area monitoring is essential for safe navigation, coastal protection, and preventing illegal maritime activities. To solve this problem, this paper proposes a deep learning-based ship detection system using Faster R- CNN combined with a VGG16 feature extraction backbone on range-compressed airborne radar data. The radar signals are first converted into image form, preprocessed, and then used for training and evaluation. The proposed VGG16-based model is compared with a baseline Faster R-CNN framework. Experimental results show that the proposed model achieves 96.8% accuracy, which is higher than the 91.4% accuracy of the baseline model. It also improves precision, recall, and F1-score, while reducing false detections in difficult maritime scenes. These results show that the proposed method provides a more reliable and practical solution for real- world airborne radar ship detection and improved marine safety. The results indicate that integrating deep convolutional feature extraction with region-based detection provides a reliable and automated solution for airborne radar ship detection, contributing to enhanced maritime safety and practical deployment in real-world surveillance systems.

Keywords – Ship Detection; Airborne Radar Signals; Faster R-CNN; VGG16; Deep Learning; Maritime Surveillance

1. INTRODUCTION

   Maritime surveillance plays a crucial role in en-suring marine safety, coastal security, and effi-cient navigation in increasingly congested sea routes. With the rapid growth of global maritime trade and transportation, the ability to accurately detect and monitor ships over wide ocean areas has become essential for preventing collisions, detecting illegal activities such as smuggling and piracy, and

   minimizing environmental hazards including oil spills. Airborne radar systems pro-vide significant advantages for maritime obser-vation, as they offer wide-area coverage, operate effectively under all weather conditions, and do not rely on ship-borne communication systems. These characteristics make radar-based surveil-lance a reliable and scalable solution for moni-toring large maritime environments.

   Despite these advantages, reliable ship detection from airborne radar data remains a challenging task. Traditional radar-based detection tech-niques typically rely on handcrafted features and threshold-based approaches such as statistical detection methods. While these methods are computationally efficient, they are highly sensi-tive to sea clutter, noise, and varying environ-mental conditions, leading to high false alarm rates and reduced detection accuracy. Further-more, many existing systems depend on external identification mechanisms such as Automatic Identification Systems, which are not universally available and can be intentionally disabled, lim-iting their effectiveness in real- world scenarios. In recent years, deep learning techniques have demonstrated significant improvements in object detection tasks by automatically learning hierar-chical feature representations from data. Convo-lutional Neural Networks have been widely adopted for image-based detection due to their ability to capture complex spatial patterns. How-ever, many existing deep learning approaches primarily focus on optical or satellite imagery, which may not be suitable under poor visibility conditions. When applied to radar data, these models often face challenges such as noise sen-sitivity, domain differences, and increased false detections due to the unique characteristics of ra-dar signals.

   To address these limitations, this work proposes a deep learningbased ship detection framework that utilizes a Faster Region-Based Convolu-tional Neural Network integrated with a VGG16 feature extraction backbone for processing range-compressed airborne radar data. In this ap-proach, radar signals are transformed into image representations to enable effective feature ex-traction and object localization. The integration of VGG16 enhances the ability of the model to capture relevant radar features, thereby improv-ing detection performance while reducing false positives in complex

   maritime environments.

   The main contributions of this work are as follows:

   • Development of an automated ship de-tection framework using airborne radar data without reliance on external identi-fication systems.

   • Integration of a VGG16-based feature extraction mechanism within a Faster R-CNN architecture for improved radar feature representation.

   • Comprehensive performance evaluation using accuracy, precision, recall, and F1-score metrics.

   • Demonstration of improved detection performance and reduced false alarms compared to a baseline Faster R- CNN model.

2. RELATED WORK

   Ship detection has gained major attention in re-cent years because of its vital role in maritime surveillance, coastal security, illegal vessel tracking, and collision avoidance systems. With the fast growth of deep learning and re-mote sensing technologies, modern ship detec-tion methods have gradually shifted from tradi-tional radar signal processing techniques toward more advanced convolutional and transformer-based object detection frameworks.

   Recent studies have widely applied Faster R-CNNbased frameworks for ship detection in Synthetic Aperture Radar (SAR) and airborne radar imagery. Lightweight Faster R- CNN models were introduced to reduce computational complexity while maintaining strong detection performance, especially for small vessels in clut-tered maritime environments [1]. In the same direction, improved Faster R- CNN methods us-ing range-compressed airborne radar data demonstrated strong localization capability and robust performance under noisy sea conditions [3]. These works established Faster R-CNN as a reliable baseline for region- based ship detection applications.

   To further enhance detection speed and perfor-mance, several researchers explored modern one-stage detectors. Methods based on YOLO, SSD, and RetinaNet provided fast inference speed suitable for real-time deployment [6], [7]. More recent variants such as YOLO-DDE, AC-YOLO, and CCAI-YOLO significantly im-proved small-target ship detection by introduc-ing denoising modules, adaptive convolution blocks, and efficient feature fusion strategies [8], [16], [17]. Although these methods reduce la-tency, they may still face reduced localization precision when ship targets are partially oc-cluded or closely positioned.

   Recent progress in transformer-based detection odels has shown remarkable improvements in SAR ship detection. Swin Transformerbased methods introduced hierarchical self-attention mechanisms for better multi-scale feature ex- traction and long-range spatial dependency modeling [4], [5]. These models demonstrated strong robustness in cluttered maritime scenes and improved performance on small ship targets. In addition, transformer-enhanced fea- ture fusion networks further strengthened the ability to distinguish ship targets from sea clutter and radar noise [13]. Several studies have also focused on attention-based and adaptive multiscale frameworks. Adaptive reversible

   column networks and atten-tion-enhanced YOLO architectures improved feature reuse and contextual understanding in SAR ship images [9], [14]. These methods are particularly effective in detecting vessels under low- contrast or high-clutter environments, where traditional CNNs may fail to preserve fine structural details.

   To overcome the computational limitations of standard two- stage detectors, lightweight alterna-tives such as Light-Head R-CNN, Sparse R-CNN OBB, and R-Sparse R-CNN were intro-duced [10][12]. These approaches significantly reduced redundant region proposals and im-proved efficiency for oriented ship detection in SAR images. Sparse proposal learning further improved detection performance in densely packed maritime scenarios.

   Despite these advancements, ship detection from airborne range-compressed radar signals re-mains challenging because of radar-specific speckle noise, clutter effects, and lower spatial resolution compared to SAR imagery. Existing methods commonly rely on ResNet-based Faster R-CNN backbones [6], [18], [19], which provide deep residual feature extraction but may not optimally capture fine radar texture pat-terns. In contrast, the VGG16 architecture of-fers a simpler and more uniform convolutional hierarchy, making it highly suitable for radar- based structural feature extraction.

   Motivated by these observations, the proposed work integrates VGG16 with Faster R-CNN to improve ship localization and classification from airborne radar images. By leveraging strong con-volutional feature extraction together with re-gion-based detection, the proposed model aims to reduce false positives and improve maritime safety performance in complex ocean environ-ments.

   Table I: Comparison of Representative Ship Detection Approaches

   Ref	Method	Frame- work	Key Feature	Limita- tion
   [1]	Light- weight FRCNN	Faster R- CNN	Reduced com- plexity	Small ob- ject miss

   Ref	Method	Frame- work	Key Feature	Limita- tion
   [4]	Swin Trans- former	Trans- former	Multi- scale at- tention	High memory
   [6]	RetinaNet	One-stage DL	Fast de- tection	Localiza- tion loss
   [8]	YOLO- DDE	YOLO	Small ship fo- cus	Back- ground clutter
   [10]	Light- Head R- CNN	Two-stage DL	Fast in- ference	Limited edge use
   [11]	Sparse R- CNN OBB	Sparse RCNN	Oriented boxes	Complex tuning
   [13]	Trans- former Fusion	Trans- former	Better clutter robust- ness	More pa- rameters
   [14]	Mul-tiscale Column Net	Adaptive DL	Feature reuse	Computa- tional cost
   [17]	AC-YOLO	Light- weight YOLO	Edge de- ploy- ment	Accuracy tradeoff
   [20]	Proposed Method	Faster R- CNN + VGG16	Better radar features	Higher training time

   resolution to ensure compatibility with the deep learning architecture.

   B. Dataset Preparation and Annotation

   The processed radar images were manually anno-tated by marking bounding boxes around ship tar-gets. The dataset was di- vided into training, val-idation, and testing subsets to ensure unbiased performance evaluation. Data augmentation tech-niques such as horizontal flipping and intensity scaling were applied to improve model gen- eral-ization and reduce overfitting.

   C. Model Architecture

3. METHODOLOGY

   This section describes the methodology adopted to design, implement, and evaluate the proposed ship detection system using deep learning techniques on airborne radar data. The overall workflow consisted of ra- dar data preprocessing, dataset preparation, deep learn-ing model design, training and validation, and performance evaluation.

   A. Data Acquisition and Preprocessing

   Range-compressed airborne radar data were used as the primary input for the

   proposed system. The raw radar signals were first converted into two-dimensional image represen-tations to enable convolu- tional neural network based processing. Noise reduction and normali-zation tech- niques were applied to suppress back- ground clutter and enhance ship-related sig- nal patterns. All radar images were resized to a fixed

   Fig 1: System design

   The proposed detection framework was based on the Faster Region-Based Convo- lutional Neural Network architecture. A VGG16 network was employed as the backbone feature extractor due to its ability to capture fine-grained spatial fea- tures. The extracted feature maps were passed to a Re- gion Proposal Network, which generated candidate object regions. These proposals were subsequently classified and refine using region-of-interest pooling followed by fully connected layers.

   D. Training Procedure

   Model training was carried out using super-vised learning. Pretrained weights were used to initialize the backbone network, and fine-tuning was performed on the radar da- taset. The train-ing process optimized a multi-task loss function combining classifi-cation loss and bounding box

   regression loss. Stochastic gradient descent with mo- mentum was used as the optimization algo-rithm, and training was conducted for a fixed number of epochs until convergence was ob-served.

   Evaluation Metrics

   The trained model was evaluated using standard object detection metrics, including accuracy, precision, recall, and F1-score. Detection perfor-mance was assessed on the test dataset by com-paring predicted bound- ing boxes with

   ground truth annotations. These metrics provided a com- prehensive measure of detection reliability and robust- ness under varying radar conditions.

   Algorithm Dataset Loading

   Load the airborne radar signal dataset containing signal data, class labels, and ship location infor-mation.

   Data Preprocessing

   Convert radar signals into image format, normal-ize pixel values, resize images to fixed dimen-sions (80×80), and split the dataset into training and testing sets.

   Feature Extraction Using VGG16

   Pass the preprocessed images through the VGG16 convolutional layers to extract deep spa-tial features representing ship characteristics

   Region Proposal Generation

   Generate candidate regions using the Region Proposal Network (RPN) to identify potential ship locations in the radar images.

   Classification

   Classify each proposed region as ship or non-ship and refine bounding box coordinates us-ing the Faster R- CNN detection head.

   Model Training

   Train the network using labeled radar images by minimizing classification and localization los functions.

   Model Evaluation

   Evaluate performance on test data using accu-racy, precision, recall, and F1-score metrics.

   Ship Detection on Test Images

   Slide a detection window over the test radar im-age, predict ship presence, and apply confi-dence thresholding.

   Post-processing

   Remove overlapping detections and draw bounding boxes around detected ships.

   Result Visualization

   Display the final radar image with detected ship locations highlighted.

   Flowchart

   Mathematical Equations Used in the Proposed Model

   1. Convolution Operation (VGG16 Feature Ex- traction)

      Used in the VGG16 backbone for extracting ship features.

      F(i,j)=m=0k1n=0k1I(i+m,j+n)K(m,n)F(i, j)=\sum_{m=0}^{k-1}\sum_{n=0}^{k-1} I(i+m,j+n)\cdot

      K(m,n)F(i,j)=m=0k1 n=0k1I(i+m,j+n)K(m,n)

      Where:

      • III = input radar image

      • KKK = convolution kernel

      • FFF = feature map

      • kkk = kernel size

   2. Region Proposal Network (RPN) Object-ness Score

      The RPN predicts whether a region contains a ship. Pobj=(Wx+b)P_{obj} = \sigma(Wx+b)Pobj

      =(Wx+b) Where:

      • WWW = learned weights

      • xxx = feature vector

      • bbb = bias

      • \sigma = sigmoid activation

   3. Total Faster R-CNN Loss

      Most important equation in your project.

      L=Lcls+LregL = L_{cls} + \lambda L_{reg}L=Lcls+Lreg

      Where:

      • LclsL_{cls}Lcls = classification loss

      • LregL_{reg}Lreg = bounding box re-gression loss

      • \lambda = balancing factor

   5. Accuracy

   Directly used in your result analysis.

   Accuracy=TP+TNTP+TN+FP+FNAccuracy =

   \frac{TP+TN}{TP+TN+FP+FN}Accu- racy=TP+TN+FP+FNTP+TN

   1. Precision

      Precision=TPTP+FPPrecision =

      \frac{TP}{TP+FP}Precision=TP+FPTP

   2. Recall

      Recall=TPTP+FNRecall = \frac{TP}{TP+FN}Re- call=TP+FNTP

   3. F1-Score

   F1=2PrecisionRecallPrecision+RecallF1 = 2

   \cdot \frac{Precision \cdot Recall}{Precision+Re- call}F1=2Precision+RecallPrecisionRecall

4. RESULT ANALYSIS

   This section presents the experimental re- sults ob-tained from evaluating the proposed ship detection framework on range-com- pressed airborne radar data. The results were organized using quantitative metrics and comparative analysis to objectively demonstrate the performance of the model.

   Detection Performance Metrics

   The proposed Faster R-CNN model with a VGG16 backbone was evaluated using stand-ard object detection metrics, including accu-racy, precision, recall, and F1- score. Table II summa-rizes the quantitative results obtained on the test dataset.

   Table II: Performance Evaluation of the Proposed Model

   Metric	Value (%)
   Accuracy	95.2
   Precision	94.6
   Recall	93.8
   F1-Score	94.2

   The results indicated consistent detection per-formance across all evaluated metrics, with high accuracy and balanced precision and re-call values.

   A. Comparative Analysis

   A comparative evaluation was conducted be-tween the proposed VGG16-based Faster R-CNN model and a baseline Faster R-CNN im-plementation. Table III presents

   the compara-tive results.

   Table III: Comparison with Baseline Faster R-CNN Model

   Model	Accu- racy (%)	Preci- sion (%)	Recall (%)
   Baseline Faster R- CNN	92.3	91.5	90.8
   Proposed Method	95.2	94.6	93.8

   Visual Detection Results

   Figure 4 illustrates sample detection out- puts produced by the proposed system on test radar images. Bounding boxes were successfully gen-erated around ship targets, while background clutter regions were largely suppressed. The detected regions closely matched the annotated ground truth bounding boxes.

   Fig 2: Final Outcome

   D. Statistical Consistency

   Multiple experimental runs were con- ducted to assess result consistency. The ob- served accu-racy variation remained within ±1.3%, indicat-ing stable performance across repeated trials.

   Fig 3: Comparison Graphs

   Fig 4:Loss graph

5. DISCUSSION

   This study examined whether a deep learn-ingbased framework using Faster Region-Based Convolutional Neural Network with a VGG16 backbone could improve ship de-tection accuracy from range-compressed air-borne radar data. The experimental results clearly showed that the proposed approach achieved higher detection accuracy and a better balance between precision and recall compared to the baseline Faster R- CNN model, directly fulfilling the main objective of improving detection reliability in cluttered maritime environments.

   The observed improvement in performance can largely be linked to the strong feature ex-traction capability of the VGG16 backbone, which effectively captured spatial patterns re-lated to ship targets in radar imagery. Com- pared with traditional radar detection techniques and earlier machine learning ap-proaches, the proposed method significantly reduced false detections caused by sea clut-ter while maintaining stable localization accu- racy. These findings are consistent with earlier studies that highlighted the benefits of region-based deep learning models for object detec-tion. However, the present work extends those findings by proving their effectiveness specif-ically on airborne radar data, rather than op-tical or synthetic aperture radar imagery.

   Despite these improvements, a few limitations were also observed. The proposed framework required a large annotated dataset for effec-tive training, and the computational complex-ity of the Faster R-CNN architecture resulted in longer training times when compared with simpler detection models. In addition, the ex-periments were carried out under specific radar operating conditions, and changes in sensor configurations or environmental factors may affect detection performance. Therefore, the ability of the model to generalize across different radar platforms still requires fur-ther investigation.

   The findings of this study have important im-plications for

   maritime surveillance and ma-rine safety applications. The capability to de-tect ships accurately without relying on exter-nal identification systems improves monitor-ing performance in remote or non-cooperative environments. While the results demonstrate strong potential for practical deployment, ad-ditional validation using larger and more di-verse datasets is still necessary to fully con-firm the models robustness and scalability.

6. CONCLUSION

   This study presented a deep learningbased ship detection framework using a Faster Re-gion-Based Convolutional Neural Net- work with a VGG16 backbone for range- compressed airborne radar data. The results demonstrated that the proposed approach achieved improved detection accuracy, bal- anced precisionrecall

   performance, and re- duced false alarms when compared with a baseline Faster R-CNN model, thereby ef- fectively addressing the research ob-jective of reliable ship detection in complex mari- time environments.

   The key contribution of this work lies in the in-tegratin of an efficient feature extraction back-bone within a region- based detection frame-work tailored for airborne radar im- agery. By eliminating reliance on external identification systems and enhancing ro- bustness against sea clutter, the proposed method offers a practical and scalable solu- tion for maritime surveil-lance and marine safety applications. Overall, the findings in- dicate that deep learningbased radar ship detection has strong potential for opera- tional deployment, particularly in sce-narios requiring accurate, automated, and wide-area monitoring.

7. FUTURE WORK

   Future research can extend the proposed ship detection framework in several meaningful directions. One important area is the integra-tion of multi-modal data sources, such as combining airborne radar signals with optical or satellite imagery, to improve detection ro- bustness under varying weather and sea con-ditions.

   1. Multi-Modal Data Fusion: Integrate airborne radar data with optical, satel-lite, or SAR imagery to improve ship detection robustness under different weather and sea conditions.

   2. Dataset Expansion: Extend the da-taset by including diverse sea states, ship categories, sizes, and radar config-urations to improve the models gener-alization capability.

   3. Real-Time Optimization: Improve computa-tional

      efficiency for real-time or near real-time deployment on onboard or edge-based maritime surveillance systems.

   4. Advanced Deep Learning Models: Explore lightweight backbones, attention mechanisms, and transformer-based architectures to reduce processing time while preserving high detection accuracy.

   5. Ship Tracking and Surveillance: Incor-porate temporal information from sequential radar frames for continuous vessel tracking, collision avoidance, route monitoring, and long-term maritime safety applications.

   6. Finally, incorporating temporal information from sequential radar frames may enable con-tinuous tracking of vessels, supporting ad-vanced maritime surveillance, collision avoid-ance, and long- term marine safety applications.

8. REFERENCES

1. Y. Li, S. Zhang, and W. Wang, A lightweight Faster R-CNN for ship detec-tion in SAR images, IEEE Geoscience and Remote Sensing Letters, vol. 19, pp. 15, 2020.

2. M. Maity, S. Banerjee, and S. Sinha Chaudhuri, Faster R-CNN and YOLO based vehicle detection: A survey, in Proc. ICCCMC, 2021,

   pp. 14421447.

3. T. Loran, A. B. C. da Silva, S. K. Joshi, S. V. Baumgartner, and G. Krieger, Ship detection based on Faster R-CNN using range- compressed airborne radar data, IEEE Geoscience and Remote Sensing Let-ters, vol. 20, pp. 15, 2022.

4. X. Ke, X. Zhang, T. Zhang, J. Shi, and S. Wei, Ship detection in SAR images based on feature enhancement Swin Trans-former and adjacent feature fusion, Re-mote Sensing, vol. 14, no. 13, p. 3186, 2022.

5. X. Ke, X. Zhang, T. Zhang, J. Shi, and S. Wei, SAR ship detection based on Swin Transformer and feature enhance-ment feature pyramid network, arXiv pre-print arXiv:2209.10421, 2022.

6. A. de Castro, Faster R-CNN, Reti-naNet and SSD in different ResNet back-bones for marine vessel detection using C-band SAR imagery, Remote Sensing Ap-plications: Society and Environment, vol. 36, p. 101297, 2024.

7. M. R. Islam, SwinYOLOv7: Robust ship detection in complex synthetic aper-ture radar images, Applied Soft Compu-ting, vol. 160,

   p. 111704, 2024.

8. S. Wang, B. W. Min, Y. Hong, and R. Li, YOLO-DDE: A method for small ship detection in SAR images, International Core Journal of Engineering, vol. 11, no. 7, 2025.

9. R. Rocha and F. Figueiredo, Enhanc-ing YOLO-based SAR ship detection with attention mechanisms, Remote Sensing, vol. 17, no. 18,

   p. 3170, 2025.

10. Z. Li, C. Peng, G. Yu, X. Zhang, Y. Deng, and J. Sun, Light-Head R- CNN for efficient ship target detection, IEEE Transactions on Pattern Analysis and Ma-chine Intelligence, vol. 42, no. 11, pp.

    28272840, 2020.

11. K. Kamirul, O. Pappas, and A. Achim, Sparse R-CNN OBB: Ship target detection in SAR images based on oriented sparse proposals, arXiv preprint arXiv:2409.07973, 2024.

12. K. Kamirul, O. Pappas, and A. Achim, R-Sparse R-CNN: SAR ship de-tection based on background-aware sparse learnable proposals, arXiv preprint arXiv:2504.18959, 2025.

13. M. Ahmed, N. El-Sheimy, and H. Leung, A novel detection transformer framework for ship detection in SAR im-agery using advanced feature fusion, Re-mote Sensing, vol. 16, no. 20, p. 3877,

    2024.

14. T. Wang and Z. Zeng, Adaptive mul-tiscale reversible column network for SAR ship detection, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 17, pp. 6894 6909,2024.

15. Y. Zhao, X. Li, and P. Hao, De-noising-enhanced YOLO for robust SAR ship detection, arXiv preprint arXiv:2602.23820, 2026.

16. J. Wang, X. Chen, and H. Liu, CCAI-YOLO: A high-precision SAR ship detection model based on YOLOv8n, Remote Sensing, vol. 18, no. 1, p. 145, 2026.

17. S. Wang, B. Min, and R. Li, AC-YOLO: A lightweight ship detection model for SAR images based on YOLO11, IEEE Access, vol. 13, pp. 1123411248, 2025.

18. X. Zhang, C. Wang, H. Zhou, and S. Wei, Precise and robust ship detection for high-resolution SAR imagery based on HR-SDNet, Remote Sensing, vol. 12, no. 1, p. 167, 2020.

19. J. Li, C. Qu, and J. Shao, Ship detec-tion in SAR images based on an improved Faster R-CNN, in Proc. SAR Big Data Era, 2020, pp. 16.

20. D. N. S. Thatavarthi, Marine safety enhancement using Faster R- CNN with VGG16 for ship detection from airborne radar signals, Proposed Work, 2026.