Praktikum 04: Computer Vision dengan Roboflow

Tujuan Pembelajaran

Memahami konsep dasar computer vision dan object detection

Mengenal Ground Penetrating Radar (GPR) dan hyperbola detection

Menggunakan Roboflow platform untuk no-code computer vision

Melakukan data preprocessing dan augmentation

Melatih model YOLO untuk object detection

Mengevaluasi performa model dan melakukan inference

Menginterpretasi hasil deteksi dalam konteks geofisika

📚 Pendahuluan

Apa itu Computer Vision & Object Detection?

Computer Vision adalah cabang artificial intelligence yang memungkinkan komputer untuk "melihat" dan memahami konten visual seperti gambar dan video. Object Detection adalah task spesifik dalam computer vision yang bertujuan untuk:

Object Detection Tasks:

1. Localization: Menemukan posisi objek dalam gambar (bounding box)

2. Classification: Mengidentifikasi jenis/kelas objek

3. Confidence: Memberikan skor kepercayaan untuk setiap deteksi

Ground Penetrating Radar (GPR)

GPR adalah metode geofisika non-invasif yang menggunakan gelombang elektromagnetik untuk mencitrakan struktur bawah permukaan. Prinsip kerjanya:

Transmisi: Antena transmitter memancarkan pulsa radar ke dalam tanah
Refleksi: Sinyal dipantulkan oleh objek/interface dengan kontras dielektrik
Penerimaan: Antena receiver menangkap sinyal reflected
Imaging: Data ditampilkan sebagai radargram (x: posisi, y: kedalaman/waktu)

GPR Hyperbola: Signifikansi Geofisika

Ketika GPR melewati objek buried (pipa, kabel, rongga, batuan), reflected signal membentuk pola hyperbola pada radargram. Karakteristik hyperbola mengandung informasi penting:

Informasi dari Hyperbola:

Apex (puncak) → Posisi horizontal dan kedalaman objek
Curvature (kelengkungan) → Kecepatan propagasi gelombang dalam medium
Width (lebar) → Ukuran dan sifat objek
Amplitude (intensitas) → Kontras material properties

Kenapa Deep Learning untuk GPR?

Challenges Manual Interpretation:

✗ Time-consuming untuk dataset besar (ratusan/ribuan profile)

✗ Subjektif - hasil bervariasi antar interpreter

✗ Sulit membedakan hyperbola dari noise/artifacts

✗ Tidak scalable untuk real-time monitoring

Keuntungan Deep Learning:

✓ Deteksi otomatis dan konsisten

✓ Cepat - ribuan images dalam hitungan menit

✓ Dapat belajar pattern kompleks dari data

✓ Scalable untuk monitoring continuous

🚀 Roboflow Platform Overview

Roboflow adalah platform end-to-end untuk computer vision yang memungkinkan workflow tanpa coding:

100% No-Code Workflow

50+ Model Formats

Free Plan Available

API Instant Deployment

Fitur Utama Roboflow

📦 Roboflow Universe

Dataset marketplace dengan 200,000+ public datasets untuk berbagai use cases

🏷️ Annotation Tools

Browser-based labeling untuk bounding box, polygon, segmentation, keypoints

🔄 Preprocessing & Augmentation

50+ augmentation techniques dengan visual preview

🧠 Roboflow Train

One-click training dengan YOLOv8, YOLOv11, YOLO-NAS, dan model lainnya

📊 Evaluation Dashboard

Real-time metrics, confusion matrix, precision-recall curves

🌐 Deployment

Hosted API, mobile SDK, edge deployment (ONNX, TFLite, CoreML)

🔧 Workflow: Step-by-Step Tutorial

FASE 1: Setup & Dataset Acquisition

1 Membuat Akun Roboflow

Buka browser dan navigate ke https://roboflow.com
Klik tombol "Sign Up" di pojok kanan atas
Pilih metode sign up:
- Sign up dengan Google Account (recommended - lebih cepat), atau
- Sign up dengan Email & Password
Pilih plan: "Public Plan (Free)" - sudah cukup untuk pembelajaran
Verify email jika diminta
Complete profile setup (nama, role, use case)

✓ Free Plan Features:

• Unlimited public projects

• 10,000 source images

• Roboflow Train access

• Hosted API deployment

• Community support

2 Navigate ke Roboflow Universe

Setelah login, klik "Universe" di navigation bar atas
Di search bar, ketik: GPR atau gpr hyperbola
Atau langsung buka URL: https://universe.roboflow.com/gpr-uurq0/gpr-liosq
Review dataset page untuk melihat:
- Total images (train/val/test)
- Number of classes
- Example images dengan annotations
- Dataset health score
- Model benchmarks (jika ada)

📊 Dataset Information to Note:

• Total images: [lihat di dataset page]

• Classes: Biasanya 1 class - "hyperbola"

• Format: Object detection (bounding boxes)

• Health score: Indicator kualitas annotations

3 Fork Dataset ke Workspace Pribadi

Di halaman dataset GPR, scroll ke bawah dan cari tombol `Use this Dataset` lalu pilih "Fork Dataset"
Pilih opsi "Fork to my workspace" (bukan download langsung)
Dialog akan muncul, beri nama project: GPR-Hyperbola-Detection
Pilih workspace (default: personal workspace)
Klik "Create Project"
Tunggu proses forking (1-3 menit)
Dataset sekarang ada di workspace pribadi Anda!

⚠️ Alternatif: Jika Fork Tidak Tersedia

Jika tombol "Fork" tidak ada:

1. Download dataset (YOLO format recommended)

2. Create new project di workspace

3. Upload images + annotations manually

4 Explore Dataset di Workspace

Buka project "GPR-Hyperbola-Detection" dari workspace
Tab "Dataset" - lihat overview:
- Total images
- Classes distribution
- Train/Val/Test split ratio
- Annotation statistics
Tab "Images" - browse sample images:
- Click image untuk melihat detail
- Lihat bounding boxes (hyperbola annotations)
- Check annotation quality
Tab "Health Check" - review dataset quality:
- Class balance
- Image size distribution
- Null annotations (images tanpa label)

🔍 Analisis yang Perlu Dilakukan:

Berapa rata-rata hyperbola per image?
Apakah annotation quality konsisten?
Apakah ada missing annotations (hyperbola tidak ter-label)?
Variasi ukuran bounding boxes (small, medium, large objects)
Image resolution - apakah seragam?

FASE 2: Data Preprocessing & Augmentation

5 Generate New Dataset Version dengan Preprocessing

Di project dashboard, klik tombol "Generate"
Atau klik "Versions" → "Create New Version"
Halaman "Generate New Version" akan terbuka dengan 3 tabs:
- Preprocessing - transformasi images
- Augmentation - create variations
- Generate - finalize & create version

Preprocessing Settings

🔧 Recommended Preprocessing for GPR Images

Preprocessing	Setting	Alasan
Auto-Orient	✅ ON	Fix image rotation berdasarkan EXIF metadata
Resize	✅ 640 × 640 px	Standard input size untuk YOLO models (v11, v8, dll)
Resize Method	Stretch to Fit	Atau "Fit (letterbox)" untuk maintain aspect ratio
Grayscale	❌ OFF	GPR sudah grayscale, tidak perlu konversi
Tile Images	❌ OFF	Hanya untuk very large images (>2000px)
Contrast	❌ OFF	Akan handle via augmentation

6 Configure Data Augmentation

Augmentation menciptakan variasi dari training images untuk improve model generalization. Klik tab "Augmentation" untuk configure.

🎨 Recommended Augmentation untuk GPR Dataset

Augmentation	Setting	Alasan
Flip Horizontal	✅ ON	Hyperbola bisa muncul dari arah manapun
Flip Vertical	❌ OFF	Orientasi vertical penting (depth ke bawah)
Rotation	✅ ±10° to ±15°	GPR profile bisa sedikit miring saat acquisition
Crop	❌ OFF	Bisa potong hyperbola, tidak disarankan
Brightness	✅ ±15% to ±20%	GPR intensity bisa bervariasi
Exposure	✅ ±10% to ±15%	Variasi kondisi acquisition
Blur	✅ Up to 1.5px	Simulate noise/low resolution data
Noise	✅ Up to 2% pixels	Realistic GPR data often has noise
Cutout	❌ OFF	Bisa hide hyperbola features
Mosaic	❌ OFF	Tidak cocok untuk GPR spatial continuity

⚙️ Augmentation Multiplier:

Jika dataset KECIL (<100 images): Pilih 3x multiplier

Jika dataset MEDIUM (100-500 images): Pilih 2x multiplier

Jika dataset BESAR (>500 images): Pilih None atau 2x

Augmentation multiplier = berapa kali augmented version dibuat per original image

7 Configure Dataset Split

Klik tab "Generate" (tab terakhir)
Review Preprocessing Summary dan Augmentation Summary
Configure Dataset Split:

📊 Recommended Split Ratios

Standard Split

Train: 70%
Valid: 20%
Test: 10%

RECOMMENDED

Small Dataset Split

Train: 80%
Valid: 15%
Test: 5%

Untuk dataset <100 images

Large Dataset Split

Train: 60%
Valid: 20%
Test: 20%

Untuk dataset >1000 images

Pastikan "Stratified split" ON (untuk balance class distribution)
Review summary - estimasi total images setelah augmentation
Klik tombol "Generate"
Tunggu proses generate (1-5 menit tergantung ukuran dataset)

✓ Dataset Version Created!

Setelah proses selesai, Anda akan memiliki:

• Version 1 (atau version berikutnya jika sudah ada versions sebelumnya)

• Preprocessing applied (resize, auto-orient)

• Augmentation applied (jika enabled)

• Train/Val/Test splits

• Ready untuk training!

FASE 3: Model Training

8 Start Model Training

Di dataset version yang baru dibuat (Version 1), klik tombol "Train with Roboflow"
Pilih "Roboflow Train 3.0" (recommended - fastest & latest)
Halaman training configuration akan terbuka

Model Selection

🧠 Choose Your Model

Roboflow RF-DETR

Fast, real-time model. Highest accuracy on COCO
Needs less data and converges earlier. Slower training
Model Size: Small

RECOMMENDED BY ROBOFLOW

YOLO26

Latest from Ultralytics
End-to-end NMS-free inference
Faster CPU inference
Model Size: Nano

FASTEST CPU

Roboflow 3.0

YOLOv8-compatible
Custom performance enhancements
Balance of speed & accuracy
Model Size: Fast

BALANCED

YOLOv11 Nano

Successor to YOLOv8
Fast, efficient inference
Model Size: Nano
Use case: Edge deployment, efisien

YANG KITA PAKAI

💡 Model Selection Guide:

Untuk praktikum ini: Pilih YOLOv11 dengan Model Size: Nano

• Nano: ringan dan cepat, cocok untuk dataset learning

• Successor YOLOv8 dengan inference lebih efisien

• Jika hasil kurang memuaskan, upgrade ke size Small atau Medium

Training Configuration

Select model: YOLOv11 → Model Size: Nano
Training settings (biasanya auto-optimized):
- Epochs: 100 (dengan early stopping)
- Batch size: Auto (Roboflow optimize based on GPU)
- Image size: 640 × 640 (sesuai preprocessing)
- Optimizer: Adam (default)
Klik tombol "Start Training"
Training job akan masuk queue (biasanya instant jika free GPU available)

⏱️ Training Time Estimates:

Small dataset (<100 images): ~5-10 menit

Medium dataset (100-500 images): ~10-20 menit

Large dataset (>500 images): ~20-40 menit

Note: Free plan menggunakan shared GPU, bisa ada waiting time

9 Monitor Training Progress

Setelah training dimulai, Anda akan diredirect ke Training Dashboard
Dashboard menampilkan real-time metrics:
- Loss curves (Box, Object, Class loss)
- Precision - seberapa akurat predictions
- Recall - seberapa banyak hyperbola terdeteksi
- mAP@0.5 - mean average precision at 50% IoU
- mAP@0.5:0.95 - average mAP across IoU thresholds
Charts update setiap beberapa epochs
Early stopping: Training otomatis berhenti jika validation loss tidak improve selama 10-15 epochs

✓ What to Look For During Training:

Loss decreasing: Training berjalan baik
Train & Val loss converging: Model generalizing well
Precision & Recall increasing: Performance improving
mAP increasing: Overall quality meningkat

⚠️ Warning Signs:

Loss stuck/plateau: Mungkin perlu learning rate adjustment
Val loss increasing while train loss decreasing: Overfitting!
Very low metrics (<50%): Dataset quality issue atau wrong model

FASE 4: Model Evaluation & Testing

10 Review Training Results

Setelah training selesai, tab "Results" akan menampilkan comprehensive metrics
Review Metrics Summary:

~78% Precision (B)

~68% Recall (B)

~68% mAP@0.5 (B)

~32% mAP@0.5:0.95 (B)

📌 Konteks Hasil — YOLOv11 Nano, ~250 Epochs

• Precision ~78%: Dari semua deteksi yang dibuat, 78% benar → false positive masih ada tapi terkontrol

• Recall ~68%: 68% dari semua hyperbola berhasil terdeteksi → ada ~32% yang masih terlewat

• mAP@0.5 ~68%: Performa solid untuk model Nano pada dataset GPR

• mAP@0.5:0.95 ~32%: Wajar untuk Nano — metric ini sangat strict (lokalisasi presisi)

• Loss curves train & val keduanya menurun konsisten → tidak ada overfitting signifikan

Advanced Training Graphs - YOLOv11 Nano GPR Hyperbola

Understanding Metrics — YOLOv11

📊 Loss Components (YOLOv11):

train/box_loss & val/box_loss

Error lokalisasi bounding box (posisi & ukuran). Turun dari ~2.6 → ~1.2 (train) dan stabil ~2.0 (val) → model makin akurat menempatkan box.

train/cls_loss & val/cls_loss

Error klasifikasi kelas objek. Turun dari ~3.0 → ~1.0 (train) dan ~5+ → ~1.1 (val) → model belajar membedakan hyperbola dari background.

train/dfl_loss & val/dfl_loss

Distribution Focal Loss — loss untuk presisi tepi bounding box (fitur baru YOLOv11). Turun dari ~2.5 → ~1.4 (train) dan stabil ~2.3 (val).

📈 Detection Metrics:

Precision = TP / (TP + FP) — dari semua deteksi yang dibuat model, berapa persen yang benar? High precision = sedikit false positives.

Recall = TP / (TP + FN) — dari semua hyperbola yang ada, berapa persen terdeteksi? High recall = sedikit hyperbola yang terlewat.

mAP@0.5 — mean Average Precision pada IoU threshold 50%. Metric utama object detection.

mAP@0.5:0.95 — rata-rata mAP dari IoU 50% sampai 95% (step 5%). Lebih strict, mengukur kualitas lokalisasi box.

Available Charts & Visualizations

Advanced Training Graphs - 10 plot: train/val box_loss, cls_loss, dfl_loss, precision, recall, mAP50, mAP50-95
Confusion Matrix - True/False Positives/Negatives breakdown per class
Precision-Recall Curve - Trade-off visualization dengan area under curve
F1 Curve - F1 score vs confidence threshold

11 Deploy & Test Model dengan Roboflow Workflows

Roboflow menggunakan sistem Workflows untuk deploy dan test model secara visual. Workflow mendefinisikan pipeline deteksi dari input image hingga output annotated.

Setelah training selesai, klik tab "Deploy" di navigation
Pilih "Deploy your model with Workflows"
Dialog "Popular Templates" akan muncul dengan pilihan:
- Detect, Count, and Visualize ← pilih ini (default & paling lengkap)
- Detect and Classify
- Small Object Detection (SAHI)
- Text Recognition
Pilih template "Detect, Count, and Visualize" — workflow otomatis terbentuk dengan nodes:
- Inputs → terima image input
- Object Detection Model (gpr-liosq-j4vtr/1) → jalankan deteksi
- Bounding Box Visualization (detection_visualization) → gambar bounding boxes
- Property Definition (count_objects) → hitung jumlah deteksi
- Label Visualization (annotated_image) → tambah label teks
- Outputs (3 outputs) → kirim hasil
Di panel kanan, pastikan model gpr-liosq-j4vtr/1 sudah terpilih (status: Selected)
Klik tombol "Create Workflow"

Deploy your model with Workflows - Detect, Count, and Visualize

🔄 Cara Kerja Workflow Pipeline:

Inputs → Object Detection Model → Bounding Box Visualization + Property Definition → Label Visualization → Outputs

• Model node menjalankan YOLOv11 Nano pada setiap image yang masuk

• Visualization nodes menggambar hasil deteksi (bounding boxes + label) secara otomatis

• count_objects menghitung total hyperbola yang terdeteksi per image

• 3 outputs: annotated image, raw detections (JSON), jumlah objek

✅ Setelah Workflow Dibuat — Cara Test:

1. Buka workflow yang sudah dibuat

2. Klik "Try it" atau "Run" untuk upload test image

3. Drag & drop GPR image ke input node

4. Workflow otomatis memproses dan menampilkan annotated result

5. Review: bounding boxes hyperbola + jumlah deteksi per image

Hasil Inference — "Test Your Workflow"

Setelah workflow berjalan, Roboflow menampilkan dua panel hasil: Visual (radargram dengan bounding box) dan Output (raw JSON detection data).

Workflow Test Result - GPR Hyperbola Detection

✅ Contoh Hasil Deteksi Aktual (dari workflow di atas):

Visual panel: GPR radargram dengan 1 bounding box ungu menandai hyperbola di area shallow (~2–4 nsec)

Output JSON:

                    // count_objects

                    "t_objects": 1,

                    // predictions[0]

                    "width": 69, "height": 54,

                    "x": 137.5, "y": 32,

                    "confidence": 0.639,

                    "class": "Hyperbola",

                    "class_id": 0

• 1 hyperbola terdeteksi dengan confidence 63.9%

• Bounding box: lebar 69px × tinggi 54px, center di (137.5, 32)

• Image size: 249 × 203 px

📖 Cara Membaca Output JSON Workflow:

t_objects (count_objects) → jumlah total hyperbola yang terdeteksi

x, y → koordinat center bounding box (pixel)

width, height → ukuran bounding box (pixel)

confidence → skor keyakinan model (0–1), nilai 0.639 = 63.9%

class → nama kelas yang terdeteksi ("Hyperbola")

detection_id → ID unik setiap deteksi untuk tracking

FASE 5: Geophysical Interpretation

12 Interpretasi Hasil Deteksi dalam Konteks Geofisika

Setelah model mendeteksi hyperbola, kita perlu extract informasi geofisika yang berguna dari bounding boxes.

Dari Bounding Box ke Parameter Geofisika

1. Depth Estimation (Estimasi Kedalaman Objek)

Parameter: Y-coordinate dari bounding box center (apex hyperbola)

Konversi:

• Y-pixel → Two-way travel time (TWTT)

• TWTT → Depth menggunakan velocity

Formula:

Depth (m) = (Velocity × TWTT) / 2

Example:

Apex at y=200 pixels, time scale 0.5 ns/pixel, velocity 0.1 m/ns:

TWTT = 200 × 0.5 = 100 ns

Depth = (0.1 × 100) / 2 = 5 meters

2. Velocity Estimation (dari Hyperbola Curvature)

Parameter: Width dari bounding box relatif terhadap depth

Konsep:

• Wide hyperbola → Low velocity material

• Narrow hyperbola → High velocity material

Relationship:

Velocity ∝ 1 / (Hyperbola Width)

Untuk precise velocity: perlu hyperbola fitting algorithm (diffraction curve matching)

3. Object Size Estimation

Parameter: Amplitude dan width dari hyperbola

Interpretation:

• Large amplitude → Strong contrast (metal pipe, large void)

• Weak amplitude → Subtle contrast (soil boundary, small object)

• Horizontal width → Lateral extent of object

Multiple Hyperbola Analysis

📊 Interpretasi Multi-Detection:

Scenario 1: Vertical alignment

Multiple hyperbolas di x-position sama tapi depth berbeda → Multiple interfaces/layers

Scenario 2: Horizontal pattern

Hyperbolas dengan spacing regular → Buried utility network (pipes, cables)

Scenario 3: Clustered detections

High density hyperbolas di area tertentu → Zone of interest (archaeological site, contaminated zone)

Real-world Applications

🏗️ Utility Mapping

Deteksi pipa, kabel, underground infrastructure sebelum excavation untuk avoid damage

🏛️ Archaeology

Locating buried artifacts, structures, atau features tanpa invasive digging

🛣️ Pavement Assessment

Detect voids, delamination, atau defects dalam aspal dan concrete roads

⚠️ Environmental Hazard

Identifying buried drums, tanks, contamination plumes di former industrial sites

🌳 Tree Root Mapping

Non-destructive root system imaging untuk urban forestry management

🏔️ Geological Mapping

Subsurface stratigraphy, bedrock depth, fault detection

💡 Best Practices & Tips

Dataset Quality Checklist

Annotations konsisten dan akurat (hyperbola apex included dalam box)

Minimum 100 annotated images untuk training yang baik

Dataset mencakup variasi: different depths, sizes, noise levels

Class distribution balanced (tidak ada severe imbalance)

Image quality acceptable (not too blurry or low resolution)

No duplicate images antara train/val/test splits

Augmentation Strategy

⚙️ Start Conservative, Iterate:

First iteration: Minimal augmentation (flip horizontal, brightness ±10%)

Evaluate results: If underfitting → add more augmentation

Second iteration: Add rotation ±15°, blur, noise

Avoid: Augmentations yang destroy GPR physics (vertical flip, extreme crops)

Troubleshooting Common Issues

Problem	Possible Cause	Solution
Low Precision (<70%)	Too many false positives - model detects noise as hyperbola	• Increase confidence threshold • Add more negative samples (images without hyperbolas) • Review annotation quality
Low Recall (<70%)	Missing many hyperbolas	• Decrease confidence threshold • Train for more epochs • Use larger model (Medium instead of Small) • Check if small hyperbolas are annotated
Overfitting (high train, low val)	Model memorizing training data	• Add more augmentation • Reduce training epochs • Increase dataset size • Use smaller model
Poor performance overall	Dataset quality issues	• Review annotations for errors • Ensure sufficient training data (>100 images) • Check image quality • Verify class balance
Training stuck/not converging	Learning rate or model architecture issue	• Try different model (YOLOv11 instead of v8) • Check for data loading errors • Contact Roboflow support if persists

Model Selection Guide (Quick Reference)

🎯 When to Use Which Model? (Tersedia di Roboflow)

Use Case	Recommended Model	Rationale
✅ Praktikum ini	YOLOv11 Nano	Fast & efficient, cocok untuk learning & dataset kecil
Highest accuracy on COCO, sedikit data	Roboflow RF-DETR (Small)	Converges earlier, cocok kalau data terbatas
Balance speed & accuracy, production	Roboflow 3.0 (Fast)	YOLOv8-compatible dengan custom enhancements
Fastest CPU inference, edge device	YOLO26 (Nano)	NMS-free inference, faster on CPU
Small object detection (tiny hyperbolas)	YOLOv11 (Small/Medium)	Upgrade size jika Nano kurang akurat

🎓 Summary & Next Steps

✅ Apa yang Sudah Dipelajari:

✓ Computer vision & object detection fundamentals

✓ Ground Penetrating Radar dan hyperbola detection

✓ Roboflow platform - end-to-end no-code workflow

✓ Dataset preprocessing & augmentation strategies

✓ YOLOv11 Nano model training & evaluation (~250 epochs)

✓ Deploy & inference via Roboflow Workflows (Detect, Count, and Visualize)

✓ Geophysical interpretation dari detection results

Langkah Selanjutnya

🔬 Experiment Further

• Try different augmentation combinations
• Compare YOLOv11 Nano vs Small vs Medium performance
• Test model pada GPR data dari different sites

📊 Advanced Analysis

• Extract bounding box coordinates untuk depth profiling
• Implement velocity analysis workflow
• Create hyperbola density maps

🚀 Integration

• Integrate Roboflow API ke GPR processing pipeline
• Build custom web app untuk GPR interpretation
• Deploy ke edge device untuk field use

📚 Explore Other Datasets

• Roboflow Universe: 200,000+ datasets
• Try seismic facies detection
• Satellite imagery analysis untuk geohazards

📖 Referensi

Platform & Documentation

Roboflow Documentation: https://docs.roboflow.com
Roboflow Universe: https://universe.roboflow.com
YOLO Documentation: https://docs.ultralytics.com

GPR & Geophysics

Jol, H. M. (2008). Ground Penetrating Radar: Theory and Applications. Elsevier.
Daniels, D. J. (2004). Ground Penetrating Radar (2nd ed.). IET.
Annan, A. P. (2009). "Electromagnetic Principles of Ground Penetrating Radar." In Ground Penetrating Radar: Theory and Applications.

Deep Learning for GPR

Jazayeri, S., et al. (2023). "Deep learning for GPR-based utility mapping." Automation in Construction, 145, 104642.
Kim, N., et al. (2020). "Vision-Based Object Detection and Tracking for Autonomous Navigation of Underwater Robots." Sensors, 20(5), 1588.
Zhang, Y., et al. (2021). "Automatic detection of underground objects using GPR with deep learning." NDT & E International, 123, 102511.