Scribd
Upload
8 views9 pages

Project PPT 01

The document outlines a project aimed at developing an automated flood detection system using satellite imagery and deep learning techniques, specifically leveraging the Sen1Floods11 dataset and NVIDIA's deep learning frameworks. The project focuses on creating a model that accurately identifies flooded areas to enhance disaster response and environmental monitoring. Key components include data preparation, model training with a ResNet18-based UNet architecture, and real-time inference using Triton Inference Server.

Uploaded by

kiplinkenterprises
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
8 views9 pages

Project PPT 01

The document outlines a project aimed at developing an automated flood detection system using satellite imagery and deep learning techniques, specifically leveraging the Sen1Floods11 dataset and NVIDIA's deep learning frameworks. The project focuses on creating a model that accurately identifies flooded areas to enhance disaster response and environmental monitoring. Key components include data preparation, model training with a ResNet18-based UNet architecture, and real-time inference using Triton Inference Server.

Uploaded by

kiplinkenterprises
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Flood Detection Using Satellite

Imagery

Presented By Under Guidance


Kanishk Mishra 01618011621. Mrs. Anjali Sardana
Assistant Professor(Department of Artificial
Intelligence and Machine Learning)
Introduction
Developing an automated flood detection system using satellite imagery and deep learning provides
critical support for disaster response and environmental monitoring. This project leverages:

➤ Satellite imagery from the Sen1Floods11 dataset containing hand-labeled flood regions
➤ NVIDIA's deep learning frameworks (DALI and TAO) for efficient image processing
➤ UNet architecture with a pre-trained ResNet18 backbone for semantic segmentation
➤ Triton Inference Server for optimized model deployment

The system processes satellite imagery to identify flooded areas with high accuracy, enabling rapid
assessment of flood extent across various geographical regions. This technology has applications in
emergency response coordination, flood risk assessment, and climate change impact monitoring.

Our approach demonstrates the complete AI workflow from data preparation through model deployment
with performance optimization for real-time flood detection capabilities.
Feasibility Study
Technical Viability
● Satellite imagery analysis has proven effective for flood detection in research settings
● Sen1Floods11 dataset contains sufficient labeled examples (spanning multiple regions)
● Semantic segmentation approaches have demonstrated success in similar environmental applications

Resource Assessment
● Data processing requirements are manageable with existing pipeline tools
● Model training complexity is reduced through transfer learning approach
● Inference time meets operational requirements for post-disaster assessment

Cost-Benefit Projection
● Development costs offset by potential reduction in manual assessment time
● Improved response coordination provides significant humanitarian benefit
● Solution can be extended to other environmental monitoring applications
Objective
The objective of this project is to develop an AI-powered flood segmentation model using deep
learning techniques. The model is trained to accurately identify and segment flooded areas in satellite or
aerial images, enabling efficient disaster management and response.

Key goals include:

● Data Preparation & Preprocessing: Organizing and splitting datasets for training and validation.
● Model Training & Optimization: Utilizing NVIDIA TAO Toolkit with a ResNet18-based UNet architecture for
semantic segmentation.
● Inference & Deployment: Running real-time inference using Triton Inference Server to ensure efficient and scalable
deployment.
● Performance Evaluation: Measuring model accuracy and improving throughput using batch processing for
optimized inference speed.

This project aims to support flood monitoring efforts by providing a fast, reliable, and automated
method for flood area detection.
Methodology
Literature Survey
📌 Traditional Flood Detection Methods
🌍 Satellite Imaging & Remote Sensing – Uses Landsat, Sentinel-1, MODIS for flood mapping.
💧 Hydrological Models – Simulates flood extent but requires manual calibration & high computation.

📌 Deep Learning-Based Approaches


🧠 CNNs & UNet – Used for flood area segmentation, offering pixel-wise accuracy.
🔗 ResNet-based UNet – Incorporates residual connections for improved performance.

📌 Pretrained Models & Transfer Learning


🚀 Models like ResNet, VGG, EfficientNet enhance flood detection.
🔄 Transfer Learning boosts accuracy with minimal training data.

📌 Edge Computing & Real-Time Inference


⚡ Triton Inference Server enables real-time flood detection.
📊 Batch Processing improves inference speed and efficiency.

📌 Applications & Impact


🌊 Early Warning Systems – Faster flood detection for better preparedness.
🚑 Disaster Response – Helps emergency teams allocate resources.
🏙 Urban Planning – Supports climate change analysis & flood mitigation.
Hardware and Software Requirements
🖥 Hardware Requirements

✅ GPU – NVIDIA RTX 3090 / A100 (Recommended for faster training & inference)
✅ CPU – Intel Core i7 / AMD Ryzen 7 or higher
✅ RAM – Minimum 16GB (32GB recommended for large datasets)
✅ Storage – At least 500GB SSD (Preferably NVMe for faster read/write)
✅ Other – High-speed internet, CUDA-compatible device

🛠 Software Requirements

💻 Operating System – Ubuntu 20.04 / Windows 10 with WSL2 (Linux preferred)


🧠 Deep Learning Framework – TensorFlow / PyTorch
🖼 NVIDIA TAO Toolkit – For model training and optimization
🚀 Triton Inference Server – For real-time model deployment
📦 Docker – To run TAO models efficiently
🛠 CUDA & cuDNN – GPU acceleration libraries
📊 Matplotlib, OpenCV, PIL – For data preprocessing & visualization
💾 NGC CLI – To download pretrained models from NVIDIA GPU Cloud
Algorithms Used

1. Model Architecture
● Encoder: ResNet-18 extracts spatial features.
● Bottleneck: Retains spatial context.
● Decoder: Transposed convolutions + skip connections
for precise segmentation.
● Activation: ReLU for layers, Softmax for final
classification.

2. Training & Optimization


● Loss Function: Cross-Entropy Loss.
● Optimizer: Adam for faster convergence.
● Metrics: IoU (Intersection over Union), Accuracy.

3. Inference & Deployment


● Convert to TensorRT for optimized speed.
● Deploy on Triton Inference Server for real-time
processing.
Thank You

You might also like