AWS-Based Deployment for a Solar Plant Detection from Satellite Imagery

Overview

This repository implements an end-to-end Machine Learning pipeline for automating the detection of solar plants across Brazil using Sentinel-2 satellite imagery. The system automates data acquisition, image enhancement, model inference, and report generation, leveraging AWS services for scalable and cost-effective deployment.

System Architecture

The workflow is orchestrated using AWS Step Functions, with AWS Lambda functions and jobs executed via AWS Batch. It includes the following components:

1. User Interface (UI): A Flask backend and a simple web application allowing users to input coordinates and define search parameters.
2. Acquisition Module (BDC_Acquisition): Fetches Sentinel-2 imagery from Brazil Data Cubes using an AWS Lambda function.
3. Image Enhancement Module: Processes raw images to enhance clarity using advanced filtering and bi-cubic interpolation techniques.
4. Prediction Module: Uses a U-Net deep learning model to perform semantic segmentation for solar plants detection.
5. Report Generation: Generates composite overlays by mosaicking input and predicted images, accessible via a downloadable report.

Repository Structure

AWS-ML_Satellite_Deployment/
│   .gitignore
│   README.md
│
├── models/
│   ├── Model_inference.ipynb   # Test Notebook for running the trained model inference
├── doc/
│   ├── Deployment_Plan.pdf 
│   ├── Deployment_Solution_Architecture.pdf
├── src/
│   ├── acquisition/
│   │   ├── BDC_Fetch.py        # Fetches images from Brazil Data Cube
│   │
│   ├── AWS.settings/
│   │   ├── step-function-definition.json  # AWS Step Function configuration (orchestrator)
│   │   ├── trust-policy.json  # IAM Trust Policy for AWS Batch roles
│   │
│   ├── detection/
│   │   ├── Dockerfile  # Docker setup for the prediction module
│   │   ├── prediction.py  # Python app for the prediction Docker module
│   │
│   ├── enhancement/
│   │   ├── Dockerfile  # Docker setup for image enhancement module
│   │   ├── Image_Enhancement.py  # Python app for the image-enhancement Docker module
│   │
│   ├── report/
│   │   ├── Dockerfile  # Docker setup for report module
│   │   ├── report.py  # Python app for the report Docker module
│   │
│   ├── UI/
│   │   ├── app.py  # Flask backend server
│   │   ├── zappa_settings.json  # settings for the deployment tool
│   │   ├── templates/index.html  # JavaScript / HTML frontend for UI
│   │
│   └── utils/
│       ├── utils.py  # Utility functions for data handling
│       ├── transaction_id_gen/
│       │   ├── counter.txt  # transaction IDs counter file
│
└── tests/
    ├── BDC_Fetch.ipynb  # Tests the Brazil Data Cube acquisition
    ├── ID_Gen.ipynb  # transaction ID generation
    ├── Image_Enhancement.ipynb  # Tests image processing module
    ├── Report.ipynb  # Tests report generation module
    ├── median_composite.tif  # Sample processed satellite image

Deployment Workflow

The system is deployed using AWS Step Functions, connecting multiple AWS services:

1. User Input (API Gateway + AWS Lambda): The user provides latitude, longitude, and distance parameters. Frontend: JavaScript + HTML | Backend: Flask + Python.
2. Data Acquisition (AWS Lambda): Fetches Sentinel-2 data.
3. Image Processing (AWS Batch Job): Enhances satellite images.
4. Model Prediction (AWS Batch Job): Identifies solar panels.
5. Report Generation (AWS Batch Job): Creates overlays and statistics.
6. Results Retrieval (S3 & Flask UI): Users can download reports via the web app.

Running the System

The demo application is hosted on AWS. Since it uses paid cloud resources, access is limited to help manage operating costs. However, if you're interested in trying the tool, feel free to request access. I’ll be happy to share an access token.

👉 Request Access to the Demo

🔑 How It Works

1. Visit the link and fill out a short form with your details.
2. You’ll receive an access token and a secure link via email shortly after.

Feel free to connect with me by e-mail or on LinkedIn if you have any questions or feedback!

📧 Email: federico.bessi@gmail.com

💼 LinkedIn: Federico Bessi

🌍 Sample Results from Satellite-Based Detection

This project includes a semantic segmentation model that analyzes Sentinel-2 satellite images to detect photovoltaic solar plants. Below are sample outputs showing the effectiveness of the model.

The left images show the input variables. The right images displays the model’s prediction in a overlay image highlighting the detection in cyan color.

🛰️ Example 1: Taboquinha DF Solar Plant

🛰️ Example 2: Hélio Valgas Solar Plant

Future Improvements

• Optimize inference speed

  • Cache intermediate results: Store BDC downloads and post-processed images in S3 to avoid redundant downloads and processing.
  • Evaluate GPU upgrade: Test faster types (e.g., g5.xlarge, p3.2xlarge) to accelerate inference, considering budget constraints.

• Post-processing improvements: Use connected component filtering to remove very small predictions, filtering noise with area-based rules.

• UI improvements: Improve the web interface for a smoother user experience, including better feedback during long processing times.

• Add time filtering: Add the ability to filter satellite images by acquisition date for more refined temporal analysis.

• Adopt S2 Cell tiling: Replace the current custom algorithm with an S2 Cells-based approach for better geographic indexing and scalability.