Welcome to my Capstone Project repository for the UCSD Machine Learning Engineering Bootcamp, mentored by Artem Yankov (Google). This project aims to develop an end-to-end ML pipeline for the automatic detection of photovoltaic (PV) power stations using satellite imagery and deep learning via semantic segmentation.
Solar energy is one of the fastest-growing renewable energy sources worldwide, especially in regions like Brazil—where vast areas of land are suitable for solar projects. However, tracking and mapping new installations through traditional methods (e.g., manual surveys or scattered records) can be time-consuming and prone to errors. This gap calls for a scalable, automated approach to accurately identify solar power stations, aiding resource planning, policy-making, and environmental impact assessments.
Beyond this practical need, a key motivation for this project is my own curiosity and desire to deepen my understanding of image detection models from both academic and real-world perspectives. By applying advanced deep learning techniques to real satellite imagery, I aim to bridge theoretical knowledge with hands-on technological implementation, contributing to the broader landscape of computer vision applications.
- Create a dataset from scratch using open-source satellite imagery curated for PV power station detection.
- Define a baseline model to establish initial performance benchmarks.
- Develop a deep learning model to identify solar plants across Brazil using satellite data.
- Test multiple popular model architectures (e.g., U-Net, U-Net++, DeepLabV3, PSPNet, and FPN).
- Enhance performance using image preprocessing techniques such as Bicubic Interpolation and Enhanced Super Resolution GANs (ESRGAN).
- Design an end-to-end pipeline, encompassing data ingestion, model training, evaluation, and deployment.
- Deploy the solution in a manner that can be demonstrated or integrated into real-world systems.
Here are a few key points:
We use open images from the Sentinel-2 program available at Brazil Data Cube, which is a research, development, and technological innovation project of the National Institute for Space Research (INPE), Brazil. It produces datasets from large volumes of medium-resolution remote sensing images for the entire national territory. The project also develops a computational platform to process and analyze these datasets using artificial intelligence, machine learning, and image time series analysis.
We created a dataset from scratch with 272 images coming from 16 photovoltaic plants of various distributions, sizes, and surroundings, aiming for good generalization. To expand our samples, we captured multiple random scenes over each plant and applied data augmentation to reduce overfitting.
| Sets | Number of Areas | Number of Images |
|---|---|---|
| Training | 11 | 201 (74%) |
| Validation | 2 | 27 (10%) |
| Test | 3 | 44 (16%) |
Labeling was performed using an ad-hoc “auto-mask” script, followed by manual edits in QGIS.
We achieved the following metrics: IoU = 94.62% and F1-Score = 97.23%, using a U-Net model pre-trained on ImageNet with an EfficientNet-B7 encoder. We also applied bicubic interpolation (2× scaling) to improve resolution.
We obtained our best results using a weighted loss function composed of:
- Binary Cross Entropy
- Focal Loss
- Jaccard Index
We also experimented with Dice Loss, but found the above combination to yield better performance overall.
We tested bicubic interpolation and Enhanced Super Resolution GANs (ESRGAN) as image-enhancement methods over the original dataset.
Below is an overview of the repository and the steps in my process. Each folder contains notebooks, scripts, or related files for that phase:
-
Step 1 - Start Planning my Capstone
Define the possible project scopes and initial feasibility assessment based on research and mentor discussions. -
Step 2 - Collect Your Data
Basic Exploratory Data Analysis (EDA) and data collection strategies. -
Step 3 - Project Proporsal
Formalize the project plan by outlining objectives, methodology, expected outcomes, and potential challenges. -
Step 4 - Survey Existing Research and Reproduce Available Solutions
Literature review and attempts at reproducing known solutions. -
Step 5 - Data Wrangling & Exploration
Data cleaning, feature engineering, and exploratory analysis. -
Step 6 - Benchmark Your Model
Initial benchmarking with a baseline model or known methods. -
Step 7 - Experiment with Various Models
Iterative experimentation with different ML/DL approaches. -
Step 8 - Scale the ML Prototype
Approaches for taking the model from proof-of-concept to production scale. -
Step 9 - AWS Deployment AWS-Based Deployment Implementation.
I’d like to acknowledge the following:
- UCSD ML Engineering Bootcamp — for providing a comprehensive and complete curriculum and unwavering support throughout the learning journey
- My mentor: Artem Yankov (Google) — for his valuable guidance and support
- Open-Source Community — for making data science accessible
- My Family — for supporting me during this learning phase 😅
If you have any suggestions, questions, or just want to connect:
- Email: [email protected]
- LinkedIn: Federico Bessi
This project is licensed under the Apache License 2.0. See the LICENSE file for details.