Scribd
Upload
79 views24 pages

Solar Panels Presentation UC3M

The document discusses detecting solar panels in satellite images using computer vision techniques. It describes gathering satellite and map images, training object detection and image segmentation models, and evaluating the models' performance at detecting solar panels. Several state-of-the-art models were trained and compared, including YOLO and UNet variants. A live demo was also created using Streamlit.

Uploaded by

Ahmad Zezo
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
79 views24 pages

Solar Panels Presentation UC3M

The document discusses detecting solar panels in satellite images using computer vision techniques. It describes gathering satellite and map images, training object detection and image segmentation models, and evaluating the models' performance at detecting solar panels. Several state-of-the-art models were trained and compared, including YOLO and UNet variants. A live demo was also created using Streamlit.

Uploaded by

Ahmad Zezo
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/ 24

Solar Panel Detection

in Satellite Images

Coordinator: David Delgado Gómez

Sergio Aizcorbe Pardo · Ricardo Chávez Torres ·


Daniel de las Cuevas Turel · Zijun He · Sergio Hidalgo López
Our team

Sergio Ricardo Daniel de las Sergio


Zijun He
Aizcorbe Pardo Chávez Torres Cuevas Turel Hidalgo López

Coordinator: David Delgado Gómez


Project Design
Feature engineering
● Obj. Det: YOLOv5 models (S, M, L, X)
vs ● Future Improvements
Neural Networks ● Img. Seg: Unet++, DeepLabV3+, … ● Live Demo

1 3 5

Problem Data Gathering Evaluation &


Model Training Results
Approach & Processing Performance

2 4
● Raw Satellite Data ● Model comparison
● Google Maps Images
● Result analysis
● Solar Panel Dataset
1. Introduction
Motivation
➢ Solar energy growth

➢ Resource optimization for energy companies

Goal
➔ Detect the presence of solar panels

➔ Approximate the surface of the cells


Feature Engineering

Input Feature Modeling Detection


extraction

👍 Insights of the data

👎 Not scalable and tends to overfit.


Why Convolutional Neural
Networks (CNNs)?

👍 Extract relevant features automatically 👎 Time and computational cost


👍 Better at generalization 👎 Large amount of training data needed
👍 Better with noisy and variable data 👎 Bad at handling rotation and
scale (data augmentation needed)
2. Data gathering
2.1 Raw Satellite data
Copernicus is the European Union's Earth
observation programme

⤬ No labeled data
✓ Multi-spectral information Sentinel-2 Satellite
⤬ Low resolution images
SNAP Toolbox

Madrid
2.2 Solar Panels Datasets
Multi-resolution dataset for
photovoltaic panel segmentation
from satellite and aerial imagery
PV01 (2021)
(0.1m)

PV03
(0.3m)
✓ Spatial Resolution to
compute the AREA
PV08 ✓ Ground truth masks
(0.8m)
⤬ Not much data
2.3 Google Maps
Compute the coordinates
(x1, y1)
and split them into
multiple processes

{
1280 px

zoom

1280 px

(x2, y2)
25 tiles of
256 x 256 pixels Convert to PNG
and save
⤬ No labels
✓ Real live scenario
⤬ Noise in some images
Data
pre-processing 80%
~2500
images
Model Training
& Validation
Solar Panels ~700
images 20%
Labeled Dataset

Google Maps
Images ~500 + 1200
images

Evaluation
3. Object Detection
vs Image Segmentation

Most optimal 👍 Generally, faster.


OBJECT
bounding box 👎 No information
DETECTION around the object about the shape

Most optimal 👍 Shape information


IMAGE of the object.
mask
SEGMENTATION around the object 👎 Slower to deploy.
3.1 Models: Object Detection

YOLO (You Only Look Once)


➔ YOLOv5 S (7.2M param.)
➔ YOLOv5 M (21.2M param.)
➔ YOLOv5 L (46.5M param.)
➔ YOLOv5 X (86.7M param.)

4 models
trained
3.2 Models: Image Segmentation
Unet++ FPN

4 models
PSPNet DeepLabV3+
trained

~20M param/model
4. Evaluation and results
Metrics

Image
Jaccard
Score (IoU)
0.89

Ground Truth
Dice Score 0.94

Prediction
(F1-Score)
Object Image
Detection Segmentation
False Negatives False Positives

➔ Unseen types of solar panels ➔ Zebra crossings w/ shadows


➔ Noise patterns in images ➔ Tennis courts
5. Streamlit Live Demo
Future
improvements Specific models
for each panel

Gathering
more data
Raw
Satellite Data
Detección de Objetos

YOLOv5 (S, M, L, X)

Model Training
& Validation Segm. de Imágenes

Unet++, FPN, DLV3+, PSPNet

Evaluation
THANK YOU!

You might also like