☀️ Solar Flare Detection from NASA SDO Imagery using Deep Learning

Automated solar flare detection from NASA Solar Dynamics Observatory (SDO) imagery using ConvNeXt Large and transfer learning.
Trained on SDOBenchmark | 86.81% F1 | 89.57% ROC-AUC | 90.51% Recall | Grad-CAM Explainability

---

📋 Table of Contents

• Overview
• Dataset
• Methodology
• Model Architecture
• Results
• Baseline Comparison
• Grad-CAM Explainability
• Limitations
• Project Structure
• How to Reproduce
• Tech Stack
• Disclaimer

---

🔭 Overview

Solar flares are intense bursts of electromagnetic radiation from the Sun's surface. Strong flares (M and X class) can:

• Knock out satellite communications
• Disrupt GPS systems
• Damage power grid infrastructure
• Irradiate astronauts in space

Early and accurate prediction of solar flares is critical for space weather forecasting and protecting Earth's infrastructure.

This project builds an AI-powered solar flare detection system that classifies NASA SDO satellite images of solar active regions into:

Class	Description
Flare	Active region will produce a C-class or stronger flare (peak_flux ≥ 1×10⁻⁶ W/m²)
No-Flare	Active region remains below flare threshold

The model achieves 90.51% Recall on the official SDOBenchmark test set — meaning it correctly identifies 9 out of 10 flare-producing active regions.

---

📊 Dataset

SDOBenchmark — Institute for Data Science, FHNW Switzerland

The dataset contains multi-channel satellite imagery from NASA's Solar Dynamics Observatory (SDO), specifically from the Atmospheric Imaging Assembly (AIA) and Helioseismic Magnetic Imager (HMI) instruments.

Property	Value
Total Samples	9,222
Training Samples	8,336
Test Samples	886
Image Channels	10 per timestep (131Å, 171Å, 193Å, 211Å, 304Å, 335Å, 94Å, 1700Å, continuum, magnetogram)
Timesteps per Sample	4
Image Resolution	256 × 256 pixels
Label	peak_flux (regression) → binary classification

Binary Label Threshold:

• Flare (1): peak_flux ≥ 1×10⁻⁶ W/m² (C-class or above)
• No-Flare (0): peak_flux < 1×10⁻⁶ W/m²

Data Split Strategy:

Split	Samples	Flare	No-Flare
Train	6,633	2,667	3,966
Validation	1,703	738	965
Test (Official)	886	527	359

Critical: Split was performed on Active Region (AR) number — not on individual samples. This prevents data leakage where the same active region appears in both train and test sets, which would artificially inflate performance metrics.

Label Distribution

Sample Images

---

⚙️ Methodology

1. Channel Selection

Each sample contains 10 imaging channels across 4 timesteps (40 images total). We selected the 3 most physically meaningful channels as RGB input:

Channel	Wavelength	Physical Significance
171Å	EUV	Coronal loops at ~1MK — primary flare indicator
193Å	EUV	Hot plasma at ~1.5MK — flare precursor structures
Magnetogram	HMI	Photospheric magnetic field — root cause of flares

These 3 channels are stacked to form a 3-channel RGB-like input compatible with ImageNet pretrained weights.

2. Temporal Strategy

Each sample has 4 timesteps. We use the last timestep — the observation closest to the flare prediction window — maximizing the information available before the event.

3. Class Imbalance Handling

Training data has a 1.49× imbalance (No-Flare vs Flare). Two complementary strategies:

• WeightedRandomSampler — oversamples Flare class during training
• Class-weighted loss — applies 1.24× higher penalty for missing Flare predictions

4. Gradual Unfreezing

• Epochs 1–2: Backbone frozen — only classification head trains
• Epoch 3+: Full model unfrozen with differential learning rates

5. Differential Learning Rates

• Backbone: 1×10⁻⁵ (preserve pretrained ImageNet features)
• Head: 1×10⁻⁴ (learn solar domain features faster)

6. Threshold Optimization

Default classification threshold (0.5) was replaced with an optimized threshold (0.4261) found using Youden's J statistic on the ROC curve. This improved F1 by +2.17% and Recall by +5.31%.

7. Advanced Augmentation

Domain-specific augmentations for solar imagery:

Augmentation	Reason
Random crop from 256→224	Implicit translation invariance
Horizontal/Vertical flip	Solar disk is rotationally symmetric
ColorJitter	Handles instrumental calibration variations
Affine transforms	Geometric invariance across observation angles
GaussianBlur	Simulates different instrument resolutions

---

🧠 Model Architecture

ConvNeXt Large — modernized CNN matching Vision Transformer accuracy with CNN efficiency.

Property	Value
Architecture	ConvNeXt Large
Pretrained On	ImageNet-22k → ImageNet-1k
Total Parameters	196.23M
Input Resolution	224 × 224
Input Channels	3 (171Å, 193Å, Magnetogram)
Drop Path Rate	0.2 (stochastic depth)
Output Classes	2 (Flare / No-Flare)
Optimizer	AdamW
Scheduler	CosineAnnealingWarmRestarts (T₀=10)
Loss	CrossEntropyLoss + Label Smoothing (ε=0.1) + Class Weights
Mixed Precision	✅ torch.amp
Gradient Clipping	max_norm=1.0
Early Stopping	patience=7 on Val F1
GPU	NVIDIA H100 80GB
Training Time	~10.3 minutes

---

📈 Results

Training Curves

Validation Performance

Metric	Score
Best Val F1	75.08%
Val ROC-AUC	~85.7%
Best Epoch	23

Official Test Set Performance (886 samples)

Reported on the official SDOBenchmark test set — completely held out and never seen during training or validation.

Metric	Default (0.5)	Optimal (0.4261)
F1 Score	84.64%	86.81%
ROC-AUC	89.57%	89.57%
Recall	85.20%	90.51%
Precision	84.08%	83.39%
Accuracy	81.60%	83.63%

Confusion Matrix (Official Test Set)

	Predicted No-Flare	Predicted Flare
Actual No-Flare	264 ✅	95 ❌
Actual Flare	50 ❌	477 ✅

Only 50 flares missed out of 527 total. High recall is the clinical priority — missing a strong solar flare can have catastrophic consequences for satellites and infrastructure.

Test Evaluation

Probability Distribution

---

🏆 Baseline Comparison

All models trained on identical data, loss function, and augmentation pipeline. Only architecture differs.

Model	Val F1	Test F1	Test AUC	Recall	Params	Epochs
EfficientNet-B0	69.16%	78.98%	82.75%	81.97%	~5M	15
ResNet50	73.05%	84.70%	88.56%	88.24%	~25M	15
ConvNeXt Large (Ours)	75.08%	86.81%	89.57%	90.51%	~196M	30

Note: Baselines trained for 15 epochs vs 30 for ConvNeXt Large. A fully converged comparison would likely narrow the F1 gap slightly, but the AUC and Recall advantages of ConvNeXt Large are consistent with its architectural superiority.

---

🔍 Grad-CAM Explainability

Gradient-weighted Class Activation Mapping (Grad-CAM) reveals which regions of the solar active region the model attends to when making predictions.

Target Layer: stages[-1].blocks[-1] — final convolutional block capturing highest-level semantic features.

Correct Predictions

Observations:

• Flare predictions: Model attends to complex coronal loop structures and high-gradient magnetic field regions — exactly the physical precursors identified in heliophysics literature
• No-Flare predictions: Attention is diffuse and scattered — consistent with quieter, less energetically complex active regions

Error Analysis

Understanding failures:

• Missed flares (P(Flare)=0.02-0.04): Attention maps resemble no-flare patterns — these active regions had unusually subdued morphology before erupting. Genuinely ambiguous cases
• False alarms (P(Flare)=0.94-0.96): Complex loop structures and strong magnetic gradients triggered high flare probability — physically reasonable even though no flare occurred in the prediction window

"Grad-CAM analysis confirms the model attends to physically meaningful regions — complex coronal loop structures and high-gradient magnetic field areas — consistent with established solar flare precursor indicators in heliophysics literature."

---

⚠️ Limitations

1. Single channel combination — only 3 of 10 available channels used. Multi-channel fusion or temporal modeling across all 4 timesteps could improve performance.

2. Binary classification only — flares are classified as Flare/No-Flare. Multi-class classification (B/C/M/X intensity levels) is a harder and more clinically useful problem.

3. Temporal information unused — each sample has 4 timesteps but only the last is used. A CNN+LSTM or 3D ConvNet architecture exploiting temporal dynamics could capture flare buildup patterns.

4. Prediction window fixed — SDOBenchmark uses a fixed 24-hour prediction window. Real operational systems need multi-window forecasting (6h, 12h, 24h, 48h).

5. Not an operational system — this model is for research purposes only and has not been validated against real-time NOAA space weather forecasting standards.

---

📁 Project Structure

Solar-Flare-Detection-Using-Deep-Learning/ │ ├── Solar_Flare_Detection_ConvNeXt_Large_SDOBenchmark.ipynb ├── requirements.txt ├── README.md ├── LICENSE │ ├── eda_labels.png ├── sample_images_fixed.png ├── training_curves.png ├── test_evaluation.png ├── probability_distribution.png ├── gradcam_correct.png ├── gradcam_errors.png └── baseline_comparison.png

Model weights (~800MB) stored on Google Drive due to GitHub size limits.

---

🚀 How to Reproduce

Requirements

pip install torch torchvision timm albumentations torchmetrics

Step 1 — Get the Dataset

Download SDOBenchmark from Kaggle: https://www.kaggle.com/datasets/fhnw-i4ds/sdobenchmark

Upload SDOBenchmark_full.zip to your Google Drive root.

Step 2 — Open Notebook

Open Solar_Flare_Detection_ConvNeXt_Large_SDOBenchmark.ipynb in Google Colab.

Step 3 — Select GPU

Runtime → Change runtime type → GPU → A100 or H100

Step 4 — Run Phase by Phase

Phase	Description
Phase 01	Environment setup & GPU verification
Phase 02	Dataset extraction & label analysis
Phase 03	Image pipeline & dataset class
Phase 04	Model architecture & training
Phase 05	Test set evaluation & threshold optimization
Phase 06	Grad-CAM explainability
Phase 07	Baseline comparison

Hardware Used

• GPU: NVIDIA H100 80GB HBM3
• Training time: ~10.3 minutes
• Also runs on A100/L4 (adjust BATCH_SIZE)

---

🛠️ Tech Stack

Library	Version	Purpose
Python	3.12	Core language
PyTorch	2.10	Deep learning framework
TIMM	1.0.26	ConvNeXt Large pretrained model
Albumentations	latest	Image augmentation
scikit-learn	latest	Metrics, class weights, ROC
Matplotlib / Seaborn	latest	Visualization
Google Colab	—	Training environment
Google Drive	—	Dataset and model storage
NumPy / Pandas	latest	Data processing

---

📜 Disclaimer

This project is for educational and research purposes only.

It is not a validated operational space weather forecasting system. For real-time solar flare forecasting, refer to NOAA's Space Weather Prediction Center (https://www.swpc.noaa.gov/).

The SDOBenchmark dataset is publicly available via Kaggle for research use.

---

👤 Author

Niteesh014

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.