AI research · IIT Kharagpur · 2021–2022
Deep Learning
Explore meta-analysis and custom deep learning models developed during research at IIT Kharagpur, using Sentinel-1 SAR imagery and PSInSAR data.
Role and context
- Contribution
- Meta-analysis, custom model development, data preparation, training, and evaluation.
- Setting
- IIT Kharagpur · research internship · June 2021–January 2022
- Scope
- Research internship at IIT Kharagpur, separate from the degree at KIIT. The full experiment archive is not included.
The problem
The satellite data contained complex spatial patterns and limited labeled examples. The work compared architectures and training methods and developed custom models for the available data.
A closer look
Training with limited labeled data
- Constraint
- The available labeled examples were limited, and the satellite imagery contained complex spatial patterns.
- Decision
- AlexNet, ResNet, and autoencoder models were compared using augmentation, fine-tuning, and pretrained ImageNet weights.
- Tradeoff
- These methods helped work with limited data, but results still depended on the dataset and evaluation split.
- Not included
- The experiments reported 90%+ accuracy. Dataset size, class balance, split details, and a reproducible evaluation archive are not included, so this is not an independently validated benchmark.
Explore the implementation
Research under Prof. Debashish Chakravarty covered meta-analysis and custom deep learning models for volcanic-deformation classification.
Model development and comparison covered AlexNet, ResNet, and autoencoders using Sentinel-1 SAR and PSInSAR data.
The experiments reported 90%+ accuracy and tested augmentation, fine-tuning, and pretrained ImageNet weights.
Weekly research reports covered metrics, confusion matrices, model comparisons, and findings.
Architecture overview
View architecture
- SAR imagery
Sentinel-1 radar and interferometric imagery provide the source data.
- Preparation
Images are prepared as consistent model inputs while preserving spatial patterns.
- Model candidates
AlexNet, ResNet, and autoencoder models are trained and compared.
- Limited-data strategy
Augmentation and transfer learning help use the limited labeled data.
- Evaluation
Metrics and confusion matrices help compare models and inspect errors.
- Classification
The models classify volcanic deformation in the imagery.
Select a step to explore the flow.
Design decisions
Data preparation
Radar imagery is prepared as model inputs while preserving spatial patterns.
Model comparison
AlexNet, ResNet, and autoencoder models are trained and compared.
Training with limited data
Augmentation and transfer learning help make use of the available labeled examples.
Evaluation
Accuracy and confusion matrices are used to compare models and inspect errors.
Results & limits
About the reported accuracy
The 90%+ accuracy was reported in internship experiments, alongside confusion matrices and model comparisons. The public summary does not establish performance on new regions or time periods.
Takeaway
Comparing models and inspecting their errors was more useful than relying on accuracy alone.