Artificial intelligence research · IIT Kharagpur · 2021–2022
Geo AI
A research internship comparing deep-learning approaches for volcanic-deformation classification using Sentinel-1 SAR imagery and PSInSAR-derived data. The emphasis is on experiment design and model comparison, with the reported result kept in context.
Role & context
- Contribution
- Research internship: data preparation, model comparison, training experiments, and evaluation updates.
- Setting
- IIT Kharagpur · research internship · June 2021–January 2022
- Public scope
- Research experience, separate from Khalid’s KIIT degree. The public case does not include the complete experiment archive.
The problem
Satellite interferometry produces complex spatial patterns while useful labeled examples are limited. The research compared architectures and training strategies without assuming that one model family would automatically be best.
A closer look
Comparing models when labeled examples are limited
- Constraint
- Interferometric imagery contains complex spatial patterns, while the available labeled examples were limited. Architecture choice alone could not resolve that constraint.
- Decision
- Khalid compared AlexNet, ResNet, and autoencoder approaches, with augmentation, fine-tuning, and pretrained ImageNet weights used to explore limited-data conditions.
- Tradeoff
- Transfer learning and augmentation offered ways to work with the available data, but evaluation still needed to distinguish model behavior from dataset and split effects.
- Evidence boundary
- The internship reported 90%+ classification accuracy. Dataset size, class balance, split methodology, and a reproducible evaluation archive are not included here, so that result should not be read as an independently validated generalization claim.
What Khalid built
Khalid developed volcanic-deformation classifiers using Sentinel-1 SAR imagery and PSInSAR-derived data under Prof. Debashish Chakravarty.
He implemented and compared AlexNet, ResNet, and autoencoder approaches for geospatial image classification.
The experiments reported 90%+ classification accuracy and tested augmentation, fine-tuning, and pretrained ImageNet weights for limited-data conditions.
Weekly updates covered metrics, confusion matrices, model comparisons, and research findings.
Public architecture
Explore the architecture
- SAR imagery
Sentinel-1 and interferometric imagery provide the raw spatial signal.
- Preparation
The data is prepared into consistent model inputs while preserving meaningful spatial patterns.
- Model candidates
Several CNN approaches are trained and compared rather than assuming a single architecture.
- Limited-data strategy
Augmentation and transfer learning help make better use of a constrained labeled dataset.
- Evaluation
Metrics and confusion matrices are used to compare behavior and identify errors.
- Classification
The final output is a deformation classification backed by the experimental comparison.
Public abstraction. Proprietary implementation details are omitted.
Design decisions
Signal preparation
Interferometric imagery is transformed into model-ready examples while retaining useful spatial structure.
Architecture comparison
Multiple CNN families are benchmarked rather than treating one architecture as a foregone conclusion.
Limited-data strategy
Augmentation and transfer learning reduce dependence on a large bespoke labeled corpus.
Evaluation
Accuracy, confusion matrices, and comparative experiments guide iteration.
Evidence & limits
How to read the reported accuracy
Weekly research updates included metrics, confusion matrices, and model comparisons. The reported 90%+ accuracy belongs to those internship experiments. The public portfolio does not provide enough experimental context to compare it with a benchmark or infer performance on new regions or time periods.
Takeaway
The research established a habit that still carries into engineering work: compare evidence and failure modes before getting attached to a solution.