Student Research Associate — Engineering Laboratory, IIT Kanpur

Disaster Damage Rating-based Infrastructure Financing Framework

Jan 2023 – Apr 2023

---

Building Damage Detection from Satellite Imagery

---

Objective

• Estimate financial aid needs after disasters by predicting damage severity from paired pre/post satellite imagery.

• Assemble a clean, generalizable dataset so models learn from useful pixels and work reliably across different regions.

Approach

• Preprocessed satellite images using OpenCV and Pillow: masked building footprints, tiled images into grids using Patchify, and filtered tiles under 5% building area to remove noise

• Benchmarked ResNet, VGG16/19, and InceptionV3 architectures; tuned hyperparameters (learning rate, batch size, dropout) and mapped damage grades to financial aid bands

• Validated model outputs and documented all preprocessing steps to ensure pipeline reproducibility for rapid deployment in new regions

Impact

• Selected InceptionV3 as the final model with 85.6% accuracy for damage classification

• Produced district-level aid estimates to inform disaster relief planning and resource allocation

• Paired pre/post views (2 images per sample) significantly boosted robustness and enabled rapid triage of high-impact zones

---

Key Learnings

---

• Dataset quality drives model performance — careful preprocessing and filtering unlock generalization across regions

• Paired imagery provides strong signals — before/after views capture damage severity more reliably than single snapshots

• Reproducibility is critical for research impact — documented pipelines enable rapid deployment in new scenarios

• Systematic benchmarking helps identify the right architectural fit for the problem

---