Posts
Writeups on things I've built and learned:
Projects
Things I've built:
- OOD Detection under Domain Shift Energy-based out-of-distribution detection with Neural Collapse structure (ETF classifiers) on CIFAR-100 and PACS. Energy gap improved from 0.03 → 0.43, AUROC 0.74 and climbing.
- GlacierHack: Glacier Segmentation Ranked #1 at IEEE GRSS GlacierHack 2025 (0.759 MCC). UNet++ with EfficientNet-B4 on a 21-image training set.
Research
Work from my internship at Gödel Earth and research at LNMIIT:
- SAR-to-Optical Translation & Segmentation Fine-tuned E3Diff (192M diffusion model) for cross-modal SAR → optical synthesis, distilling 1000 steps into a single DDIM pass. Extended into a multi-task network for joint image synthesis and 4-class land-cover segmentation. Benchmarked 9 EO foundation models across 71 sites in 5 countries — SegFormer led at 70.5% mIoU.
- Cloud Removal via SAR-Gated Inpainting (OLMoEarth) Adapted OLMoEarth (119M multimodal ViT) for cloud removal under ~45% occlusion. Built a SAR change detection gate to selectively blend clean historical pixels with neural synthesis — reduced MAE by 30% (29.6 dB PSNR).
- InSAR Sand Dune Detection Built a 5-gate statistical pipeline to detect aeolian sand accumulation threatening solar PV infrastructure across the Thar Desert, using Sentinel-1 coherence stacks. Validated against ERA5 wind vectors via 2D FFT — angular alignment error of 1.7°.
- Few-Shot Remote Sensing Classification (RGFS) Coupled prototypical metric learning with a masked reconstruction auxiliary objective to prevent prototype collapse. Accepted at IEEE InGARSS 2025.
Info
B.Tech–M.Tech Integrated in Computer Science at LNMIIT, Jaipur (2023–present). Working at the intersection of generative vision, multimodal fusion, and Earth observation.
Interests: Diffusion Models, SAR Remote Sensing, Multimodal Fusion, Out-of-Distribution Robustness
Skills: PyTorch, Hugging Face, ONNX, Modal, FastAPI, Rasterio, Python, Git, Linux
Contact: jaiswalmj09@yahoo.com