Challenges Solved
- Super-Resolution Pipeline: Engineered a 4x super-resolution pipeline (SR-GAN) for hyperspectral satellite imagery.
- Training Acceleration: Achieved 30x training acceleration on 1TB+ datasets via CUDA-level optimizations and multi-GPU distribution.
- Data Preprocessing: Optimized data preprocessing in C++/Python to handle spectral resampling bottlenecks.
Signal
HPC / CUDA Optimization / Research Rigor
Research Overview
Conducted at the Space Applications Centre (SAC), ISRO, this research addresses the resolution limitations of hyperspectral sensors. By applying Deep Learning to satellite imagery, we can extract significantly more detail for environmental and agricultural monitoring.
Methodology
- SR-GAN Architecture: Utilized a Generative Adversarial Network with a focus on spectral consistency, ensuring that the upscaled imagery maintains the original hyperspectral signatures.
- HPC Utilization: Leveraged multi-GPU clusters and high-performance computing resources to handle the massive data volume.
Technical Depth
- CUDA-Level Optimization: Wrote custom CUDA kernels to accelerate specific compute-intensive operations in the hyperspectral processing chain.
- Memory Management: Implemented advanced data shuffling and memory-mapped IO to prevent bottlenecks during 1TB+ dataset training.
- Quality Metrics: Evaluated performance using both traditional (PSNR, SSIM) and domain-specific (Spectral Angle Mapper) metrics.
Links