---
language:
- en
license: mit
library_name: keras
tags:
- remote-sensing
- hyperspectral
- uncertainty-quantification
- conformal-prediction
- mamba
- foundation-models
metrics:
- accuracy
- kappa
- coverage
---

# DIAS 65-Band Hyperspectral Model Registry

This repository contains trained neural network checkpoints for the 65-band DIAS benchmark suite, spanning foundational CNNs, state-space models, vision transformers, and multi-modal geospatial foundation models.

## Benchmark Performance Summary (65-Band DIAS)
| Architecture | Test Accuracy | Cohen's Kappa ($\kappa$) | Marginal Conformal Coverage | Checkpoint Files |
| :--- | :---: | :---: | :---: | :--- |
| **SpatialGCN** | **97.05%** | **0.9644** | **95.40%** | `SpatialGCN_best.keras` |
| **MambaHSI** | **99.99%** | **0.9999** | **96.27%** | `MambaHSI_{Small,Base,Large}_best.keras` |
| **DOFA Foundation** | **100.0%** | **1.0000** | **95.15%** | `DOFA_Hiera_Fusion_best.keras`, `Hiera_ViT_best.keras` |
| **GeoRSCLIP** | **100.0%** | **1.0000** | **95.01%** | `GeoRSCLIP_TextHead_best.keras` |
| **AlexNet CNN Baseline** | 94.63% | 0.9352 | 95.10% | `AlexNet_CNN_best.keras` |
| **GFNet Global Filter** | 97.43% | 0.9691 | 95.20% | `GFNet_best.keras` |
| **ViT_UNet Hybrid** | 95.13% | 0.9413 | 95.05% | `ViT_UNet_best.keras` |

## Authors & Attribution
- **Principal Author**: Danesh Selwal (`selwaldanesh@gmail.com`)
- **Co-authored by**: Claude (`claude@anthropic.com`)