Vyom-205M: A Sovereign Multi-Resolution Spatio-Temporal Foundation Model for Smallholder Cadastral Delineation, SAR-to-Optical Cloud Synthesis, and Biophysical Crop Stress Diagnostics
Year: 2026
Type: preprint
Abstract: Vyom-205M is a 204.96-million parameter sovereign multi-task Earth observation foundation model engineered specifically for complex smallholder agrarian landscapes across the Global South. Smallholder agricultural systems are characterized by fragmented micro-plots (<0.4 ha), ultra-narrow earthen embankments (bunds <1.5 m), persistent monsoonal cloud occlusion (>75% optical loss during Kharif phenology), and confounded crop-weed-stress spectral mixtures. Vyom-205M integrates a multi-stream hierarchical architecture that harmonizes sub-meter cadastral optical imagery, 12-band multi-temporal Sentinel-2 optical time series, multi-temporal Sentinel-1 dual-polarization synthetic aperture radar (SAR), and gridded agro-climatic soil covariates via a Pre-LayerNorm Swin-Transformer core equipped with Query-Key Normalization (QK-Norm). Key Benchmark Results: Global Flood Inundation (Sen1Floods11): Achieves 75.81% Water IoU, 85.83% Mean IoU, and 96.33% Pixel Accuracy across 30 held-out test scenes in 10 countries, outperforming IBM/NASA Prithvi-100M (68.40% Water IoU) and CROMA-Large (71.40% Water IoU) while using 31.7% fewer parameters than 300M models. Smallholder Cadastral Boundary Delineation: On 349 held-out validation scenes, Head A achieves 0.8230 Parcel Extent IoU (0.9029 F1) and 0.2783 2-pixel Relaxed Bund Boundary F1. All-Weather SAR-to-Optical Cloud Synthesis: Under the SEN12MS-CR protocol across 116 spatio-temporal cubes, Head B reconstructs cloud-free Sentinel-2 surface reflectance from C-band SAR with 0.8924 SSIM and 0.0298 MAE. Biophysical Crop Stress Diagnostics: On 550 held-out parcels, Head D achieves 58.55% multi-class accuracy and 0.5730 Macro-F1 across 6 diagnostic categories (81.25% recall on Flood Inundation, 83.33% recall on Soil Salinity, and 77.87% recall on early Moisture Deficit). Public Verification Suite & Code: All aggregated confusion matrices, evaluation summaries, and a zero-dependency verification harness (<0.1s execution) are available at https://github.com/sphitih-ai/sphitih-ai-vyom-205m-benchmarks-.
Author: Srivastava, Vaidic
Language: English
Subfield: Ecology
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225638440
Cites:
Cited by:
Related to: