B.Tech Capstone Project Review • Mid-Term Defense

BharatSpectral

Democratized Spectral Intelligence for Indian Earth Observation Through Physics-Informed Foundation Modeling and Public Geospatial Infrastructure

🔬 Research Pillar
BharatSpectral-MAE: First hyperspectral Foundation Model custom-engineered for Indian smallholder fragmentation, intercropping, and 3-season phenology via 7 novel architectural innovations.
🌐 Product Pillar
BharatSpectral Platform: Zero-cost, public WebGIS delivering biochemical analytics (Nitrogen, Soil Organic Carbon, Water Quality) using serverless edge inference and zero-egress tile streaming.
🇮🇳 National Alignment
Digital Public Infrastructure: Harnesses IndiaAI AIRAWAT DGX A100 HPC, ISRO AVIRIS-NG & HysIS missions, and open data rails to empower farmers and district collectors without paywalls.
Department of Computer Science & Engineering HPC: IndiaAI AIRAWAT • Sensors: ISRO AVIRIS-NG / HysIS • NASA EMIT
Context & Motivation

The Indian Geospatial Paradox: Data Abundance vs. Field Paralysis

Why 140 million smallholder farmers remain locked in guesswork despite India's top-5 space program

🛰️ Observational Abundance (Spaceborne Potential)
  • Top-5 Global Space Power: ISRO operates elite earth observation constellations and airborne campaigns (AVIRIS-NG India, HysIS) capturing petabytes of spectral data.
  • Massive Public Investment: Hundreds of crores invested in remote sensing hardware, yet operational application remains confined to academic papers and closed research labs.
  • The "Data Graveyard" Syndrome: AVIRIS-NG datasets reside as raw 5–10 GB binary ENVI/HDF files on ISRO Bhoonidhi. Completely uninterpretable by agronomists or district officials.
🌾 Field Paralysis (The Smallholder Reality)
  • 1.08 Hectare Reality: 86% of Indian landholdings are smallholders. A single misdiagnosed crop infection or fertilizer delay can cause financial devastation.
  • The Enterprise Paywall: Commercial hyperspectral analytics (e.g. Pixxel Aurora) target enterprise agribusiness at thousands of dollars/month — inaccessible to rural Panchayats.
  • The Software & Compute Barrier: Hyperspectral analysis currently demands heavy desktop software (ENVI, ERDAS) or complex Python GIS coding in Google Earth Engine.
  • Our Mission: Translate high-dimensional spectroscopy into an open, zero-cost, instant mobile WebGIS platform for every Indian citizen.
Spectroscopy Principles

Biophysical vs. Biochemical Intelligence

Why 10-band multispectral satellites detect failure too late, and how 200+ narrow bands diagnose root causes

MULTISPECTRAL SENSING (Current Status Quo)
Sentinel-2, Landsat-8/9, ISRO LISS-IV (10–12 Broad Spectral Bands)
  • Intelligence Level: Biophysical only (NDVI, NDRE, EVI).
  • What it Measures: Detects THAT a crop is losing vigor or biomass.
  • The Fatal Blindspot: Cannot determine WHY. Nitrogen starvation, fungal leaf rust, soil salinity, and moisture stress look spectrally IDENTICAL across broad 100 nm bands.
  • Irreversible Yield Loss: By the time broad-band NDVI drops, chlorophyll breakdown is extensive and 15–25% yield loss is already locked in.
HYPERSPECTRAL INTELLIGENCE (BharatSpectral)
AVIRIS-NG, NASA EMIT, ISRO HysIS (200–425 Continuous Narrow Bands, 5nm)
  • Intelligence Level: Biochemical & Molecular spectroscopy (380–2500 nm).
  • Pre-Symptomatic Nitrogen (720 nm): Isolates the exact Red-Edge inflection shift driven by cellular nitrogen binding 7–14 days before visible yellowing.
  • Yellow Rust Spores (680 nm vs 710 nm): Separates fungal mycelium damage from water stress, halting whole-field pesticide overuse.
  • Soil Organic Carbon (2200 nm): Measures the SWIR Al-OH and clay-organic absorption doublet directly from orbit without physical soil core sampling.
Literature Gap & Verification

Why Existing Foundation Models Fail on Indian Landscapes

SOTA models (SpectralGPT, HyperSIGMA, SS-MAE) suffer catastrophic domain shift when applied to Indian agriculture

⚠️ The Canonical "Indian Pines" Fallacy: Foundation models claim success on "Indian Pines" (1992). Despite its name, this dataset was captured in Indiana, USA over giant rectilinear monocultures. It shares ZERO agronomic, ecological, or spatial characteristics with Indian agriculture!
1. Extreme Fragmentation
Average plot is 1.08 ha (millions <0.5 ha). At 30–60m pixel resolution (EMIT/HysIS), every pixel contains multiple crops and boundaries. Western models assume pure pixels; Indian data requires sub-pixel unmixing at every point.
2. Intercropping Mixing
Indian farmers co-plant 2–3 species simultaneously (e.g. Sorghum + Pigeon Pea). The resulting canopy spectra are non-linear mixtures completely absent from Western monoculture benchmarks.
3. 3-Season Phenology
India experiences Kharif (monsoon), Rabi (winter), and Zaid (summer). The same GPS coordinate shows completely divergent phenology. Single-season models suffer catastrophic seasonal drift.
4. Multi-Sensor Gap
No existing model harmonizes airborne AVIRIS-NG (4–8m GSD, 425b) with spaceborne EMIT (60m GSD, 285b) and HysIS (30m GSD, 220b). Scale-MAE ignores the spectral bandwidth dimension completely.
Ecosystem Audit & Verification

National Remote Sensing Landscape: The Public Infrastructure Gap

Exhaustive 2026 audit of Indian geospatial platforms verifying that no citizen hyperspectral engine exists

Platform / Entity Data Modality Access Model Web HSI Inference? The Translational Gap
ISRO Bhuvan / Krishi-DSS Multispectral (LISS, Sentinel-2) Public Web Portal None (No HSI) Biophysical stress only; cannot diagnose biochemical root cause.
ISRO VEDAS (AVHYAS) Hyperspectral (AVIRIS-NG) Desktop QGIS Plugin Offline Only Requires heavy local workstation install; zero browser/citizen access.
ISRO Bhoonidhi Raw HSI Data Catalog Public (Registration) None (Raw ENVI) Distributes raw 5–10 GB binary cubes; no automated analytics engine.
Pixxel Aurora Hyperspectral (Constellation) Commercial B2B SaaS Yes (Proprietary) Enterprise paywall; inaccessible to smallholders and public research.
Google Earth Engine (GEE) PaaS (Hosts NASA EMIT) Freemium PaaS User Must Code Requires Python/JS GIS scripting; no pre-trained smallholder AI models.
BharatSpectral (Ours) Multi-Sensor HSI (200–425 b) 100% Free Public Infra Yes (Real-Time Edge) The ONLY open Foundation Model + zero-cost WebGIS platform in India.
🎯 Verified Finding: No system exists globally or nationally that combines a physics-informed Foundation Model engineered for Indian smallholders with zero-egress, citizen-accessible public WebGIS deployment.
System Architecture

The Dual-Pillar Framework: Synergizing AI Research & Public Infrastructure

Creating Democratized Spectral-Semantic Intelligence (DSSI) from laboratory spectroscopy to citizen fingertips

🔬 Pillar 1: AI Foundation Research
BharatSpectral-MAE Engine (AIRAWAT HPC DGX A100)
  • Multi-Sensor Pre-Training: Ingests airborne AVIRIS-NG (425 bands, 4–8m), spaceborne EMIT (285 bands, 60m), and HysIS (220 bands, 30m).
  • 7 Named Innovations: SSPE, RNRL, SHT, AAM, ECSA, Ph-LoRA, and FASU explicitly solve Indian sub-pixel and phenological challenges.
  • BharatHSI-Bench: First open, standardized Indian hyperspectral agricultural benchmark with ground truth labels.
  • Teacher-Student Distillation: Compresses heavy Vision Transformers into lightweight edge models.
🌐 Pillar 2: Public Geospatial Infrastructure
BharatSpectral Platform (Cloudflare Serverless)
  • Zero-Egress Streaming: Multi-terabyte Zarr datacubes and Cloud-Optimized GeoTIFFs (COG) hosted on Cloudflare R2 with $0 egress fees.
  • Serverless Spectral Inference (SSI): Quantized ONNX student model executes inside Cloudflare Workers V8 isolates within 128 MB RAM.
  • Citizen MapLibre WebGIS: Fast, responsive web frontend allowing farmers, KVK officers, and researchers to query any coordinate without GIS tools.
  • Actionable Intelligence: Outputs single-click diagnostic maps: Nitrogen (kg/ha), Soil Carbon (%), Algal Bloom Toxicity alerts.
Research Core

BharatSpectral-MAE: The 7 Named Architectural Innovations

Physics-informed mechanisms engineered specifically for Indian smallholder Earth Observation

[SSPE] Scale-Spectral Positional Encoding
Jointly encodes GSD (4m–60m), spectral bandwidth, and mixture entropy across 3 sensors. Extends Scale-MAE into the hyperspectral regime.
[RNRL] Reflectance-Normalized Reconstruction Loss
Normalizes self-supervised reconstruction by target magnitude, preventing low-reflectance water bodies (<5% SWIR) from being discarded as noise.
[SHT] Spectral Harmonic Tokenizer
Sensor-adaptive 3D tokenizer allocating fine-grained tokens in diagnostic Red-Edge zones (700–750 nm) and coarse tokens in continuum regions.
[AAM] Atmospheric Absorption Masking
Simulates atmospheric water vapor (1350–1420 & 1800–1950 nm) and CO2 absorption gaps, forcing the Transformer to learn radiative transfer physics.
[ECSA] Endmember-Constrained Self-Attention
Regularizes self-attention via spectral unmixing similarity, attending to physically matching crop signatures across fragmented plot boundaries.
[Ph-LoRA] Phenology-Conditioned LoRA
Dynamically modulates PEFT adapter weights by season (Kharif/Rabi/Zaid) and growth stage embeddings, preventing seasonal spectral drift.
[FASU] Foundation-Augmented Spectral Unmixing: Direct sub-pixel unmixing head operating on pre-trained representations without retraining standalone autoencoders. No Published Precedent
Software Architecture

Product Pillar: Breaking the 3 Structural Cloud Bottlenecks

How our serverless architecture reduces marginal operating costs to near-zero for sustained public access

1. Data Volume Bottleneck
THE PROBLEM:

A single scene is 2–10 GB. Serving 200+ raw bands to thousands of concurrent citizen users crashes standard WebGIS tile servers.

OUR SOLUTION:

Convert scenes into Cloud-Optimized GeoTIFFs (COG) and chunked Zarr datacubes. The browser requests ONLY the exact bounding box and 3–5 diagnostic wavelengths via HTTP Range requests, slashing payload sizes by 98%.

2. Egress Cost Bottleneck
THE PROBLEM:

AWS S3 and GCP charge $0.08–$0.12/GB for outbound data egress. Streaming gigabyte-scale spectral cubes to the public creates thousands in recurring cloud debt.

OUR SOLUTION:

Cloudflare R2 object storage with guaranteed $0 data egress fees. Public users can pan, stream, and query spectral cubes infinitely without incurring bandwidth penalties.

3. Inference Compute Bottleneck
THE PROBLEM:

Hyperspectral models (200+ bands) demand high-end GPU clusters ($1,500+/mo), making public citizen deployment economically unsustainable.

OUR SOLUTION:

Serverless Spectral Inference (SSI): Distilled ONNX student model running directly in Cloudflare Workers V8 isolates within strict 128 MB RAM limits, providing sub-second inference at the edge.

Engineering Pipeline

The End-to-End Data Journey: From Raw Photons to Citizen Action

Detailed workflow connecting ISRO/NASA satellites to browser-based edge inference

Stage 1: Preprocessing
  • ISRO AVIRIS-NG & NASA EMIT ingestion
  • Bad Band Removal (BBR) dropping 1350–1420 & 1800–1950 nm
  • Radiative transfer surface reflectance conversion
  • Analysis-ready Zarr & COG creation on Cloudflare R2
Stage 2: AIRAWAT HPC
  • IndiaAI AIRAWAT DGX A100 nodes
  • PyTorch DDP distributed scaling
  • Self-supervised pre-training with 7 innovations (SSPE, RNRL, SHT, AAM, ECSA)
  • bfloat16 mixed precision & gradient checkpointing
  • Ph-LoRA fine-tuning for Kharif/Rabi
Stage 3: Distillation
  • Teacher: Spectral-MAE Transformer
  • Student: Mobile 3D-2D CNN
  • Transfers "dark knowledge" and absorption sensitivity
  • INT8 post-training quantization
  • ONNX runtime compilation optimized for 128 MB V8 isolates
Stage 4: Edge Delivery
  • Cloudflare R2 zero-egress tile hosting
  • Serverless Spectral Inference (SSI) on Workers
  • Next.js + MapLibre GL JS client portal
  • Sub-second diagnostic maps: Nitrogen, Carbon, Cyanobacteria
  • Operates on 4G/5G mobile browsers
Real-World Impact: Use Case 1

Precision Agriculture: The Smallholder Farmer in Ludhiana

Pre-symptomatic nitrogen deficiency and yellow rust diagnosis 7 to 14 days before visible damage

🌾 Persona: Gurpreet Singh (Ludhiana, Punjab)
  • The Challenge: Gurpreet cultivates 2.4 acres of wheat. Over-application of urea has degraded his soil and spiked input costs, while stripe rust (yellow rust) threatens his crop every February.
  • The Status Quo Dilemma: Multispectral apps (NDVI) only detect distress after leaves turn yellow. By then, fungal mycelium has penetrated the tissue and 20% yield loss is already locked in.
  • BharatSpectral Intervention: Gurpreet opens BharatSpectral on his phone. The system queries recent EMIT/AVIRIS-NG passes and unmixes sub-pixel signatures over his plot coordinates.
  • Pre-Symptomatic Diagnosis: Pinpoints the Red-Edge inflection shift at 720 nm (nitrogen deficiency) vs 680 nm absorption drop (yellow rust spores) 10 days before visual symptoms appear.
📈 Quantifiable Economic & Agronomic Outcomes
  • ₹3,500 / Acre Fertilizer Savings: Variable-rate nitrogen prescription maps target exact deficiency pockets, cutting urea broadcast by 30%.
  • Yield Protection (15–22% Preserved): Early targeted fungicide spraying in micro-clusters halts yellow rust epidemics before whole-field infestation occurs.
  • Groundwater Protection: Prevents toxic nitrate leaching into Punjab's over-exploited Malwa aquifer.
  • Zero Jargon: Color-coded prescription map on mobile WhatsApp/PWA ("Zone A: Apply 8kg Urea; Zone B: Healthy").
Real-World Impact: Use Case 2

Soil Health & Governance: District Collector & KVK in Karnal

Automating Soil Organic Carbon (SOC) and Soil Health Card verification across entire districts

🏛️ Persona: Dr. Anita Verma (KVK Officer, Karnal)
  • The Bottleneck: Mandated to issue 25,000 Soil Health Cards annually. Physical soil sampling requires laboratory wet-chemistry (Walkley-Black titration) taking 4–6 weeks per sample. Over 80% of plots remain unverified.
  • Residue Burning Crisis: Stubble burning in October degrades topsoil organic matter, but blanket fertilizer subsidies mask progressive soil degradation.
  • The BharatSpectral Solution: Uses Ph-LoRA bare soil attention to map Soil Organic Carbon (SOC) and clay mineralogy directly from 2200 nm SWIR absorption doublets across the entire district at 10m resolution.
  • Automated Soil Cards: Pairs hyperspectral reflectance directly with National Soil Health Card databases, generating continuous spatial soil health layers without physical transport delays.
📊 Policy & Agricultural Impact
  • 100% District Coverage vs 5% Manual: Replaces sparse point-sample interpolations with exhaustive, continuous wall-to-wall soil carbon mapping.
  • Stubble Burning Impact Tracking: Quantifies topsoil carbon volatilization post-fire, providing district magistrates with empirical data to reward zero-burn farmers.
  • Rationalized Subsidies: State agriculture departments can redirect subsidized fertilizer based on true biochemical deficiencies rather than political quotas.
  • Carbon Market Verification: Provides the spatial baseline required for Indian smallholders to participate in international soil carbon credit programs.
Real-World Impact: Use Case 3

Environmental Protection: Water Quality Inspector in Varanasi

Detecting toxic cyanobacterial blooms and industrial effluent plumes in low-reflectance inland waters

🌊 Persona: Rajesh Tripathi (Pollution Control, Varanasi)
  • The Inland Water Challenge: Water absorbs >95% of incoming solar radiation in SWIR (reflectance <5%). Standard AI models and multispectral satellites treat water as dark noise.
  • Toxic Cyanobacteria vs. Algae: Multispectral sensors measure broad chlorophyll-a (665 nm), mistaking harmless green algae for life-threatening cyanobacterial blooms releasing microcystin liver toxins.
  • BharatSpectral Breakthrough: Reflectance-Normalized Reconstruction Loss (RNRL) forces the model to learn subtle spectral variations in low-reflectance water bodies.
  • Phycocyanin Detection: Isolates the specific 620 nm phycocyanin absorption dip, detecting toxic blooms 5 days before fish kills and water treatment shutdowns occur.
🛡️ Public Health & Ecological Outcomes
  • Water Intake Protection: Early warning alerts dispatch to Varanasi municipal water treatment plants to switch coagulants and activate carbon filters.
  • Industrial Effluent Tracing: Narrow-band spectral unmixing traces chromium and chemical dye plume dispersion from Kanpur/Unnao tannery drains.
  • Namami Gange Support: Provides National Mission for Clean Ganga with verifiable, transparent water quality layers (Turbidity, CDOM, Chl-a) without manual boat sampling.
  • Zero Lab Lag: Reduces environmental compliance reporting from 14 days of wet-lab incubation to instant sub-minute web inspection.
Real-World Impact: Use Case 4

Disaster Resilience & Crop Insurance: PMFBY Claim Verification

Objective sub-pixel quantification of crop lodging, drought desiccation, and flood inundation

⚖️ The Insurance Bottleneck (PMFBY)
  • Crop Cutting Experiment (CCE) Delays: Pradhan Mantri Fasal Bima Yojana relies on manual CCEs. Conducting millions of physical field cuts takes 3–6 months, leading to prolonged payout disputes.
  • Subjective Litigation: Disagreements between insurance companies and state governments over drought or hailstorm severity frequently freeze compensation funds.
  • Flash Drought Canopy Water Loss: Multispectral sensors detect drought only after plant canopies turn brown. Hyperspectral 970 nm and 1200 nm liquid water absorption bands measure cell turgor pressure drop in real time.
  • Sub-Pixel Crop Lodging Detection: High cyclonic winds flatten crops. BharatSpectral unmixes soil-canopy structural geometry shifts, separating flattened crops from standing fields.
⚡ Automated Claim Settlement Impact
  • Payouts in Days, Not Months: Instant satellite-derived loss assessment enables direct benefit transfers (DBT) to farmers' bank accounts within 72 hours of catastrophic weather.
  • Sub-Hectare Plot Granularity: FASU sub-pixel unmixing accurately resolves damage on plots as small as 0.2 hectares, ensuring smallholders are not excluded by coarse pixel averaging.
  • 100% Tamper-Proof Audit Trail: Publicly verifiable, open-access hyperspectral records eliminate fraudulent claims and political tampering.
  • Disaster Relief Coordination: State disaster management authorities can immediately prioritize relief supplies to the exact tehsils with critical crop biomass destruction.
Master Execution Roadmap

End-to-End Project Timeline: Engineering Phases 1 Through 5

Systematic progression from raw data ingestion to national-scale public infrastructure deployment

Phase 1: Ingestion
DATA SCORING & PIPELINE
  • AVIRIS-NG, EMIT & HysIS data ingestion
  • Bad Band Removal (BBR) & Radiative Transfer
  • Chunked Zarr & COG creation on Cloudflare R2
  • BharatHSI-Bench benchmark ground truth curation
  • SOTA failure documentation
Phase 2: Baseline
BENCHMARKING & HPC
  • 1D-CNN, 3D-CNN & HybridSN baselines
  • Classical unmixing (VCA & FCLSU endmembers)
  • AIRAWAT Slurm batch orchestration scripts
  • Quantitative evaluation metric baseline
  • Strict memory guardrails
Phase 3: Foundation
CREATIVE CORE & DISTILLATION
  • BharatSpectral-MAE with 7 innovations
  • Self-supervised pre-training on DGX A100 nodes
  • Ph-LoRA fine-tuning for Kharif/Rabi phenology
  • Teacher-Student Knowledge Distillation
  • INT8 quantization
Phase 4: Platform
WEBGIS ENGINEERING
  • Next.js + MapLibre GL JS frontend on Cloudflare
  • Cloudflare R2 zero-egress tile streaming API
  • Serverless Spectral Inference (SSI) on Workers
  • Sub-pixel agricultural map visualization
  • Sub-128 MB RAM execution
Phase 5: Release
VALIDATION & OPEN SCIENCE
  • End-to-end latency & accuracy validation
  • Cloud cost verification ($0 egress vs AWS)
  • Open-source weights on AIKosh & HuggingFace
  • Comprehensive capstone thesis & documentation
  • Final viva defense
Evaluation Framework

Quantitative Benchmarking & Success Metrics

Rigorous empirical standards across AI accuracy, sub-pixel unmixing, and platform performance

🎯 AI Accuracy & Unmixing Targets
  • Classification Accuracy (BharatHSI-Bench):
    • Overall Accuracy (OA): Target > 92.5%
    • Average Accuracy (AA): Target > 89.0%
    • Cohen's Kappa Coefficient (κ): Target > 0.90
  • Cross-Scene Generalization: Must demonstrate +15% to +25% OA gain over Western-pretrained SpectralGPT and HyperSIGMA.
  • Sub-Pixel Unmixing (FASU):
    • Abundance Root Mean Square Error (RMSE): < 0.08
    • Spectral Angle Distance (SAD): < 0.05 rad
⚡ System Performance & Economic Feasibility
  • Serverless Latency (SSI on Workers):
    • Inference Time: < 5.0 seconds per km² tile
    • Browser Tile Render (MapLibre): < 200 ms
    • Edge Memory: Strictly under 128 MB V8 isolate ceiling
  • Zero Egress Cost Validation:
    • Outbound Data Transfer Fee: Exactly $0.00 / month on Cloudflare R2
    • Total Hosting Cost: < $50 / month vs $1,200+ / month for equivalent AWS EC2 + GeoServer deployment
  • Citizen Accessibility: 100% responsive on standard Android 4G/5G mobile browsers.
Conclusion & Vision

BharatSpectral: Democratizing Spectral Intelligence as Digital Public Infrastructure

From closed scientific repositories to nationwide citizen empowerment

7 Architectural Novelties
Engineered the first foundation model built specifically for Indian smallholders. SSPE, RNRL, SHT, AAM, ECSA, Ph-LoRA, and FASU close the empirical gap that causes Western models to fail.
3 National Benchmarks
Establishes BharatHSI-Bench, Multi-Sensor Indian Datacubes, and Paired Spectral-SoilHealth records on AIKosh, freeing Indian academia from its 30-year reliance on outdated foreign datasets.
Public Digital Good
Translates complex aerospace spectroscopy into a zero-cost, high-speed WebGIS tool accessible to any farmer, extension worker, or policymaker on any device without paywalls.
🇮🇳 Alignment with National Missions: Directly advancing the IndiaAI Mission, Digital Public Infrastructure (DPI), and National Mission on Sustainable Agriculture — proving that world-class AI research can deliver direct social utility to the common citizen.

🎙️ Presenter Talking Points ✖