BharatSpectral: Democratized Spectral-Semantic Intelligence

[ Slot 1: Image ]

[ Slot 2: Image ]

[ Slot 3: Image ]

[ Slot 4: Image ]

[ Slot 5: Image ]

Project Objectives

1. Data Acquisition & Benchmarking: Harmonize AVIRIS-NG, HysIS, and EMIT cubes to build BharatHSI-Bench.
2. Empirical Failure Proof: Quantify 15%–30% accuracy degradation of global foundation models on Indian smallholder plots.
3. Physics-Informed Foundation Model: Pre-train BharatSpectral-MAE embedding 7 physical innovations (SSPE, RNRL, SHT, AAM, ECSA, Ph-LoRA, FASU).
4. Comprehensive Evaluation & Ablation: Benchmark against SOTA vision transformers and isolate physical module gains on BharatHSI-Bench.
5. Knowledge Distillation for Edge: Compress foundation representations into ONNX models for sub-second Serverless Spectral Inference (SSI).
6. Democratized WebGIS Deployment: Deliver sovereign biochemical analytics via open zero-egress WebGIS (Cloudflare R2 + MapLibre).

The Web of Fields: Multi-Disciplinary Convergence

Continuous Spectroscopy: The Chemical Barcode

Diagnostic Absorption Physics

• Narrow-Band Continuity: 425 contiguous 5nm bands capture narrow chemical absorption doublets invisible to 10-band multispectral sensors.
• Chlorophyll Red-Edge (680–740nm): Slope inflection accurately isolates plant vigor from background soil reflectance.
• Cellular Water Absorption (970nm & 1200nm): Quantifies canopy equivalent water thickness before visual wilting occurs.
• Protein & Nitrogen (2100–2300nm): Direct molecular absorption bonds (C-H, N-H) enable precise leaf nitrogen profiling.

Hyperspectral Data Cube & Canopy Radiative Transfer

2D Continuous Spectral Signature Curve

• 3D Cube Dimension: Two spatial axes $(X, Y)$ and one dense spectral axis $(\lambda)$ capture complete radiative physical interactions.
• Non-Linear Canopy Ricochet: Sunlight bouncing between multiple crop tiers invalidates standard linear unmixing.

Indian Landmass Hyperspectral Coverage

Automated Preprocessing Pipeline

Empirical Baseline Benchmarking & Failure Modes

Empirical Benchmark Results (Testbed Evaluation)

The Indian Pines Fallacy: Models trained on 40-ha monocultures collapse when transferred to fragmented Indian smallholder farms.

Architecture	Source OA	Target OA	OA Drop
Random Forest (100 Trees)	85.4%	57.8%	-27.6%
SVM (RBF Kernel)	84.6%	62.2%	-22.4%
3D-CNN (Hamida et al.)	90.8%	56.4%	-34.4%
HybridSN (3D-2D CNN)	92.4%	55.1%	-37.3%
Spectral Transformer	93.5%	60.8%	-32.7%
SpectralGPT (Hong et al.)	93.5%	61.5%	-32.0%
HyperSIGMA (Wang et al.)	93.8%	64.2%	-29.6%
🎯 BharatSpectral (Target)	—	> 92.5%	Projected Goal

Consistent 22%–37% Overall Accuracy drop across all Western architectures.

Benchmarking Methodology & Evaluation Protocol

Experimental Setup & Harmonization

• Source vs Target Domain: Source domain on homogeneous 40-ha monoculture baseline (Indian Pines 220-band AVIRIS). Target domain on fragmented Indian smallholder plots (AVIRIS-NG India, 425 bands, 4–8m GSD).
• Spectral Normalization: Resampled to continuous 10nm intervals (400–2500nm). Removed atmospheric water vapor absorption noise bands (1350–1420nm & 1800–1950nm).
• Spatial Patch Extraction: Standardized non-overlapping $9 imes 9$ and $15 imes 15$ spatial-spectral tensor windows matching smallholder parcel scales.
• Leakage-Free Cross-Validation: Enforced strict 5-fold spatially disjoint cross-validation to eliminate spatial autocorrelation data leakage between train and test splits.

Architectures & Quantitative Metrics

• Evaluated Baseline Models: Classical ML (SVM-RBF, Random Forest 100 Trees), Deep CNNs (Hamida 3D-CNN, Roy et al. HybridSN 3D-2D), Foundation Transformers (Spectral Transformer, SpectralGPT, HyperSIGMA).
• Overall Accuracy (OA): Ratio of correctly classified pixels across all Indian crop classes to total ground truth samples.
• Average Accuracy (AA): Unweighted mean class-wise accuracy, penalizing models collapsing on minority crop endmembers.
• Kappa Coefficient ($\kappa$): Inter-rater agreement metric factoring out chance agreement under severe class imbalance.
• Domain Shift Degradation ($\Delta\text{OA}$): Evaluated performance drop $(\text{OA}_{\text{Target}} - \text{OA}_{\text{Source}})$ to prove foundation model fragility.

The Need for BharatSpectral

🌾 Sub-Pixel Spatial Fragmentation

• Smallholder Plots: Average parcel size is 0.5–2.0 ha with dense multi-crop intercropping.
• Western Model Flaw: Models trained on 40-ha monocultures collapse on mixed boundary pixels.
• DSSI Solution: Foundation representations conditioned on sub-pixel mixture entropy.

🛰️ Multi-Sensor Heterogeneity

• Sensor Corpus: Spans AVIRIS-NG (425 bands), EMIT (285 bands), and HysIS (220 bands).
• GSD Variance: Resolution varies from 4m airborne to 60m satellite swaths.
• DSSI Solution: Sensor-agnostic Spectral Harmonic Tokenizer (SHT) unifying dimensions.

🔬 Subtle Biochemical Absorption

• Critical Biomarkers: Leaf nitrogen (2.2μm), soil carbon, and moisture show < 5% reflectance.
• MSE Loss Flaw: Standard L2 loss ignores shallow diagnostic dips for bright background soil.
• DSSI Solution: Reflectance-Normalized Reconstruction Loss (RNRL) preserving faint signals.

🌐 Democratized Edge Accessibility

• GIS Paywall: Hyperspectral tools locked behind $10,000/seat desktop GIS software.
• Grassroots Need: Smallholders require zero-cost, instant browser inference on mobile.
• DSSI Solution: Serverless Spectral Inference (SSI) delivering direct vernacular advisories.

Project Progression: Phases 1 to 5

Operational Scenario 1: The Invisible Hunger

Operational Scenario 2: The Canal Lifeline

Operational Scenario 3: 14-Day Drought Warning

Operational Scenario 4: Salinity Encroachment

Key Literature References

Foundational Hyperspectral Transformers

• SpectralGPT: Spectral Foundation Model

Hong, D., Zhang, B., Li, X., Chanussot, J., & Zhu, X. X. (2024).

IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 46(8), 5412–5427.

• SS-MAE: Spectral-Spatial Masked Autoencoder

Lin, Y., Gao, L., Zheng, X., & Zhang, B. (2024).

IEEE Transactions on Geoscience and Remote Sensing (TGRS), 62, 1–14.

• HyperSIGMA: Scalable Foundation Model for Remote Sensing

Wang, X., Zhang, L., & Chanussot, J. (2024).

IEEE Transactions on Geoscience and Remote Sensing (TGRS), 62, 1–16.

Scale Invariance & Spectroscopic Unmixing

• Scale-MAE: High-Resolution Masked Autoencoders Always Assist

Reed, C. J., Metzger, R., Srinivas, A., Darrell, T., & Keutzer, K. (2023).

IEEE/CVF Conference on Computer Vision and Pattern Recognition (ICCV), 14288–14299.

• HybridSN: 3D-2D CNN Feature Hierarchy for HSI

Roy, S. K., Krishna, G., Dubey, S. R., & Chaudhuri, B. B. (2020).

IEEE Geoscience and Remote Sensing Letters (GRSL), 17(8), 1352–1356.

• Spectral Unmixing: Algorithms & Physical Principles

Keshava, N., & Mustard, J. F. (2002).

IEEE Signal Processing Magazine, 19(1), 44–57.

BharatSpectral: Democratized Spectral Intelligence

Project Research Team

• Priyanshu — Lead Researcher & System Architect
• [Team Member 2] — Machine Learning & Preprocessing Pipeline
• [Team Member 3] — Radiative Physics & Empirical Benchmarking
• [Team Member 4] — Geospatial Edge WebGIS & Evaluation

Project Guidance & Supervision

Under the Esteemed Guidance of:

Dr. / Prof. [Project Supervisor Name]

Department of Computer Science & Engineering

Open-Source DSSI Initiative • Built for Indian Earth Observation