• 400–700 nm (VNIR): Plant pigment absorption (Chlorophyll a/b, Carotenoids).
• 700–750 nm (Red-Edge): Steep cellular structure cliff. NOTE: SHT (Spectral Harmonic Tokenizer allocates fine tokens here).
• 970 & 1200 nm (NIR): Cellular liquid Equivalent Water Thickness (EWT).
• 1.4 & 1.9 μm: Atmospheric water vapor zero-transmission gaps. NOTE: AAM (Atmospheric Absorption Masking).
• 2.1–2.3 μm (SWIR): Organic protein nitrogen, soil organic carbon (SOC), clay mineral lattices.

• Multi-Bounce Scattering: Sunlight enters intercropped sorghum + pigeon pea and ricochets between leaves and soil before sensor reception.
• Failure of Linear Unmixing: Classical GIS (LSU/FCLS) assumes linear superposition ($r = \sum a_i e_i$). Ricochets cause non-linear cross-talk ($RMSE_A > 0.15$).
• NOTE: SSPE Scale-Spectral Positional Encoding: Jointly encodes GSD ($4–60\, ext{m}$), bandwidth, and mixture entropy.
• NOTE: FASU Foundation-Augmented Spectral Unmixing: Decomposes complex non-linear canopy mixtures using pretrained foundation representations.

• Spectroscopy Physics: Provides ground truth absorption laws, radiative transfer equations, and atmospheric absorption gap constraints.
• Foundation Model AI: Provides self-attention capacity, masked autoencoding pretraining, and non-linear parameter-efficient adaptation (LoRA).
• WebGIS Public Infrastructure: Eliminates cloud costs through Cloudflare R2 zero-egress storage and browser-edge ONNX WASM inference.
• Core Convergence: BharatSpectral (DSSI) delivers democratized biochemical intelligence directly to citizen devices.

• AVIRIS-NG India: 425 continuous bands (380–2510 nm), 4–8m GSD flightlines over Anand (Gujarat), Godavari Basin (AP), and Punjab tracts.
• ISRO HysIS: 220 bands (VNIR/SWIR), 30m spaceborne GSD.
• NASA EMIT: 285 bands (381–2493 nm), 60m GSD on the International Space Station.
• NOTE: RNRL Reflectance-Normalized Reconstruction Loss: Prevents gradients from collapsing in low-reflectance SWIR bands (<5% reflectance).

Removes atmospheric path radiance ($L_{path}$) via 6S radiative transfer code, then converts raw 16-bit integer Digital Numbers into $[0.0, 1.0]$ Top-of-Canopy surface reflectance tensors.

Prunes zero-transmission water absorption windows (1350–1450 nm and 1800–1950 nm) retaining 200 standardized bands, then generates $9 imes 9 imes B$ and $15 imes 15 imes B$ smallholder patches.

The Indian Pines Fallacy: Models trained on 1992 Indiana monocultures fail in India due to 1.08 ha plot fragmentation, intercropping, and 3-season phenology drift.

• GIS SAM Baseline: Cosine angle ignores absolute reflectance magnitude, causing a 36.4% false match rate (confuses shadow/moisture with crop stress).
• GIS Linear Spectral Unmixing (LSU): Produces high residual error ($RMSE_A = 0.2775$ on boundaries; 54.2% pixels fail threshold) due to non-linear canopy scattering.
• NOTE: ECSA Endmember-Constrained Self-Attention: Regularizes Transformer attention weights using spectroscopic unmixing priors.
• NOTE: Ph-LoRA Phenology-Conditioned LoRA: Modulates adapter weights by Kharif, Rabi, and Zaid phenological stage embeddings.

Data ingestion pipelines verified across 3 sensors; baseline architectures evaluated on Raspberry Pi 5; 22%–37% domain shift collapse empirically proven; GIS spectroscopic failure modes audited.

BharatSpectral-MAE pretraining; LoRA adapter tuning; multi-sensor harmonization; followed by Phase 4 & 5 serverless edge deployment on Cloudflare R2 + Workers WebGIS.

• Zero App Install: 100% web browser execution in mobile Chrome/Safari.
• Zero Egress Cost: Cloudflare R2 stores COG tiles with $0 egress fees.
• Edge Inference: Quantized ONNX WASM model executes sub-pixel unmixing directly inside the farmer's browser in <15 milliseconds.
• Plain Language Advisories: Translates spectral nitrogen deficit into direct KVK advice: "Apply 12 kg Urea in Northern plot; skip Southern plot".