- In this physics-grounded synthetic demo, every crop-stress model faces the same objection: you are just classifying your own data. We built a 424-band diagnosis engine and checked its weighted wavelengths against the physics 🧵
- Multispectral NDVI flags a corn block as stressed but not why. One amber blob, no cause. Knowing something is wrong is a long way from knowing what to spray.
- SpectraRx reads the full 424-band reflectance curve (400-2515 nm, VNIR+SWIR) at every pixel. A 1-D spectral CNN types each curve as healthy, nitrogen, water, or early tar spot. That part is routine. The real question is whether it learned physics or just noise.
- So we compute band saliency: which wavelengths drove each verdict, measured independently of the features the model trained on. If it learned the real signal, the salient bands should fall on the diagnostics a remote-sensing scientist would name. They do.
- Tar spot saliency peaks in sampled 530 and 535 nm bins nearest the PRI diagnostic near 531 nm. Nitrogen saliency is on the 680-720 nm red edge (REIP). For water stress, the system separately checks documented SWIR diagnostic features; those checks are not top saliency peaks.
- On the demo's held-out synthetic set, the broadband incumbent is near 1.0 F1 on nitrogen, while tar-spot F1 is 0.61 vs our 0.77 and water is 0.63 vs 0.97. The spectral inspector exposes the evidence behind each call instead of treating the class label as self-justifying.
- Across stress types, macro-F1 is 0.887 for the 424-band CNN vs 0.69 for Sentinel-2. The scope: a held-out 1,280-pixel set on an unseen cultivar and soil (a real distribution shift), physics-grounded synthetic scenes. A mechanism result, not an open-field guarantee.
- The typed diagnosis becomes a variable-rate map snapped to a 27 m boom, plus an EU Farm-to-Fork IPM-style record template with spectral evidence. Ambiguous pixels abstain; high-abstention zones route to ground-truth scouting.
- A model a spray boom spends money on should be interrogable, not trusted on its softmax, so we expose the salient bands per zone.
- A runnable demo, not a deployed pipeline. Live sensor and farm-management integrations are not implemented. Inspect a zone's 424-band curve and saliency trace: https://veriprajna.com/demos/hyperspectral-agriculture-ai #RemoteSensing #HyperspectralImaging #PrecisionAg
Published on X · August 20, 2026
On social media
See this post on its original platform
In our archive