NDVI camera build — Raspberry Pi guide.

The Astro Pi tutorial we adapted from uses the Rosco #2007 Storaro Blue filter, which is what the Astro Pi computers on the ISS use. Blue passes through, NIR passes through, red and most green are blocked. The red channel of the resulting image captures NIR, the blue channel captures visible blue.

However, at typical drone survey altitude (60–120m AGL), there's a real tradeoff: Rayleigh scattering of blue light by the atmosphere reduces NDVI accuracy for distant subjects. The Astro Pi works fine because the ISS is hundreds of kilometres up — the scattering averages out. Our drone is closer to the ground, so the scattering biases the blue channel.

The red filter alternative — using a Roscolux #19 fire red filter — is documented in the Step 2 appendix. Some Cohort 02 alumni have switched. We document blue here because (a) it matches the original Astro Pi tutorial, (b) Rosco #2007 is more widely stocked in PH theatre supply houses, and (c) for a starter rig, both produce usable data. If you flying mostly above 100m AGL, the red variant is worth the experiment.

Cost. MicaSense RedEdge-MX or DJI P4 Multispectral cameras start around ₱180,000 and go up rapidly. A Lumipad alumnus running a microenterprise can't recoup that capital cost in Year 1 unless they're working at scale we don't see in the network.

Quality differences are real but smaller than the price suggests:

• Commercial multispectral has 4–5 distinct narrow-band sensors (red, green, red-edge, NIR, sometimes blue). The Lumipad rig has effectively 2 (NIR ≈ red channel, VIS ≈ blue channel).
• Commercial sensors are more accurately calibrated and less affected by atmospheric variation.
• Commercial systems include factory-calibrated reference targets and certified processing pipelines.

For research-grade work or precision-agriculture applications where decisions involve significant capital (variable-rate fertilising on a 500-hectare plantation), commercial cameras are worth their cost. For the small-coop survey work most Lumipad alumni do, the Pi NoIR rig is the right tool.

Yes. NDVI is a generic measure of vegetation health — chlorophyll behaviour is similar across plant species. Cohort 02 alumni have used the rig successfully on:

• Coconut — works well, large canopy makes for stable NDVI signatures.
• Rice — works during the vegetative and reproductive stages; not useful at maturity (canopy senesces).
• Banana — works well; large leaves are easy to map.
• Mango — works for canopy-level health; not useful for fruit detection.
• Vegetables (small-scale) — possible but less reliable; small individual plants don't generate strong canopy signatures from a drone.

For partner-org deployments with very different crops (oil palm in Malaysia, rubber in Vietnam, vineyards anywhere), the rig works in principle but reference benchmarks need adapting. Email engineering@lumipaddrones.com if you're calibrating for a new crop type — we'll add it to the public reference set.

It can happen — vibration, power glitch, or low battery. The capture script writes images directly to the microSD as they're captured (no large in-memory buffer), so a reboot loses at most the current frame.

If it reboots, the auto-start service brings the script back up within ~30 seconds. The lost capture window depends on flight altitude and speed — typically a 10×10m square that the FC's grid will overlap on the next pass. Stitching algorithms generally cope with this well; no client has ever rejected a survey because of a single missing frame.

The bigger concern: a Pi that reboots repeatedly indicates a power problem (UBEC failing, bad solder joint on the 5V tap). Run the bench tests in Step 3 between flights if you suspect intermittent power.

Rosco #2007 is theatre lighting gel — designed for indoor stage lights, not direct tropical sun. Cohort 02 data: filters last 12–24 months of regular use before noticeable degradation. Symptoms: reference target shots come back warmer (more reds in the blue channel), NDVI maps drift towards lower values overall.

Replace prophylactically every 18 months or when you see calibration drift. New filter squares cost ~₱50 each (a single A4 sheet provides 100+ squares); the limiting factor is the ordering trip, not the cost.

Yes, and Cohort 02 has one alumnus doing this experimentally. Two Pi Zero 2 W modules, two Pi NoIR cameras, one with blue filter and one with red. The dual-camera setup is what the Public Lab community calls a "split-camera" rig — produces better NDVI than either filter alone because it captures pure NIR (red filter) and pure VIS (blue filter) on separate channels.

Trade-offs: doubles the weight (~190g instead of 95g — still within the v1's payload margin), doubles the build complexity, and requires camera-to-camera frame alignment in post-processing. We don't recommend this for first builds; come back to it after you've run 10+ surveys with a single-camera rig.

Documentation for the dual-camera variant is in the build kit's experimental/ folder.

The Astro Pi tutorial is excellent for the camera modification, library installation, and capture code. It assumes a stationary indoor camera looking at a dataset of pre-captured Earth-from-space images.

What this Lumipad guide adds:

• Mounting on a flying drone (Step 4) — the Astro Pi guide doesn't cover this at all.
• Field calibration with reference targets (Step 5) — important for survey work where each flight has different lighting.
• FC integration and trigger via PWM (Step 4) — for flight-controlled capture timing.
• PH-specific component sourcing and pricing.
• Real-world use cases and accuracy expectations from Cohort alumni.

Both are useful. Read the Astro Pi tutorial first to understand the underlying technique; use this guide for the practical drone implementation.

Two reasons. First, the Fastie colormap was specifically designed for vegetation imagery — its breakpoints align with the NDVI ranges that matter (0.0, 0.2, 0.5, 0.7) so that the visual distinctions match the analytically meaningful thresholds. Matplotlib's "RdYlGn" is a generic diverging colormap; it produces visually similar maps but the colour transitions don't align with NDVI's biological meaning.

Second, the Fastie colormap is the de-facto standard in the open-source vegetation-imaging community (Public Lab, Infragram, the Astro Pi project). Using it makes Lumipad surveys directly comparable with imagery from those communities.

The 256-row LUT is included in the build kit as fastiecm.py. If you want to experiment with alternative colormaps for client-facing reports, swapping is one line — but keep the Fastie version archived as your scientific baseline.