A security camera makes a picture for a person to look at. An agricultural camera makes a measurement: a record whose whole value is that it can be compared to the same view yesterday, last week, and last season. Frame it that way and the buying question changes completely. The skill is not picking a better camera. It is making one frame comparable to the next, and that is a discipline, not a purchase.
Where this page comes from
We have not built or field-tested these stations yet. Everything here is reasoned from physics, published research, and manufacturer specifications, and the page says so where it matters. The resolution figures below are arithmetic, not observation. As builds happen and get tested, the pages get updated. We would rather tell you that than pretend.
01A camera is not buying you photos.
Every feature that makes a good security camera (auto exposure, auto white balance, autofocus, noise reduction, HDR) exists to make each frame look good to a human. But every one of those features changes between frames, and change is exactly the signal you are trying to detect. If the camera brightens Tuesday's frame and cools Wednesday's colors, the yellowing you needed to catch is buried under the camera's own helpfulness.
So the discipline that makes imaging useful in a grow is the opposite of what the consumer market sells: lock everything down, keep the geometry fixed, put a reference in the frame, and shoot on a schedule. The camera becomes just another instrument, and its frames flow through the same Collect · Have · Use path as any sensor reading: the station captures and pushes the frame onward, your endpoint keeps it, and the analysis happens downstream. Keeping the raw frame matters more than it sounds. Analysis methods improve every year, and a saved frame can be re-analyzed with a better method in three years. A number extracted at the camera and thrown away cannot.
02The two questions a camera answers.
Keep these separate, because they want different hardware, different cadence, and different money.
- "What is happening in that place?" Is the cooler door shut, did the tray flood, is the pump running, what visits the pen at night. Operational state, mostly yes-or-no. Low resolution is entirely adequate, and the value is in how many eyes you can afford to hang.
- "What is the plant telling me?" Growth rate, canopy closure, color shifts, early disease, pest pressure, water stress. Agronomy. This half demands resolution, faithful color, and locked-down capture, and it is the half the consumer market serves worst.
Most of the disappointment in grow cameras comes from buying hardware for the first question and expecting answers to the second. A $50 camera watching a whole room will tell you the room is fine and never show you the mildew colony, because at that framing the colony is smaller than a pixel.
03"Infrared" is three different things.
This is the most commonly confused point in the whole field, and vendors blur it on purpose because "infrared" sounds like one capability. It is three, and they share almost nothing.
| Near-infrared (NIR) | ~700–1000 nm | Reflected light, just past what your eye sees. Ordinary camera sensors already detect it; a filter normally blocks it. Remove the filter (a "NoIR" camera) and you can see it. Healthy leaf tissue reflects NIR strongly; stressed tissue less. Cheap: $30–80 on top of a station. |
|---|---|---|
| Shortwave infrared (SWIR) | ~1000–2500 nm | Also reflected, but past what silicon sensors detect at all. Water absorption lives here, which makes it the genuine moisture band. Needs an exotic sensor: $10,000 and up per camera. Know it exists, then move on. |
| Thermal (LWIR) | 8–14 µm | Emitted heat, not reflected light. A completely different sensor measuring surface temperature. From $15 (one pixel) to $250 (a phone clip-on) to $400 and up (a real thermal module). |
The commercially important consequence: the "IR night vision" advertised on every camera is NIR illumination, a floodlight your eye cannot see. It is not thermal. It measures nothing. A camera with IR night vision and a camera with a thermal sensor are unrelated claims, and buying one when you needed the other is the most common imaging mistake there is.
One more NIR fact that surprises everyone: LED grow lights emit essentially no NIR. Point a NoIR camera at a lit grow room and it sees nothing useful. To do NIR work indoors you must bring your own infrared light, which is actually better than sunlight for the job, because your light never varies and the sun always does.
04The number that matters is millimeters per pixel.
Megapixels are a marketing number. The number that decides what a camera can actually see is how many millimeters of leaf each pixel covers: the field of view divided by the pixels across it. Here is the arithmetic for a camera framing a 4-foot bench:
| ESP32-CAM (2 MP) | 1600 px wide | 0.76 mm per pixel |
|---|---|---|
| ESP32-S3 + OV5640 (5 MP) | 2592 px wide | 0.47 mm per pixel |
| Raspberry Pi HQ Camera (12.3 MP) | 4056 px wide | 0.30 mm per pixel |
| 12 MP IP camera | 4512 px wide | 0.27 mm per pixel |
Now the biology. A spider mite is about 0.4 mm. Thrips run 1 to 2 mm. An aphid is 2 to 3 mm, and an early powdery mildew colony about the same. Detecting something reliably takes roughly 8 to 10 pixels across it, not 2 or 3. So at 0.76 mm per pixel an aphid is three pixels and a mite does not exist, and no software downstream can recover what the sensor never captured. At 0.27 mm per pixel that aphid is nine pixels and the mildew colony is ten. That is a real threshold crossing, not a marketing increment.
Here is the part that saves you money: framing buys resolution more cheaply than sensors do. A 2 MP camera twelve inches from a leaf sees far more than a 12 MP camera across the room. So the question is never "how many megapixels." It is "how small is the thing I need to see, and how many pixels will land on it." Run your own numbers before you buy anything:
05Four things that outrank megapixels.
- The lens is usually the real limit. A $130 camera puts a 12 MP sensor behind a molded plastic lens that cannot resolve fine enough detail to feed it. You get 12 million pixels holding maybe 5 megapixels of actual information. This is the biggest single gap between a sealed camera and one that takes a lens you choose.
- The camera's own processing deletes your signal. Closed cameras apply noise reduction tuned for security footage, and noise reduction specifically targets small high-frequency detail, which is a precise description of a thrips at nine pixels. On a closed camera you cannot turn it off.
- Spec-sheet resolution is not always real resolution. Some sensors reach their headline number through pixel-grouping tricks and give back less raw detail than the number implies. Two cameras with the same megapixel count can capture meaningfully different information.
- Shape matters. Security cameras shoot a wide letterbox shaped for a driveway. A bench from above is squarish and a plant from the side is tall, so a third of the letterbox frame is floor and wall you paid for. A 4:3 sensor is simply the right shape for this work.
06Where the camera rides matters as much as what it is.
| Fixed station | $8–150 per view | One view, forever, on a schedule. The only platform that produces a truly comparable time series without effort, because the geometry never changes. The primary platform here. |
|---|---|---|
| The phone in your pocket | $0 | Better sensor and optics than anything in the fixed tier, already paid for. With a reference card in the frame it becomes a real instrument. The most underrated platform in agriculture. See the Station Card. |
| Rail or gantry scout | $500–5,000 | A camera on a track visiting fixed stops. Per-plant coverage with fixed geometry, and the practical way to run infrared work in the dark cycle. Future territory; the concept is placed here so you know where it sits. |
| Drone | $1,500–10,000+ | Genuinely useful above roughly 20 to 50 acres of row crop, plus licensing and flight time. For a two-acre diversified grower or an indoor operation, a drone is the wrong tool, and plenty of vendors will not tell you that. |
| Satellite | Free | Public programs give free imagery at 10 to 30 meters per pixel. Useful for field-scale trend on big acreage; useless under about 5 acres, where one pixel is bigger than the whole plot. Free is worth knowing about either way. |
| Implement-mounted | Commercial | Cameras on sprayers doing weed detection and targeted application. Real, active, and priced for large operations. Placed so you can locate what you have heard of. |
07The honest tiers.
Each tier is named for the job it is honestly good at, which is the only spec that matters. These are our engineering conclusions from published specifications, stated as such. Every tier now has its own build page on the Configs shelf: click through a row for the parts, the discipline, and what software is downloadable today versus coming.
| Your phone + a reference card | $0 | The diagnostic tier. Resolution on demand, wherever you point it. Needs a human, so its weakness is cadence, not quality. |
|---|---|---|
| Raspberry Pi + HQ Camera + a lens you choose | ~$90 | The fixed plant station. Full manual control of exposure and focus, the capture control that makes 200 days of frames comparable. Needs mains or PoE power. |
| Raspberry Pi + Camera Module 3 | ~$40 | The budget plant station. Same discipline, less optical control. Lock the autofocus once and never touch it; focus hunting across a season destroys comparability. |
| ESP32-S3 camera on solar | ~$15 | The off-grid eye. Sleeps at almost nothing, wakes, shoots, pushes, sleeps: runs indefinitely on a small panel and one battery cell, which a Pi simply cannot do. For phenology, stand establishment, and damage events in the field, where 2 to 5 megapixels answers the question. |
| ESP32-CAM, hung densely | ~$8 each | The witness tier. Is the tray flooded, is the door shut, is the pump running. Twenty of these across a site is a different capability from three good ones. State, not measurement: just do not call it agronomy. |
| The IP camera you already own | $0–150 | Weatherproof, powered, mounted, already installed. Nearly every wired one can hand over a single still on request or push one to your endpoint on a timer, no video stream involved. Wide context and place-watching; its closed processing keeps it out of the plant tier. |
| Thermal, from spot to clip | $15–250 | A ladder, not a product: a one-pixel spot thermometer, a coarse 32×24 grid for a single bench, a 256×192 clip for the phone you already carry. Thermal is its own story, told at CWSI: the image only means something next to air temperature and humidity, which your sensor layer already measures. |
Why the $8 board is not the flagship
Eight dollars is the board, not the station. A deployed station is an enclosure, a mount, power, a cable run, and an hour of your labor: call it $60 to $120 either way. The cheap board saves about $30 on the one component that decides the data quality for the station's whole life, and its dated sensor loses exactly the things the plant question needs: color fidelity and resolution. Where there is no power, it wins outright, and it is not close. Where there is power, spend the $30.
08Capture discipline: the rules that make frames comparable.
None of this costs money. All of it is the difference between a folder of pictures and a record you can act on.
- Lock exposure, white balance, gain, and focus. Set once at the working distance. Auto anything varies between frames, and variation is the signal you are trying to detect.
- Scheduled stills, not streams. A frame at 10:00 every day under the same light is a time series. Continuous video is a security habit wearing an agronomy costume, and it will eat your network for nothing: one still every fifteen minutes is trivial bandwidth where streaming 4K never is.
- Record the light state with every frame. Lights on, lights off, daylight. A frame that does not say is a frame you cannot compare.
- A reference in every frame. The Station Card carries the scale, the color reference, and the station's identity, which is what lets a phone, a fixed station, and next year's replacement camera produce data you can lay side by side.
- Never mix the light worlds. A NoIR camera belongs in the dark with its own infrared light. A color camera belongs in the light. A NoIR camera in a lit room gets the worst of both: wrecked color and no usable infrared signal.
- Thermal cannot shoot through glazing. Greenhouse poly, glass, and tent windows are opaque to heat radiation. Not degraded: opaque. A thermal camera outside the glazing is measuring the glazing. Every thermal eye goes inside the envelope, no exceptions.
09What it costs to keep.
Keeping the raw frames is the whole point, and it is the first design choice here that makes storage a real line item, so here is the arithmetic up front. High-quality 12 MP stills every fifteen minutes run on the order of 2 GB per camera per month. Thirty cameras is roughly 60 GB a month, 700 GB a year. That is cheap, a single ordinary hard drive per year, but it is not free, and a grower planning thirty stations should know it before, not after.
10A camera sees people.
A temperature probe watches a number. A camera in a grow space watches whoever walks through it, and the moment there is hired labor in the frame, where the images live, who can see them, and how long they are kept are real questions with legal weight in some places. This is part of why grower-owned matters: on your own storage, under your own retention rules, those questions have answers you control. In a vendor's cloud they have whatever answers the terms of service say this year. No camera vendor writes this paragraph, which is exactly why it is here.
11Reading any vendor's spec sheet.
We are not going to tear down products by name. We would rather hand you the questions, because the questions do not go stale:
- Does "thermal" come with an array size? If a product says thermal imaging and prints no resolution, assume the coarsest part available. A vendor with a good array prints the number.
- Is "IR night vision" being sold as thermal? They are unrelated. Many products have both and describe them as one thing.
- What does "true VPD" mean here? Air VPD from temperature and humidity is what every controller already does. Leaf-level VPD needs leaf temperature, which is real and better, but only if it is computed as a difference against air by the same instrument. Ask how.
- Can you get your images out? A snapshot URL, a scheduled push to your own endpoint, a real export. If the plant's whole visual history lives in the vendor's cloud in a format you cannot take with you, you are renting your own records.
- What happens when the company stops supporting it? A cloud-dependent camera is a subscription with a lens on it.
The shortest version
An image is a measurement, so treat the camera like an instrument: lock the settings, fix the geometry, put a reference in the frame, shoot stills on a schedule, and keep the raw frames somewhere you own. Buy resolution by framing before you buy it by megapixel, know that night-vision IR is not thermal, and match the camera to the question: cheap eyes for watching the place, a disciplined station or the phone in your pocket for watching the plant.
Frequently asked questions.
What camera resolution do I need to see pests like spider mites or aphids?
Think in millimeters per pixel, not megapixels: the width the camera sees divided by its pixel count. Reliable detection takes roughly 8 to 10 pixels across the pest. An aphid (2 to 3 mm) needs about 0.3 mm per pixel, which a 12 MP camera reaches across a 4-foot bench and a 2 MP camera reaches only a foot from the leaf. A spider mite (0.4 mm) is out of reach of room-scale framing entirely; that is a close-up job for a phone or a tightly framed station.
Is IR night vision the same as thermal imaging?
No. IR night vision is an invisible floodlight (near-infrared light the camera can see and you cannot) that makes an ordinary picture in the dark. Thermal imaging uses a completely different sensor that measures the heat surfaces give off. Night vision cannot read temperature, and a thermal sensor does not make night-vision pictures. Vendors often blur the two under one word; the spec that separates them is a stated thermal array resolution.
Can I use a regular security camera to monitor plant health?
For watching the place (doors, pumps, floods, animals), yes, and most wired ones can hand your own endpoint a scheduled still with no cloud involved. For measuring the plant, a security camera fights you: its automatic exposure and processing change between frames, its lens rarely delivers its sensor's resolution, and its wide letterbox wastes a third of the pixels. It is genuinely useful context, but the plant record wants a camera you can lock down.
Do I need a drone for a small farm?
Almost certainly not. Drones earn their cost above roughly 20 to 50 acres of field crop, where covering ground is the problem. Under a few acres, fixed cameras and the phone in your pocket cover the same questions for a fraction of the cost, on a schedule a drone cannot match. Free public satellite imagery covers the field-scale trend question at zero cost, though only usefully above about 5 acres.
Why do my grow photos look different every day even though nothing changed?
Because the camera is adjusting itself: auto exposure, auto white balance, and autofocus re-decide on every frame, so the same plant photographs differently day to day. That is fine for snapshots and fatal for comparison. Lock the exposure, white balance, and focus, keep the camera and lighting fixed, and put a color and scale reference in the frame, and the day-to-day change you see becomes the plant's, not the camera's.
Can a thermal camera see plant water stress through a greenhouse wall?
No. Greenhouse glazing (poly, glass, or film) is opaque to the heat radiation thermal cameras read, so from outside you are measuring the wall's temperature, not the canopy's. The readings will look plausible and mean nothing. Thermal cameras must be inside the growing envelope, and the temperature they read only becomes a stress signal next to air temperature and humidity measured in the same place.