A harvesting robot has a fraction of a second to decide whether a piece of fruit is ready. In that time, the vision system must distinguish ripe from unripe fruit, identify defects, and separate produce from the background.
Color cameras handle part of that task. Many of the most useful indicators of crop health and ripeness sit outside the visible spectrum. Chlorophyll breakdown, water stress and early fungal infection all change how a plant reflects light outside the visible band before they change how it looks.
The challenge is integrating that spectral capability into a platform with strict limits on space, power, weight, and reliability.
A standard color sensor samples the scene through three broad overlapping filters spanning roughly 400 nm to 650 nm. That is enough to reproduce what a person would see, and not much more.
Chlorophyll absorbs strongly around 670 nm, while healthy leaf tissue reflects strongly beyond 750 nm, where the internal air-cell structure of the mesophyll scatters light. The transition between the two, the red edge near 700 nm to 730 nm, shifts position as chlorophyll content changes, and it shifts before the leaf looks any different. Water content appears further out, in absorption features near 970 nm and 1450 nm.
During ripening, chlorophyll degrades while carotenoids and anthocyanins accumulate, altering reflectance across roughly 500 nm to 600 nm. Much of this information is lost when reduced to RGB data.
Two samples with measurably different chlorophyll concentration can return near-identical RGB values. Adding a narrow band centered around 720 nm can separate them clearly.
One approach is to dedicate a separate camera to each wavelength band.
On a bench that works. On a moving platform it brings problems that scale badly.
Parallax comes first. Two cameras separated by 40 mm, looking at fruit 300 mm away, see the scene from different angles, and the resulting disparity varies with the distance to each object in the frame. A single calibration transform cannot correct it, because the correction depends on depth. Agricultural imaging works at short range, which is where the error is worst.
Sequential capture has the same problem in time rather than space. A filter wheel or a switched illumination scheme collects bands at different instants, and anything moving between frames, including the robot itself, breaks the registration.
System-level complexity increases as well. Each camera adds a lens, a mount, a cable, a share of the interface bandwidth, a power draw, a calibration file someone has to maintain, and another sealed enclosure keeping out dust and washdown water.
Relative alignment between cameras drifts with temperature and vibration, so a system registered at commissioning is not necessarily registered a season later.
For agricultural robots operating on moving platforms, simultaneous acquisition matters as much as spectral content. Capturing all bands in a single exposure avoids registration errors caused by vehicle motion, changing working distance, or moving targets.
Micropatterned filters use a different approach. Instead of one filter per camera, multiple filter regions are patterned photolithographically onto a single substrate, which then sits in front of, or directly on, one sensor.
Three filter architectures are commonly used in agricultural imaging:
1. Mosaic Arrays
2. Stripe Filters
3. Linear Variable Filters
Torrent patterns dielectric, metallic, and color filter arrays on a single substrate, with custom bands across UV, VIS, NIR and SWIR, and up to nine bands on one part.
Patterning can be performed on glass or semiconductor wafers using semiconductor-style mask aligners handling wafers up to 8 inches, supporting cost-effective volume production.
A small number of design parameters can strongly influence system performance. Band selection should start with the target reflectance features, not a filter catalog. A 20 nm band centered at 717 nm and a 20 nm band centered at 740 nm look similar on a datasheet and behave differently on a canopy.
Band count trades against spatial resolution. Nine bands at 0.55 MP each may be right for a ripeness classifier and wrong for a system that must also locate a stem accurately enough to cut it.
Angle of incidence matters more than most first-time users expect. A dielectric bandpass filter shifts toward shorter wavelengths as incidence angle increases, following the standard relation for a thin-film stack of effective index n. At 20 degrees with an effective index of 1.8, the shift is close to 1.8 percent, which is 14 nm at 800 nm. The exact shift depends on filter design and stack structure, but the effect is significant enough that angle-of-incidence must be considered early in the optical design.
For a 10 nm passband, this can exceed a full bandwidth across the field of view. Common solutions include telecentric optics, wider passbands, or spatially graded filter designs.
Separation between the filter plane and the photodiode drives spectral crosstalk. Rays at the edge of the cone from one filter cell walk laterally as they propagate, and where that walk exceeds the pixel pitch they land under a neighboring band. At f/2.8 the marginal ray sits roughly 10 degrees off axis, so 400 microns of cover glass and microlens stack allow around 70 microns of lateral travel. On a 3.45-micron pixel that is twenty pixels across. Reducing the standoff is the whole reason pixel-level patterning and direct-to-sensor integration exist.
Substrate and coating durability decide whether the part survives the application at all. Field equipment sees humidity cycling, temperature swings and chemical cleaning. Ion-assisted deposition and sputtered dielectric stacks are dense enough to hold their spectral position under those conditions.
Alignment tolerance and registration to the pixel grid set the practical floor on feature size.
The same approach can be applied to several other agricultural and imaging applications. Weed and crop discrimination for targeted spraying uses much the same red-edge and NIR bands, with the added requirement of working at vehicle speed.
Disease detection generally needs narrower bands and better radiometric stability, because the signal is a small change inside an absorption feature rather than a large change in overall reflectance. Post-harvest sorting lines have the luxury of controlled illumination and fixed geometry, which relaxes the angle-of-incidence budget considerably.
Remote sensing from UAV and satellite platforms places greater emphasis on size, weight, and power reduction, making single-sensor approaches particularly attractive.
Read more: How multispectral imaging cameras are transforming agriculture
Whether a patterned filter is the right choice depends on the required spectral bands, working distance, optical speed, and production volume. These decisions are best made before the mechanical design is finalized, because filter geometry and sensor selection constrain one another.
Our micropatterned filter capabilities include custom band definition, filter layout, and wafer-level production across UV to SWIR. Where the filter needs to sit closer to the photodiodes than the sensor package allows, image sensor modification is a related route that changes which geometries are available.
Contact our technical sales team to discuss your optical requirements: sales@torrentphotonics.com