In 2014 one of our OMAFRA summer students designed a short-and-gritty demonstration using a backpack sprayer, a variable-speed fan and some water-sensitive paper positioned downwind at 1.5 metre intervals. The intent was to illustrate how sprayer operators could reduce the potential for off-target drift by recognizing and accounting for three factors:
Apparent wind speed (i.e. the sum of wind speed and travel speed)
Boom height (i.e. release height)
Droplet size (i.e. nozzle spray quality)
Apparent Wind Speed
Spray operators know they should not spray when the air is calm or when the wind is too high, but they often forget that the nozzles experience “apparent wind speed” which means driving 10 km/h into a 10 km/h headwind is essentially spraying in a 20 km/h wind.
The result of spraying with a Medium spray quality in 10 km/h and 15 km/h wind: water-sensitive papers indicated that there is more downwind drift in higher winds.
Boom Height
Spray operators raise their booms to ensure their nozzles clear the crops, but this contributes to off target drift and greatly reduces coverage – particularly when using twin-fan style tips. Dr. Tom Wolf explains how to set your boom height here, or you could watch one of our Exploding Sprayer Myths videos on the subject.
The result of spraying with a Medium spray quality in a 10 km/h wind at 50 cm and 100 cm from the ground: water-sensitive papers indicated that downwind drift increases as the boom gets higher.
Droplet Size
The coarser the spray quality, the less likely the spray will drift off target. Remember, for a given volume, shifting to larger droplets means fewer droplets. Application volumes may have to increase to compensate for potentially reduced coverage.
The result of spraying with a Medium spray quality versus spraying with an Extremely Coarse spray quality: water-sensitive papers indicated that there is more downwind drift from smaller droplets.
Take-Home
This demo used percent coverage as a metric, which is convenient but greatly underestimates drift. So even when the spray window is small and the spray has to go on, take a moment to drop the boom, use a coarser droplet size and if it’s too windy, just don’t spray.
WUR Drift Calculator
There are many drift calculators available for home use. Some require more expertise than others to get a reliable result. This free downloadable calculator from Wageningen University & Research was made available in 2021. It can quantify spray drift deposits onto surface waters and non-target terrestrial areas near a sprayed field or orchard
The calculator uses statistically obtained regression curves to calculate spray deposition next to the sprayed field. The spray drift curves are based on the latest experimental data for field crops, fruit orchards and avenue tree nurseries.
In June 2013 we ran a ginseng spraying workshop and we learned as much as the growers did. Ginseng is notoriously difficult to spray:
It is highly susceptible to pathogens given the high humidity and still conditions generally found under the shade structure.
It forms a solid ceiling of leaves that resist spray penetrating to the stem and crown below and makes under-leaf coverage very difficult to achieve.
Many growers have (wisely) walked away from the old Casotti sprayers, which have been shown to give erratic coverage at best. They have adopted the Arag Microjet system with it’s characteristic orange shields. The >$80.00 CAD price tag for each nozzle is due to the brass mixing valve and swivel joint, as well as import costs from Italy. Contrary to popular belief, it does not use air-assist, or air-induction – it is strictly hydraulic. It does tend to create a ‘wake’ of air movement at high pressure. This phenomenon is called air entrainment and it is caused by large droplets travelling at high speed.
Classic Arag microjet nozzles.
This nozzle is essentially the business-end of a spray gun. The way it is used in ginseng it works more-or-less like a hollow cone disc-core assembly. This begs the question “Why not use the cheaper and more readily available ceramic disc-core?” We set out to compare the two options using water sensitive paper set within the canopy. These yellow, paper targets turn blue when sprayed, clearly showing spray coverage.
Location of water sensitive papers in the ginseng canopy.
Determining rates
The first step was to determine the output rate for each nozzle. Generally, nozzle manufacturers provide rate tables showing how much volume a nozzle emits by time (e.g. US gallons per minute) at a given pressure. Finding these tables for the 1.5 millimetre Arag Microjet proved difficult. When we finally found one, it was discovered the rates were established for 200 to 850 pounds per square inch. This is excessively high pressure for a typical boom sprayer, so tables had to be developed for lower pressures.
Classic Arag microjets have a mixing valve that opens the spray up into a hollow cone (valve handle left or right), or collapses it into a tight stream (valve handle middle). The valve position also changes the rate. It can never be shut off completely, and it’s hard to adjust consistently.Determining nozzle rate using the Innoquest Spot-On SC-4.
Further, given the odd design of the mixing valve, it was determined that moving the handle ~10 degrees left of centre, or ~10 degrees right of centre, gave a difference of as much as 60%. The table below shows the outputs for a 1.5 millimetre nozzle with the handle in both positions and the two graphs show the results… well… graphically. Outputs were determined using the Innoquest Spot-On SC-4, but the frothing effect created by the nozzles may have created minor errors. Each rate is the average of a minimum of three samples.
Valve Setting
Pressure (psi)
Avg Output (gpm)
Pressure (bar)
Avg Output (L/min)
10 degrees left
40
1.02
2.76
3.86
10 degrees left
50
1.1
3.45
4.16
10 degrees left
60
1.25
4.14
4.73
10 degrees left
70
1.25
4.83
4.73
10 degrees left
80
1.38
5.52
5.22
10 degrees left
90
1.4
6.21
5.3
10 degrees left
100
1.45
6.89
5.49
10 degrees left
110
1.6
7.58
6.06
10 degrees left
120
1.75
8.27
6.62
10 degrees left
150
1.87
10.34
7.08
10 degrees left
200
2.2
13.79
8.33
10 degrees right
40
0.65
2.76
2.46
10 degrees right
50
0.7
3.45
2.65
10 degrees right
60
0.8
4.14
3.03
10 degrees right
70
0.85
4.83
3.22
10 degrees right
80
0.9
5.52
3.41
10 degrees right
90
0.9
6.21
3.41
10 degrees right
100
1
6.89
3.79
10 degrees right
110
1.07
7.58
4.05
10 degrees right
120
1.1
8.27
4.16
10 degrees right
150
1.25
10.34
4.73
10 degrees right
200
1.37
13.79
5.19
Average 1.5 mm ARAG Microjet output at a range of pressures and two valve settings in US Imperial units.Average 1.5 mm ARAG Microjet output at a range of pressures and two valve settings in Metric units.
Comparing nozzles
Using the grower’s typical ground speed of 5 km/h (~3 mph) and operating pressure of 6.9 bar (100 psi), we found four TeeJet disc-core combinations that emitted a hollow cone pattern and approximately the same output as the Arag Microjets. The five nozzles sets tested were:
ARAG Microjet® 1.5 mm = ~0.95 US g/min avg at 100 psi
TeeJet® D8-DC25= 0.97 US g/min at 100 psi= ~97° cone
TeeJet®D7-DC45= 0.97 US g/min at 100 psi= ~81° cone
TeeJet®D4-DC46= 0.88 US g/min at 100 psi= ~33° cone
TeeJet®D6-DC45= 0.93 US g/min at 100 psi= ~81° cone
We did not use nozzle drop hoses (aka drop arms or hose drops) because it has already been firmly established that they are absolutely required to achieve under leaf coverage See OMAFRA factsheet 10-079 and this article.
Observations
While there were some complications with setting up the papers for the demo, we observed the following:
The output of each Microjet nozzle can be as much as 50% more or less than expected without being visually detectable and output for each nozzle must be confirmed before spraying. Therefore, outputs should be confirmed before every application.
Microjets at 100 psi emitting ~890 L/ha (~95 US gallons per acre) gave satisfactory coverage on all upward facing targets, but unsatisfactory under-leaf coverage. This has been demonstrated many times before.
The TeeJet D7-DC45 combination emitting a similar rate gave satisfactory coverage on all upward facing targets, but unsatisfactory under-leaf coverage. They may be a viable alternative to the Microjets.
Nozzle drops are advised to achieve under-leaf coverage.
The demo also raised some questions:
Did the TeeJet disc-core push the canopy apart as much as the Microjet? The audience noticed there was some leaf-shadowing where the cards did not get complete coverage using disc-core. This might have been coincidence, or it may not have. This question will be addressed in a research trial next season, but for now, the D7-DC45 appeared to give similar coverage to the Microjet.
Can nozzle drops be avoided if pressure is raised to 27.5 bar (400 psi)? Thanks to one grower trying this experiment in his garden after the demo, we saw some under-leaf coverage is possible at such high pressures, but this occurred at the cost of a lot of noise, diesel fuel and considerable wear on the ceramic Microjet discs. The grower tested these tips and discovered they needed replacement after only two years of use. Nozzle drops are cheaper, easier and result in considerably more spray in the under leaf positions.
We saw what minimal and excessive foliar coverage looked like, and determined how much variability there was from one nozzle to another. A significant question was “How much spray can be saved when using a more accurate application?” and the answer is yet to be determined, but could be well in excess of 10% of the typical spray volume. Given that this crop can be sprayed more than 100 times over it’s 3 or four years before harvest, this represents significant savings in pesticides and refill time.
Additional – Newer ARAG Microjet Design
Since this work was performed, growers have been exploring a newer option from ARAG.
They are an improvement over the older version insofar as they are more easily calibrated and held at a given rate thanks to a lock nut. They still employ a 1.5 mm diameter ceramic disc, but this can be changed for a 1.0 or 1.2 quite easily. They are still somewhat finicky when trying to set a consistent spray quality and rate from nozzle to nozzle, but are better than the mixing-valve option.
Custom-made ginseng sprayer. A standard design with newer, cheaper and easier-to-use ARAG microjets.
Special thanks to Syngenta Canada for providing lunch, to C&R Atkinson Farms Ltd. for hosting, to TeeJet for supplying the disc-cores and water-sensitive papers, and to Dr. Sean Westerveld, Dr. Melanie Filotas and OMAFRA summer student Megan Leedham for contributing to the workshop.
This is the final part of our three-part article discussing methods for digitizing and processing water sensitive paper. You can read part one here and part two here.
Morphological operations
We can now move on to the larger shapes, or “morphology” of the objects in our binary image. Our goal is to quantify deposits by interpreting these shapes. Once again, these operations are powerful processing tools, but we must acknowledge three overriding limitations:
1. Inconsistent stains
Sometimes deposits do not create a consistent blue colour – they can get lighter or take on a greenish-yellow hue towards the perimeter of the stain. During thresholding, the outer edge can be accidently eroded, leaving behind an object with a jagged edge. This may lead us to underestimate the percent area actually covered. In the case of tiny stains, it might eliminate them entirely and lead us to underestimate deposit density.
2. Overlaps
It can be difficult to determine if an object represents a stain from a single droplet or is the result of multiple, overlapping deposits. This becomes significant when the surface of the WSP exceeds ~20% total coverage. The resulting objects may or may not have hollow centres where droplets do not overlap entirely. Misidentifying overlaps leads us to falsely conclude that an object is the result of a single, coarser droplet rather than multiple finer droplets.
3. Ellipses
Non-circular stains are formed when droplets scuff along the surface. Two droplets with the same volume encountering a paper at different angles can create stains with significantly different areas. We may wrongly conclude that the droplets that created them were coarser than they truly were. One approach is to use Feret’s Diameter (aka Caliper Diameter) by measuring the widest spans on the X and Y axes and taking the average. Another approach is to interpret the ellipse as a series of circular stains. Or we can decide to only acknowledge these objects when calculating percent area covered, but omit them when calculating deposit density or predicting original droplet size. Each strategy is a compromise, so it is important to be consistent and transparent when reporting results.
Three common problems when analysing water sensitive paper.
We’ll explore two morphological operations that can help us separate fact from fiction: Granulometry and Dilation-and-Erosion. We’re introducing these operations as part of the processing and detection step, but they may also overlap with the measurement step in our three-step process.
Granulometry
We can estimate the range of object sizes and get a sense of how they are distributed on the paper by filtering or “sieving” the image. Imagine pouring a mixture of sand and rocks through a series of ever-finer sieves. Doing so allows you to separate particles based on size exclusion. A granulometry function compares each object to a series of standardized objects with decreasing diameters. This isolates objects of a similar size and bins them in that size range. This is a powerful operation, but accuracy is lost when stains overlap to form larger objects. In this case, we move on to Dilation and Erosion.
Dilation and Erosion
Think of dilation as adding pixels to the boundary of an object. This makes tiny objects bigger, fills in any interior holes and can cause objects to merge. The number of pixel-wide dilations required to make objects contact one another can be used as a measure of deposit density.
Erosion removes pixels from the outer (and sometimes inner) boundaries of an object. This eliminates tiny artifacts that may not actually represent stains. It can also split non-circular objects into multiple parts before shrinking them into multiple nuclei (aka centroids). These last-remaining points are not necessarily the centre of a stain, but the pixels furthest away from the original boundary.
When a non-circular shape has more than one nucleus, they likely represent individual droplets that combined to form the larger stain. We can then use these nuclei to measure deposit density, such as in a Voronoi partition which triangulates each nucleus in relation to the two closest neighbours.
Many image processers use both these operations sequentially. When an image is eroded and then dilated (a process called “Opening”), smaller objects are removed, leaving the area and shape of remaining objects relatively intact. Dilating and then eroding (a process called “Closing”) fills in small holes and merges smaller objects, once again leaving the area and shape of remaining objects relatively intact. We can use both of these functions to help smooth an image prior to measurement.
(Top) Opening operations erode and then dilate the image. Moving left to right, the smaller objects tend to disappear. (Bottom) Closing operations dilate and then erode the image. Moving left to right, smaller objects either disappear or merge and holes are filled in
Distance Transformations
Distance transformations are advanced operations specifically used to separate objects that are densely packed. While not typically used when analyzing WSP, distance transformations are another means of identifying object nuclei. They are another means for teasing apart objects that are likely the result of overlapping deposits and then mapping their relative sizes and positions.
Measurement
The calculation of the area covered by deposits is straightforward. The pixels belonging to objects (the deposits) and those belonging to background are summed and then the fraction is converted to percent area covered. Research has shown that the image resolution does not significantly impact percent coverage assessments and has suggested that all image analysis software tends to produce similar results (+/- 3.5% observed when the same threshold was applied to multiple papers). This is acceptable because it’s within the variability inherent to spraying.
We ran a similar experiment wherein we analyzed the same piece of WSP using four methods. Here are a few facts about the software we used:
DropScope produces images between 2,100 and 2,300 DPI. Currently, it ignores ellipses and doesn’t count anything spanning less than ~35 µm (3 pixels).
We set ImageJ to ignore any object spanning less than 3 pixels, which at 2,400 DPI was 30 µm in diameter.
We are unaware of Snapcard’s processing methods except that the software was benchmarked using ImageJ. Developers note it will underestimate the percent area covered if the image is out of focus. (Note: As of 2026, this app may no longer be supported by the GRDC).
The images shown in the figure below were cropped from screenshots produced by each method. The actual ROI analyzed was ~3 cm2 for SnapCard, 3.68 cm2 for DropScope and 2.0 cm2 for both Epson/ImageJ methods. Our results indicate an +/- 4% difference in percent area coverage. This variability reflects the results of a 2016 journal article that compared SnapCard with ImageJ and other leading analytical software. That study claimed no statistically significant difference in percent coverage detected (standard deviations were about 20%). However, the ImageJ results tended to trend several percent higher than SnapCard. We saw this as well. And so, while resolution may not have a significant impact on percent area covered, there does appear to be some correlation.
Percent area covered as reported by three image analysis systems. Only a minor difference was observed when resolution was doubled using the Epson/ImageJ method.
Resolution definitely affects deposit counts. Particularly in applications that employ finer droplets. Consider the difference between detecting or missing 1,000 30 µm diameter objects. It may only amount to a fraction of a percentage of the surface covered, but +/- 1,000 objects on a 2 cm2 area is significant in terms of deposit density.
Output
Once a WSP image (or set of images) has been scanned, pre-processed, processed and measured, we will receive some manner of output. Some software packages create an attractive report with images, graphs and key values. These reports include percent coverage and many provide droplet density. Deposits may be binned by size, or spread factors are used to calculate the original droplet diameters and even estimate the volume applied by area. Other software packages provide raw data that can be imported into a statistical program or spreadsheet program like Excel for further analysis. Some software packages provide both.
How far can we take this?
Blow-by-blow data analysis is beyond the scope of this document, but how much weight should we give to coverage data obtained using WSP? The answer depends on the metric in question, but in all cases we must first acknowledge the three overriding caveats. Take it as said that they apply to everything that follows:
Different brands (and even different production runs) of WSP can produce significantly different coverage metrics. When conducting experiments, use a single brand of WSP. Better still, use papers from the same production batch whenever possible.
The same of piece of sprayed WSP can produce significantly different results depending on the software and protocol used to analyze it. When conducting experiments, use the same software and assessment protocol and be transparent about the process when communicating results.
WSP coverage may not reflect the coverage achieved on an actual plant tissue surface. It is suitable as a relative index (I.e. papers can be compared to papers, but not to tissues) but the spread factor changes with surface wettability and the surface tension of the liquid sprayed. Note the differences in percent area covered in the following experiment with an organosilicone super-spreader:
Difference in deposit spread on water sensitive paper versus a leaf surface using an organosilicone super-spreader and UV dye. The same volume was applied in each case and while the area increased two-fold on WSP it increased ~10-fold on an actual leaf. Image reproduced from work by Robyn Gaskin, Plant Protection Products, New Zealand.
Recall that we started this document by listing the four pieces of information commonly sought using WSP. They were listed in order of reliability, and now we can explain why.
The percent surface area covered: We have established that this is the most reliable piece of data. Droplets do not spread on WSP the way they do on plant surfaces, so it will underestimate actual coverage. The results vary by analytical method, but it’s likely not dependent on resolution and still falls within the variability inherent to spraying. This metric gives us valuable and actionable information. We can say whether or not we hit a target, and evaluate whether a sprayer change resulted in more or less deposit.
The density of deposits on the target area: We have established that that there are limits to the reliability of this metric. It is affected by the analytical method used and can be greatly underestimated when resolution is poor or when deposits overlap in high numbers. Also, it will never reliably reflect deposits under 30 µm. Nevertheless, under controlled conditions this information does have value and is of great interest in enquiries about drift and contact fungicides.
The size of the droplets that left the stains: This metric is highly questionable except under controlled conditions. The many assumptions about surface tension, droplet speed, and droplet evaporation make it impossible to make definitive statements about spray quality. Finer droplets are greatly underestimated in this equation. Therefore, while there may be some value in using WSP as a relative index, this metric is a crude indication at best.
The dose applied to the target surface: This metric has not been discussed up to this point, but is quickly and easily dismissed. Let’s assume that a droplet with a high concentration of an active ingredient will leave a stain that is the same area as another droplet with a lower concentration. This will lead some to suggest that as long as the original concentration is known, we can back-calculate the dose (which is the amount of active on a given area). However, one droplet has the same volume as eight droplets that are half it’s diameter. This cubic relationship means that if they all deposit, the larger droplet will cover roughly 1/2 the surface area as the eight smaller droplets. Therefore, the smaller droplets spread the same amount of active over a greater area. Spread factor muddies this a bit, but ultimately it means that dose cannot be estimated from area covered. Dose is better assessed using collectors that permit the residue to be removed, such as Petri dishes, Mylar sheets, pipe cleaners, alpha cellulose cards, or glass slides.
And so, the image analysis process described here is powerful and effective when used with water sensitive paper as long as the limitations are acknowledged. The same process can also be used with dyes and specialized collectors such as Kromekote to permit even greater resolution. But that’s another story.
Marçal, A.R.S., Cunha, M. 2008. Image processing of artificial targets for automatic evaluation of spray targets. Trans. of the ASABE. 51(3): 811-821.
Moor, A., Langenakens, J., Vereecke, E., Jaeken, P., Lootens, P., Vandecasteele, P. 2000. Image analysis of water sensitive paper as a tool for the evaluation of spray distribution of orchard sprayers. Aspects of Applied Biology. 57.
Panneton, B. 2002. Image analysis of water‐sensitive cards for spray coverage experiments. Applied Eng. in Agric. 18(2): 179‐182.
Salyani, M., Zhu, H., Sweeb, R.D., Pai, N. 2013. Assessment of spray distribution with water-sensitive paper. Agric. Eng. Int.: CIGR Journal. 15(2): 101-111.
SnapCard website. University of Western Australia and the Department of Primary Industries and Regional Development, Western Australia. (Note: As of 2026, may no longer exist).
Syngenta. 2002. Water‐sensitive paper for monitoring spray distributions. CH‐4002. Basle, Switzerland: Syngenta Crop Protection.
Turner, C.R., Huntington, K.A. 1970. The use of a water sensitive dye for the detection and assessment of small spray droplets. J. Agric. Eng. Res. 15: 385-387.
OK, fine. We confess to the shameful use of click-bait in our title. Nevertheless, it’s absolutely true: Drift can be good. The reason this statement is unsettling is because of the lack of context, which is really what this article is about.
The majority of sprayer-related information available to ag stakeholders relates to horizontal boom sprayers. Most of it is relevant to broadacre field crops and often pertains to herbicide applications. If you’re unsurprised, or still struggling to grasp our point, it’s likely because you’re part of that world. But consider everyone else.
Agricultural spraying is diverse and many usage patterns are grossly underrepresented. As a result, those operators struggle to find relevant information. And what information is most readily available? Yes – broadacre herbicide spraying. Even the experts (i.e. agronomists, salespeople and consultants) often make the error of responding to specialty crop questions with field crop answers. Or relatedly, they assume their entire audience is comprised of field croppers and fail to use a disclaimer before making sweeping statements. The problem (and it is a problem) is so pervasive that we often hear about specialty crop operators taking training courses intended for field crop applicators… because that’s all that’s available.
So how different can spraying be for a given crop? Surely a droplet is a droplet and the laws of physics don’t care what you grow? This is true, but droplet size, spray volume, distance-to-target, environmental conditions, sprayer type and product formulation combine in complicated ways. The result is that the best advice for one operation can be disastrously wrong for another.
Case in point: Drift is good. If we were persnickety (and we are) we would suggest that moderate drift is the lateral movement of spray with the prevailing wind, and that this helps ground and/or disperse spray in a predictable direction. It’s not a bad thing. But we know that drift quickly becomes bad in high winds and especially when there are sensitive downwind areas. We won’t even talk about dead calm.
However, in controlled environment applications (e.g. greenhouses) operators use very small droplets in high numbers and absolutely rely on air circulation to expose all crop surfaces to the spray. Without drift, most stationary foggers would only have a limited and localized effect. And in airblast operations, a wind that would never deter a field sprayer operator would derail an airblast operator. Wind is much faster with elevation and airblast sprayers use very small droplets that span considerable distances to the tops of tall targets. In their world, droplet size is not the primary means for drift mitigation – it’s air alignment.
The following table is a relative comparison of key factors in spraying. Note how different they are depending on the usage pattern. And if this isn’t diverse enough, recognize that we didn’t include a column for vegetative management (e.g. roadsides, industrial and forestry) or aerial application (which might even be split into piloted and remotely-piloted).
And so where does this leave us? We have two pieces of advice:
If you are looking for information on spraying, take the time to find out who your source is and understand the context of the information. More often than not it will have relevance, but in some cases it could be completely wrong for you.
If you are providing information on spraying, be clear who the information is intended for. While we don’t propose caveats amending every statement, context is always appreciated. A sentence in an article, or a brief interjection during a presentation, might help someone that doesn’t know what they don’t know.
This presentation was delivered virtually for the 2021 OMAFRA Controlled Environment Webinar Series. If you’d like to learn more about strategies for spraying in closed environments, settle in and give a watch. It’s 45 minutes plus questions.
So, a minor error in the presentation. The image of the ascospore was not quite right. This new version (below) is correct. Perspective can get tricky at the micron-scale of resolution.
This is part two of our three-part monster-article on methods for digitizing and processing water sensitive paper. You can read part one here.
Image analysis software
There are many choices of software designed to analyze digitized WSP images (E.g. Optomax, Stainalysis Freeware, DropVision, ImagePro Plus, DropletScan, AgroScan, DepositScan, UTHSCA ImageTool). Some were developed for aerial applicators to evaluate entire swaths and others to focus on single collectors. Some are more user-friendly than others, some cost money, and some are no longer supported. All of them employ algorithms (a set of rules a computer follows when making calculations) that often make image processing decisions. Sometimes these algorithms are pre-set, which can be convenient but may also restrict our analysis.
ImageJ is a free, open-source application developed at the National Institute of Health by Wayne Rasband to adjust and analyze high-resolution images of small structures. There’s a variation called “Fiji” (Fiji Is Just ImageJ) which bundles ImageJ with tools specifically intended for biologists. Happily, they are equally valuable for analyzing WSP. The interface can be intimidating, but only because there are so many functions that we won’t be using. The learning curve is worthwhile because the user has complete control over the analysis.
The ImageJ menu. Version 1.53e.
Three steps to image analysis
No matter the software, the operations used to analyze a digital image tend follow a three-step progression:
Pre-processing: We select an ROI (a Region Of Interest) in the image and perform a few preliminary operations to improve image quality and contrast. In selecting a specific region, we can avoid unwanted flaws like drips or fingerprints as well as crop the image to some standard size for scaling purposes.
Processing / Detection: Point and Morphological Operations are used to refine the image and establish a threshold so we can differentiate between deposits and the unstained background. The ideal outcome sees the original colour image converted to a binary (typically black and white) image.
Measurement: We use ready-made computational routines to quantify some value. Typically, the percent area covered by deposits, but possibly the count and density of those deposits and perhaps even an estimate of the original droplet size.
Let’s explore each of these steps.
1. Pre-processing step
Pre-processing establishes the scale of the image and allows us to isolate the specific region we want to analyze. Perhaps the water sensitive paper was folded during sampling and we want to analyze each half separately. Perhaps we want to avoid obvious imperfections that would interfere with our results. In some cases, pre-processing might also include adjusting the image brightness to improve the contrast between stains and the yellow background.
Flaws and imperfections become obvious when water sensitive paper is examined under magnification. Part of pre-processing is to select a region of interest that represents typical coverage and does not include artifacts that might interfere with the analysis.
2. Processing / Detection step
Processing and detection can take time because of the degree of computation involved. The higher the resolution and the larger the ROI, the longer it will take. Depending on what you want to measure, it might be acceptable to sacrifice some accuracy for speed.
We begin by determining which pixels represent part of a deposit stain and which represent part of the unstained background. We can accomplish this through global point operations called Thresholding and Filtering. If you haven’t already noticed, image analysis includes has a lot of jargon: “global” means the entire image and “point” refers to our focus on individual pixels. Ultimately each pixel is assigned one of two values, reducing the image to a binary (or 1-bit) format.
Once we have a binary image, we explore the shapes of the deposits (which are sometimes referred to as objects) to determine the limitations of what we’re confidently able to measure. Morphological operations are used to refine or modify these shapes in order to smooth jagged edges and identify whether an object is the result of a single deposition or multiple overlapping deposits.
3. Measurement step
Depending on the image analysis software, the user may be limited in what they can measure. The spectrum ranges from a single value (usually percent area covered) to in depth data relating to each object in the image. The latter might appear in a pre-formatted report, or as a CSV (Comma Separated Value) file for further exploration in spreadsheet format.
Thresholding
A thresholding operation sorts all the pixels in an image by some characteristic, and then allows us to set a threshold dividing them into two camps. In our case, we want to divide them into “stained” and “unstained”. The process is almost like taring a scale, where anything above the weight of the container is identified as the weight of the contents.
Thresholding is like taring a scale. Just as the weight of the container is isolated from the total weight of the container and contents, the stain colour is isolated from the background colour.
The HIS thresholding operation
ImageJ’s Colour Threshold operation is only one way to threshold an image, but it serves as a good example. This method uses HIS (Hue, Intensity and Saturation) to separate the deposit stain colours (blue-green) from the background colour (Yellow). As discussed, each pixel is represented by one or more 8-bit values. In this case pixels represent 0 – 255 hues, 0 – 255 intensities and 0 – 255 saturations. That may seem intimidating, but we mostly focus on hue.
When the Colour Threshold operation is selected, ImageJ sorts all the pixel values in the image into a binned histogram (where the Y-axis is the pixel count and the X-Axis is the range of pixel values).
a. Hue
We begin by considering the hue, which is simply another word for colour. An image with no stains would produce a histogram with pixel colours producing a distinct peak in the yellow range. An image with stains would also display peaks in the blue-green range. The user then segments the background from the foreground by manually setting the threshold between these peaks.
(Top) When the threshold overlaps the background yellow hue (set to 30 here), some portion of the background is falsely identified as a stain. (Bottom) When the threshold is adjusted to fall in between background hue and stain hues (set to 36 here), a sharper distinction is made.
The hue thresholding process is less reliable (or can outright fail) when WSP has coverage in excess of 50%. This is because the color of the intermittent unstained areas changes as the distance between stains decreases. Think of it as blue bleeding into the yellow. The result is that the level of contrast between stained and unstained regions is inconsistent, making it difficult to confidently differentiate between “stained pixels” and “unstained pixels”. Similar issues arise when humidity causes the background colour to change but this tends to be more uniform and easier to threshold.
If the hue threshold it is set too low, stains will appear smaller and lose their shapes. If it is set too high, stains will appear larger and the gaps between adjacent, separate stains can disappear. This can have a significant impact on deposit count and distribution assessments, resulting in the loss of thousands of tiny, distinct deposits. Threshold accuracy has less impact on the determination of percent area covered. Research has shown that the use of a single threshold for multiple papers gives an absolute error of +/- 3.5% area covered. This is considered well within the intrinsic variability of spray coverage data.
b. Intensity
Sometimes referred to as “Value”, intensity can be thought of as pixel brightness. No threshold is required here because capturing the entire 256 pixel value range improves the contrast between colours.
c. Saturation
Finally, saturation (a measure of the difference between red, green, and blue levels) is a useful thresholding adjustment when WSP has been exposed to humidity. Humidity does not affect WSP’s ability to resolve stains, but as we mentioned it can cause the background to take on an overall greenish hue. An increase in low end saturation limit can increase the contrast between the stain colour and a less-distinct background colour.
When the HIS thresholds are set, ImageJ converts the pixels values closer to the foreground (stains) to black and those closer to background (yellow) to white. Users can invert this if they wish, or even make the stains red. The important part is that we now have a binary image that makes a clear distinction between the stains and the unstained background. Ideally, thresholding should be performed for each consistent set of samples.
Learn how HIS thresholding is being used to perform weed recognition functions on sprayers in this article.
This 2 cm x 2 cm ROI was HUI thresholded. You can double check your accuracy by having ImageJ “show outlines” which outlines and numbers each distinct object. Zoom in to see if any artifacts remain (or were inadvertently created) and go back to make minor thresholding adjustments if needed.
Pixel filtering
We won’t belabor filtering because it isn’t often required when analyzing WSP. Filtering operations compare pixel values to those of their neighbours and then replace those values with some form of weighted average. This reduces the relative differences between pixel values, smoothing the image and reducing noise (at the cost of lost detail).
