Best Practices to Stitch Crop Imagery
Drone-flown maps are fast becoming a part of farm intelligence. Stitch crop imagery more effectively with these tips!
Jul 5, 2016
Building a mosaic: The practice of creating one large image from many small images, is integral to how many growers, agronomists, and crop consultants currently use drone data. Many geotagged photos are transformed into a single “georeferenced orthomosaic,” corrected for spatial anomalies and perspective to give a proper “top-down” view of the subject at hand. From there, the image can be exported into an appropriate vector or raster dataset format. In the time that we’ve been in the “drone business,” we’ve seen a lot of failed stitches over a multitude of crop types and myriad of different platforms. It’s best to get past any harbored denial early on; because if you stitch crop imagery, some of those fields are gonna fail to come together. It’s the nature of the beast. That said, knowing where the pitfalls are in-process can make your workflows that much more reliable and manage everyone’s expectations when things don’t go well.
There is no Cookbook.
Ag professionals in general are very good at following instructions to the letter. Heck, they gotta be to deal with a variety of hardware, software, and services (not to mention the colors that go with them). However, with stitching crop imagery, one size doesn’t ever fit all. Just like almost every other aspect of farming, important decisions about your aerial survey are going to have to be made in the field, and you’re gonna have to be flexible. In lieu of taking meticulous notes and number-crunching the statistical likelihood of success (ain’t nobody got time for that), your experiences are key.
How to be more successful more often.
Enough of the soapbox talk. What follows are considerations we look at with every mission that have given us more success when we go to stitch crop imagery together:
Stitching software likes to see odd things in your photo pile. Those unique inconsistencies can mean the difference between a good stitch and leaving your pickup with an empty tank of gas at the end of the day with nothing to show for it. Be it cloud-based or desktop, stitching software picks up on those unique things in your photo pile. Stitches tend to fall like a house of cards if variability isn’t dotted across a field (bet you never thought all those red ant hills in your alfalfa could be a good thing, huh?). Is your field bounded by a road? Include it in the flight. Don’t really need the corners around the center pivot? Capture ‘em anyway as that variation could make all difference toward a successful stitch. Truth be told, all those extra short turn-arounds on the edges of a circle don’t really help battery life anyway (that’s a discussion for another day).
There’s evidence out there that perhaps the optimal measure of crop health in aerial photography happens when GSD, or pixel size is equivalent to leaf area. That ain’t where it’s at however for an orthomosaic. Unless you’ve got a good reason to fly low (stand counts, etc.) fly as high as the law allows and get as much area as you can into every shot. The more tie points stitching software can find between overlapping images, the more likely it is you’ll have a useable product. Think about the last time your spouse handed you a camera at a family picnic. If you wanna get more folks in the photo, the easiest way to do that is to back up, right? What’s more, most farm implements work on scales of feet, not inches. Flying at 400 foot above ground level still gives GSDs of around 4 inches or better in many cases; and even if native resolution was still a foot or two per pixel, we’d still have to dumb the data down to make it functional for the likes of analysis software such as SMS, SST or Apex. I would fully expect that as we move forward into the new age of Part 107 flight in the U.S. that many of the waivers that we see will be for altitude exception over crops. I understand tasseling corn stitches way better at 700’ AGL than 400 feet (but you didn’t hear that from me).
Fly dense (but not too dense).
There comes a point of diminishing returns with respect to acquiring too much data. The goal is to maximize flight times. Flying higher also gives us the added bonus of being able to cover more ground [a lot] faster; but short of the regulations changing, what we can do is dial in image overlap. In some mission software, this is one parameter; in others it’s broken down into sidelap (cross-track) and frontlap (along-track). A blog post by way of DroneDeploy does a fantastic job of illustrating the difference in percent overlap, so I’m not gonna cover old ground here; but understanding overlap can mean more efficient flying. In the context of a typical DJI multirotor aircraft equipped with a stock camera, shutter frequency (and in turn frontlap density) ultimately falls back to ground speed. Flying slow really chews through battery on a multicopter; and it goes without saying that where sidelap is concerned, less rows means less flight time.
Take for example, two flights on the same field: Front and sidelap both at 80% gives us a successful stich to be sure, but tweaking the side and frontlap to 65% and 75% respectively gives us almost four and a half additional minutes of flight time that could be better put to use elsewhere.
You should also expect optimal overlap to vary on the same field as a season progresses and crops change. More homogenous areas need more quite a bit more overlap to stitch (if they do at all); and what worked for wheat in the spring probably won’t work for milo in the fall.
Stitching software has a love/hate relationship with change. Point cloud works because of difference, but when those differences are temporal and not spatial, “GIGO” rears its ugly head. Stitching software can get quite confused if the similarity ain’t there between images that should be; and in our experience unexpected temporal change is the the number-one cause of failed crop image stitch when nothing seems wrong with a flight. Temporal change offenders include:
- A shift in wind direction mid-flight that blows crop canopy in another direction.
- Center pivot in motion.
- Vehicles or livestock on the move.
- And, the no. 1 bane of aerial image mosaics: Slowly moving, dense, spotty cloud cover.
It’s probably your fault.
The longer a drone is in the air, the larger the dataset. Larger datasets are inherently more difficult to process, so it goes to figure that those big acreage flights open yourself up to error. It may take more time, but smaller acreage missions may make more sense to cordon-off problem areas so that the rest of your data isn’t “poisoned” by a handful of bad images.
I can’t tell you how many times I’ve seen folks blaming software for their own mistakes. “I did it the exact same way last time” is a common refrain; but is that really the case? Look at your workflow. Look at the land. Look at your environment and mission variables. Chances are [99.999999%] that at least one thing is different from last time you flew. Recognize what changed. Take notes if you have to; and over the course of your seasonal flights, you should start to see success more often.
..and hey – If you have any can’t-miss tips to stitch crop imagery successfully, we’d love to hear about it in this article discussion on the forums!