I was wondering if anyone could share their experiences with estimating populations on photos they've taken with UAVs. Software or particular methods that you have found to work. I currently use a Y6 tri-copter from Event 38 with a GoPro Camera as well as a Canon SX 260 converted to near-infrared. I've used Agisoft Photoscan with great success in photo-stitching and dronemapper. I've dabbled with Fiji/image j, and Microsoft ICE but I've found they seem to have more of a learning curve for me.
Any ideas if population estimates work well in any of these programs or in others that I have not mentioned?
Hi Morgan. I have some internal tools I have made for counting plants from high resolution imagery.
If you can resolve the individual plants in the imagery, then we can do a fairly accurate whole population count, with the main sources of error being any completely overlapped plants and errors from the mosaic. If the individual plants are not resolvable, then we may be able to do some straightforward math based on density and vegetation area in order to get an estimate.
If you have some imagery that you'd like to share, I'd be happy to run my software against it and see what kind of results we can get.
I should be able to find some imagery to take a look at. What file format do you prefer? And does the flight altitude matter, the crop, or whether it is a stitched mosaic or an individual shot?
Cool. Sounds great!
Any format is fine, though GeoTiff is preferable if you have the georeferencing information.
Uncompresssed (raw/bitmaps) are better too, even at a 2cm/pixel GSD the compression will often start to blur the fine details that we might want to delineate the individual plants.
For altitude, the lower the better, but its all tradeoffs really. The lower you are the more likely you can resolve the individual plants, though the data size starts getting unwieldy. I have written my software to handle multi-gigabyte mosaics however, so its not too big an issue. It just takes longer to run and longer to transmit the data.
What kind of crops do you have? My prototype was build around some imagery of potato fields shortly after the plants emerged. The post-emergence time is good because the plants have generally not completely overlapped and formed an indistinguishable canopy. A dense mature corn or soy field may be a bit more challenging (though maybe not impossible).
For starters, it would be easiest however to just run against an individual shot. That will be sufficient to see what we are up against and get an idea of how well the algorithm works for that application.
If you want to upload some images to Dropbox or Google Drive and send me a link via direct message that would be fine. Or email me tcg [at] precisionsilver.com
Looking forward to it. Thanks
I was able to get some good results from one of your corn images.
In this one, if you zoom in, you can make out the individual corn plants.
With this level of detail, we can do a pretty good job at automatically identifying and counting all of the plants. It is a challenging problem, however, and if the plants have formed more of a canopy it will start being much more difficult to tell them apart. As it is there are still a few tricky cases that are a matter of interpretation. I estimate we are correct to within a few percent, but this will require a controlled collection with ground truth to verify. Within this small sub-image it identifies 83 plants, my hand count from the same area was 85.
The full image from the GoPro is actually a bit too wide-angle to be useable over the whole image, the perspective changes from overhead to side-looking as you get to the edges of the image.
I would recommend you set the GoPro to the narrowest Field of View setting for your future collections. This will help with the mosaic as well because it is easier to match the images if they are all from the same perspective (ie looking more-or-less straight down). Also be sure to do everything you can to reduce vibrations and keep the lens clean, there are some other smearing artifacts that are affecting the image quality as well, and every little bit of detail helps in the tough cases like this.
Running this process on a larger section of the image, where the perspective is still mostly straight down, it identifies 1,132 plants. Of course I didn't try a hand count to verify this =)
I can identify a few errors where it adds a plant where I would not have, or misses one I would have counted. In all however, I think the error rate is low and relatively unbiased, so the number is probably close. (If anyone wants to work together for a controlled experiment, let me know, I'm happy to talk!)
So, what's next? I can run these algorithms against a full mosaic if you have one and get a population count for an entire field (so long as it has the same level of detail).
What do you all think?
Taylor: very interesing work, I was thinking how to count plants with images, your program is commercialy available?
Thank you! This software is not commercially available yet, but I am working for it to be available soon.
I am happy to run a few images here or there for you or anyone who would like, this helps me to develop the software, but also if people find it useful I like to help out.
If you have a lot of imagery I could also do this as a fairly low cost service as well. Feel free to PM or email me (tcg at precisionsilver.com) and we can discuss.
If you need some images I can send to you, I have sunflower, corn, soybean, wheat and potato if you need something in special
That would be great! I'll take one of each. Feel free to send me a link on Dropbox or Google Drive. Or if you need it I can send an account to upload to.
I have gotten good results with corn and potato plants so far, but the soybeans I have seen I have not been able to separate the individual plants (I think it will have to be pretty early after emergence and very high-spatial-resolution/low-altitude images). Sunflowers and wheat sounds great too. Thanks
please send me an account, I upload different situations and you choose what is usefull. I'm going to try to obtain soybean photos as you need to try
This is great. Can you create a shapefile with points for each plant detected?
Hi Eric. Sure, that can be done. It requires having geo-referencing for the image/mosaic, so this usually means a GeoTiff file as opposed to just a .jpg. And if you want the coordinates to be really accurate you need to do some extra work in image acquisition/mosaic-ing.
These look great! Very interesting work. I don't think I have a full mosaic that is at quite that level of detail right now. Next season I will be sure to check out my Field of View setting and check how vibration is to get a better image. Hopefully I can pin down the best flight methods to get useable photos for this application, and I will let you know when I do. I'd be very interested to learn if/when your program becomes commercially available. I will send some more images as they come, and as I find them.
Thanks again for taking a look I really appreciate it!