1
$\begingroup$

I have a plant field with potatoes. I have made measurements on ground with a sensor and I have measured soil salinity (Electric Conductivity) and I have min EC and max EC. I have also UAV images of the plant field. I want to map soil salinity values on the ground with the UAV image. I have read papers, found through google searching, and they use Machine Learning algorithms (ex. https://www.sciencedirect.com/science/article/pii/S0273117721007845). Can I avoid ML algorithms and implement it with another rationale/algorithm?

$\endgroup$

2 Answers 2

3
$\begingroup$

Can I avoid ML algorithms and implement it with another rationale/algorithm?

If you can do it at all (with or without ML), it is because there is some signature that you can see from the air with your camera that is distinctive -- i.e. is there some combination of leaf color and texture (or soil color) that indicates salinity and not poor nutrient content or the wrong pH or too dry or too wet or whatever.

Can you look at the ground and tell? (Seriously, I don't know.) Is there some really smart geezer in your area who can? If so, use that skill as a starting point.

If folks are trying machine learning on it, it's because they've tried other techniques and didn't have much luck, or because such techniques are known but require too much processing, or because ML is just so popular now that we'll throw anything into the ML meat grinder and find out what flavor of sausage comes out.

If you can do it by rules-based AI (i.e., good ol' traditional machine vision) it gets back to knowing what signature to look for. It may be that the information just isn't there in a visible-light, RGB image. You need to have something that's different on the salty soil vs. the sweet soil, and that something needs to show up in your sensor. If the salty soil doesn't do something distinctive to the RGB image then you're just out of luck.

I'd dig for older papers. Maybe see if you can exclude "machine learning" from your searches. Maybe see if you can find a professor of plant or soil science and have a chat. If you're in the US, call your state's agricultural extension office -- they may be able to help you directly, or hook you up with said prof. If you're not in the US, see if your country has an equivalent.

I also wouldn't let anything other than money stop me from considering other sensors. I suspect that there's radar frequencies where the reflection is different for conductive vs. low conductivity soil. It may be that overlaying an RGB image with a radar map will have enough information for you.

$\endgroup$
1
$\begingroup$

Tim's answer is excellent, as someone who happens to work in remote sensing I could very much relate to the "how do you currently tell them apart" sentiment. But yes, it is routinely done using things ranging from just color cameras to SAR. If working with plain RGB, the easiest approach is to convert it to the HSL space and devise a metric there, good ol' clusterization ("naïve" ML) approaches still work fine. This will likely not be optimal, but could somewhat work as a cheap option.

One of the most direct indicators of soil salinity is the ~1900-2000nm spectral reflectance (see e.g. this article, freely available on ResearchGate). The main challenge with UAV-based spectral cameras (particularly hyperspectral, as most of them are either scanning or... well, not that great) would be geometric correction. Still, it is routinely and successfully done for simple DEMs, which is the case for most soils. Perhaps there are frame-based SWIR cameras on the market with just that band for this specific problem, I do not know this area as well, unfortunately. But those would be perfect for your problem. At any rate, this is an entire rabbit hole of its own, and I would advise you to find a remote sensing lab and ask if they could help you.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.