What follows already exists on the thread whose link is below. I open a new thread so that the title is more evocative to facilitate the search for those who are looking for information.
P4 Multispectral controller | DJI FORUM

We realized this by comparing rasters made by ourselves with a P4M and rasters also made by ourselves with the Pika-L hyperspectral camera. I'll give you an example, on the same group of tree, on the same day, same resolution, a few minutes apart, we find an average NDRE=0.68 with the Pika-L, and NDRE=0.31 with the P4M.


You suspect it but I specify it for those who are not yet aware : DJI has never answered. On the other hand, the Pika-L is returned to the US for calibration, and on return the problem persists. So it's definitely a problem coming from the P4M images.

I have prepared something more visual for you below. Both rasters are normalized 0-1. I cut one raster to highlight the inconsistency of the data. Pika-L is on the Left / P4M on the right. Both are (NIR-RE)/(NIR+RE) rasters 840nm and 730nm.


With NDVI the data tends to be identical so that the two rasters merge, the difference is not obvious at all.


The NIR band is the same, so I conclude that the RedEdge is not on 730nm.

I try to replace it with closest hyperspectral band without success. What i can say is that by multiplicate by 2 the NDRE value from P4M its +/- good.
I mean, by walking fast, nobody can see the difference.

Of course by comparing with Pika-L or Sentinel data, you immediatly noticed that something is wrong.

Last Thursday DJI offered us a webinar presentation of the Mavic M3M. Organized by Randal Warnas Executive Vice President of enterpriseuas.com (not a DJI company but a third party). I am attaching the link to the transcript of this webinar at the bottom of this post.
Before the webinar I sent him an email concerning the defect of the RedEdge band observed on the P4M, in order to know if this has been corrected on the M3M or if the same error has been made.

The webinar consisted of a question and answer exchange where Randal addressed his questions to Kyle Miller (DJI-Solution Engineer).

Around 13:30 the RedEdge question was addressed by Randal, thanks to him. Response from the DJI representative:

"I'am not surprised with the phantom 4 multispectral a bit, the one massive upgrade, you gonna see, the P4M had 2MP imagery, these (the M3M) each camera is 5MP. So we are looking at 2 and a half time the GSD so a bit more priority in there. You are also going to notice, we are capturing 4 bands in the Green, Red, RedEdge, Near-infrared, you can see which band are going after. The P4M was +/- 32nm we are thightening data quite a bite becoming much closer to industry standard with Micasense Sequoia things like that. So the accuracy of the band is greater the GSD is also greater."

According to the specification from the DJI site, screenshot below, it is the same characteristics of sensors. Only the NIR has changed. Unfortunately it was the band that worked well on the P4M. We are entitled to expect the same error on this new M3M.
Moreover the P4M is not +/-32 but +/-16 like the M3M.

To err is human, let Kyle have the right to answer us and provide more details if he wishes. I'm going to send him an email now, giving him the link to this thread in order to invite him to come and express himself.

P4M :


M3M:

The email is sent:


Hi,

Thank you for your presentation during the M3M webinar last Thursday. Even if the vision of the multispectral imagery that you bring is a little vague. But anyway, it's still not an exact science.

I'm the person who asked Randal about the faulty P4M RedEdge band. You partially answered the question. First of all thank you because you are the first DJI'employee to express yourself on the subject.

However, the answer is not really the one expected and moreover there are inconsistencies.

I talk about it on the DJI forum in the P4M section, I invite you to come and bring more precision, and to correct certain things if necessary:

Faulty Red-Edge Band | DJI FORUM

The goal is not to destroy the image of DJI company or DJI's multispectral range. We obtain excellent results with the Red and NIR bands of the P4M. In addition, I find it very good that your products make multispectral accessible for half the price of what is done in other brands.

The idea is to find a solution, explanations... Maybe it's us users who misuse the product.

Waiting to see you answer us on the forum.


Regards

i was there, although I came in a bit late and clearly after your question was addressed. I asked the same question as well but it's no surprise they didn't respond as it was a dupe by that time. As to his response to your question during the webinatr, he gave an answer but it was not the answer to the question you asked.
Interested to hear their after-webinar response to this ongoing inquiry...

patiam Posted at 2-27 15:39
i was there, although I came in a bit late and clearly after your question was addressed. I asked the same question as well but it's no surprise they didn't respond as it was a dupe by that time. As to his response to your question during the webinatr, he gave an answer but it was not the answer to the question you asked.
Interested to hear their after-webinar response to this ongoing inquiry...

I finally managed to get a complete dataset from M3M.
It is provided by Pix4D thanks to them:  DJI Mavic 3 Multispectral dataset | Pix4D

As usual we will go to the essentials. I have no sentinel data corresponding to the same date, nor data from another camera taken on the same day.
Nevertheless what seems is that the problem is exactly the same as on the P4M.

In detail :

The data was captured over several agricultural plots. I don't know what kind of culture it is. The field that covers most of the raster is made up of relatively young and sparse plants, so not ideal for comparison. There are two much more vigorous plots in the West, so I isolated one of these to extract the values.



The histograms are quite telling:
The NDVI seems to be correct, it shows healthy cultures, which is consistent with what can be seen on the RGB orthophoto.


Let's go to the NDRE, that's where there is something strange. The values are extremely low.
Take a look at scientific publications to get a third-party opinion, these values should potentially be at least double considering the high NDVI index.


Again I don't have comparative data for this dataset, I'll look it up.
But there is clearly a problem.


-Either the problem comes from me, there could be something wrong in my workflow, which is not impossible.
-Either the problem comes from DJI which does not have the necessary know-how to manufacture multispectral sensors.


Anyway, in the absence of a response from DJI, if you want to do multispectral imaging, avoid the M3M.
At the limit for a farmer who wishes to work only in NDVI, yes why not. Because the NIR and Red bands seem to return consistent results with the use of the sun sensor. In addition, the RGB sensor is of good quality.

Scientists, run away!

One last thing, Pix4D provides images of a reflectance target in the dataset. Don't worry about that, without calibration data, you might as well photograph a sheet of paper, the result will be the same...

The sun sensor does his job very well. Provided it is oriented correctly. If the sun is low, remember that the GNSS antenna of the additional RTK module can generate shadows on the sensor depending on the trajectory of the drone. This was not the case on P4M.

Hello,

is there any news on this very interesting point?

Forester123 Posted at 6-18 21:05
Hello,

is there any news on this very interesting point?

No news from DJI. It would seem that they are unable to say why they made this choice of band. The positive point for them is that by comparing the data with a hyperspectral camera, I find values consistent with what is announced.
For example below it is the totality of the spectrum of the crown of a pine from 400 to 1000nm. In RAW relectance value what DJI returns is twice what Sentinel-2 or a Micasense camera returns.




The problem with RedEdge is that the data range varies extremely. If you take NIR it's already more linear so it's easier to compare even if the bands aren't quite the same.
I have never tried with Parrot Sequoia but given the specs announced, we are not far from the values of the DJI camera.

I don't know why there is so much difference. If anyone knows, he is welcome to explain to us.

Bump. Is DJI ever going to respond to these questions?

patiam Posted at 12-27 18:03
Bump. Is DJI ever going to respond to these questions?

No, apart from the awkward and furtive response during the webinar on M3M, no response.  

I just think that no one at DJI has any real expertise in this area.  They probably bought a cheap sensor off the shelf to put on their drone, without really knowing what it was.  

On GSPRO it produces photos with lots of colors which seem to match the health of the plants -> quality examination OK -> placed on the market -> collect money...

We can't blame them too much, they're a drone manufacturer, they're good at it.  For non-RGB sensors there is still a way to go.

What is scandalous is not having someone who could justify the choice of this sensor.  The opportunity to be contacted by those who designed it would allow us to find out more.

We can blame them. They marketed several products as "Enterprise" level "Professional" tools for scientific applications. They provided sub-standard documentation, no clarity on calibration or validation methodology, and no informed or intelligent response to repeated questions regarding the specifics of the sensor source and real-world capabilities. They required the P4 MS to use a GCS (GS Pro on iOS) that they have abandoned without a replacement. They have failed to provide an RID solution for the platform as well.

DJI does make drones that (mostly) fly well. But if they wish to be taken seriously (and not just take our money) they need to better support their non-hobbyist customers.

We can blame them plenty.