Reflectance/Radiometric Calibration on MicaSense Sequoia

[selimerguden]

Hey Mark,
The calibration targets and the albedo values are being used when you enable the necessary calibration steps. For some reason you still need to normalize the data by dividing the band by the max value. That is the workflow that I was told by Alexey.

Furthermore if you want to make sure that your calibration target and the albedo values are being included, I would process the same dataset once including the calibration images and the albedo values, once excluding the calibration images and the albedo values. I have tested it handful of times and there was a significant difference between the reflectance values. I am more familiar with Pix4D multi-spectral processing, I have processed over 1500 projects but I have done enough multi-spectral projects with Agisoft Photoscan as well.

[selimerguden]

Hey Gustavo,
I am not sure if 65000 value is actually the DN number, because that value appears in the Raster Calculator histogram once the data is being processed. However the data still needs to be normalized regardless. Once it is being normalized the reflectance values makes sense since they are between 0 and 1 for the Green, Red, RedEdge and NIR bands and between -1 and 1 for the NDVI band.

I have also processed many of the same multi-spectral datasets in Pix4D and compared the reflectance values to Agisoft Photoscan reflectance values. Pix4D and Agisoft Reflectance values are very comparable, especially after the 1.4 update. There was however a very slight shift between Agisoft Photoscan and Pix4D reflectance values.

Like I mentioned earlier this was the workflow that I was told by Alexey, dividing the max value by the band, so something like B1/65500. I have processed about 200 multi-spectral projects using this technique and it has been working so far and I have processed maybe 30-40 of those same datasets in Pix4D and compared the reflectance values. So I can confirm that the workflow actually makes sense at least from my own experience.   

[gtalckmin]

Hello Selim,

My contribution to the understanding of the mechanics of MicaSense sensor/workflow would be:
- Sequoia TIF files have a bit depth of 16 bits. Thus, 2^16 = 65536. I would assume that the 65000 is the max value found in your imagery data set ( the white reference, possibly?); given that the sequoia optimizes the sensor to use the whole range of bit depth.
- You would not be far off from the meaningful values when normalizing/dividing by the maximum value of each band; after all, the white reference should be 1 or the max DN value.
- You would have a problem if the max DN value would be equal to MicaSense's white reference (thus, setting it to 1), as its reflectance value is actually much lower than 1 (and each panel has a specific value, I reckon that is one of the reasons why there is a QR code).
- Another point is that, without using other exif metadata, one would be not subtract/include dark current aka black level. Thus, the coefficient/slope would be different (possibly why you found different values of reflectance between different software).
- There is also vignetting effects,which are not being accounted for.

I believe that you are familiar with, but personally I found this post to be very helpful: https://support.micasense.com/hc/en-us/articles/115000351194-RedEdge-Camera-Radiometric-Calibration-Model .

To me, is not clear whether PhotoScan has incorporated this calibration as described in the link above. If not, I would argue that it should.

It may be worthwhile to incorporate some of these references in your workflow - if you think is worthwhile. After all, reflectance is a physical property and, therefore, should not vary based on which software one is using.

Cheers,
Gustavo

[selimerguden]

Hey Gustavo,

That max value varies between bands. I have seen it anywhere between 35000 to 65000.  And you are definitely right each calibration target has its own specific albedo value, which I believe is linked to the QR code. The reflectance value difference between softwares is very compareable though. That being said there is also that black box within each software, that I am not sure what is happening in their end. Moreover thanks for sharing the link, I am actually familiar with it.
Selim

[mdc]

Hi Gustavo,

Sorry about the late reply. Yes,  I do have the reflectance values for the Airinov target:

Green 17.1%
Red 21.5%
Red Edge 26.6%
NIR 36.9%

Mark

[mdc]

Hi Selimerguden,

I've tried processing the images as you suggested, but I still get get some odd results.

1. If I process the images without applying any reflectance corrections then the data values for each band in my site range from:

Green: 5875 to 32348
Red: 3408 to 43059
Red Edge: 6214 to 55145
NIR: 5765 to 56416

Which all seem pretty reasonable.

2. Processing using the calibration values only (with just the 'use reflectance panel' option ticked) gives the ranges:

Green: 61731 to 65586
Red: 63224 to 65573
Red Edge: 65516 to 65536
NIR: 65515 to 65536

Which surely cannot be correct?

3. Processing using the calibration values and the sun sensor (both the 'use reflectance panel' and 'use sun sensor' options ticked) gives the ranges:

Green: 1197 to 9005
Red: 884 to 12678
Red Edge: 2190 to 29271
NIR: 2359 to 36428

Although these seem a bit more realistic in terms of the ranges, the values themselves seem odd.

It seems to me as though something is fundamentally wrong here. Any thoughts?

Mark

[treeperson]

Hi everyone,

I am having a similar problem. I'm also using the Airinov calibration target. When I enter the reflectance values and click calibrate, all my photos turn solid white. I have tried changing the brightness, but they just turn grey. When I process the photos, I get very odd results. I get similar ranges as mdc. The orthomosiac comes out very strange as well. Screenshots of dense and orthomosaic, with and without calibration are attached.

 Has anyone figured out how to solve this problem?

[Alexey Pasumansky]

Hello all,

As reported in the similar thread, there's the issue in that latest PhotoScan Pro version (1.4.1) related to the reflectance calibration procedure.

The issue will be fixed in the next version update, but it can be quite easily workarounded in the actual version:
variant 1 - before running reflectance calibration operation, set on the flag "normalize band sensitivity" for each band of the multispectral camera in the Camera Calibration dialog,
variant 2 - run reflectance calibration two times in a row without resetting the calibration results.

[william]

Just thought I would chime in here: to the people who recommend dividing each band by the max value, I think there's an important distinction.

1. Dividing each band by the overall max value seems like it will work.

2. Dividing each band by the band max value will completely mess up your colours. It will not preserve the ratio of colour intensity between bands. This certainly won't work.

[selimerguden]

Hey William,

I have processed about 200 multi-spectral projects using dividing the max value of each band to the band and it has been working so far and I have processed maybe 30-40 of those same datasets in Pix4D and compared the reflectance values and they are very comparable, visually and values as well. There is a very minor shift between the Pix4D and the Agisoft Photoscan values, something around 0.05.  Dividing each band to the max value was suggested by Alexey and I believe it is because it is a floating dataset. For the Slantrange for example it is not a floating dataset and it was 255 bit so every band needed to be divided by 255 in order to normalize it, which I was told by Slantrange technical support guy.

You have suggested that dividing it by the max value for each band wouldn't work, have you actually tested it? I am trying to understand the full picture.

Moreover I have looked at the histograms prior to normalization for each band both in QGIS and Agisoft Photoscan and after the normalization for each band both in QGIS and Agisoft Photoscan and ratio stays the the same for each band. I haven't  looked at the ratio between bands however if the ratio is staying the same for each band then in theory it should stay the same between the bands as well.

[william]

A simple example:

Suppose you want to compute NDVI (nir - red) / (nir + red) for a particular pixel P.

Here is some fake data:

colour,min,max,P
nir,4000,50000,10000
red,2000,10000,2000

before normalizing, NDVI = (10000 - 2000) / (10000 + 2000) = 0.67

if we normalize by the overall max (50000), we are left with
nir_P = 10000 / 50000 = 0.2
red_P = 2000 / 50000 = 0.04
NDVI = (0.2 - 0.04) / (0.2 + 0.04) = 0.67

if we normalize per band, we are left with
nir_P = 10000 / 50000 = 0.2
red_P = 2000 / 10000 = 0.2
NDVI = (0.2 - 0.2) / (0.2 + 0.2) = 0

The correct value for NDVI is 0.67, but per-band normalization completely ruins it!

[selimerguden]

Hey William,

 I don't perform normalization on NDVI since it has been normalized within the formula.

I would perform something like:

Green/max
NIR/max
Red/max
RedEdge/max

for each band individually

Moreover I am talking about a range so it wouldn't be a single number.

An example could be something:

NIR  (min)0.30 - 0.67(max)

I never ended up with a zero value. Unless of course they both have same min and same max values for NIR and Red then you could end up with a zero for the NDVI range, which hasn't happened to me though.

[william]

That's all right, as long as you understand that after doing per-band normalization, you can certainly only analyze each band independently. If you combine the normalized bands you will find the colours are wrong.

[jsal490]

Hello All, It was (and still is) unclear to me how the different combinations of the check-boxes in the 'Calibrate Reflectance' dialog influence NDVI results of the resulting orthos. Therefore, I conducted a comparison of NDVI values from 3 separate projects for the following combinations of the check-boxes:

1) 'Use Sun Sensor' Checked
2) 'Use Reflectance Panels' Checked
3) 'Use Sun Sensor' and Use Reflectance Panels' both checked.


As noted by other users my images and resulting orthos were very washed out when I calibrated using options 2 and 3. The NDVI values for these were similar, but nonsensical (basically shadowed areas were the only pixels that contained positive NDVI values). I've attached my results displayed as a classification of NDVI with a range of -1.0-0.35 for dead vegetation (red) and 0.35-1.0 for live vegetation (green). Also included is the average NDVI value for each project. I used the same calibration panel images and photoscan workflow settings to process the three projects. I'm using a reflectance panel from MicaSense, a CSV containing reflectance values provided from MicaSense, and a Parrot Sequoia.

Is it recommended to use both check-boxes when you have calibration panel imagery collected for your project both pre and post-flight?

What do these settings actually do to the exif data?

How can you tell if your calibration was successful (right now I just notice the washed out imagery and the sensitivity values changing in the 'camera calibration' dialog)? I notice it normally takes 1-2 seconds for photoscan to perform the calibration.

Thanks

[gtalckmin]

Hello everybody,

I am still having problems with the radiometric correction of the MicaSense Sequoia.

- Firstly, the software never identifies the AirInov grey target;
- The workflow for masking of grey targets and posterior reflectance factor setup is not clear ;
- Sometime I get all black or all white imagery.
- The output range (0 to 1, 8  bits or 10 bits) should be an option on the software.

Is there a tutorial by Agisoft showcasing how can this be done? I reckon a number of users are facing the same problem.

--- SORTED/SOLVED:
Hello everyone,

Quick documentation as these problems seem to apply to a number of users:
- Sequoia creates 3 sets of images (I assume different exposure intervals), allowing for a gradient of DNs.
- Sometimes one of these images is  under/overexposed and Photoscan cannot read the QR Code in all 3 images.
- One should mask (select+right click+add to selection) the area  of the reference panel on all channels. For such, one should  click the chunk + "Set Primary Channel"  + "Green/Red/RedEdge/NIR".
- Adjust band values according to the reflectance of your panels (Micasense are usually around 0.6 and AirInov have the values printed on the panel).
- Reflectance correction should be executed prior to the workflow.
- No problems were found when also checking the "use sun sensor" option.

By doing such, problems such as complete dark or complete white imagery was solved.

Kind regards