Greetings, I am an MSc student in South Africa who has been using the Altum sensor mounted abroad the DJI Matrice 300 drone. We have collected numerous images over temporal scales along with field data points to run a random forest regression. We mosaicked and conducted radiometric correction of the Altum images in Pix4Dfields and are using the orthomosaic image for analysis. The process we use in Pix4Dfields is:
Select drone image folder, select the weather type and leave the GSD off and leave the total megapixels on the default (50 megapixels).Once processing is completed we export the orthomosaic image and surface model and use it for analysis in ArcGIS and QGIS.
However, when we have ran our model in R-statistical software (Rattle) using the random forest regression for each of our variables the accuracies and R squared values are below 0.1. We have been trying to better our accuracies since November 2020 and have had no success. This is why we think the issue may be in our processing technique in these initial stages.
Any insight into bettering our accuracies or any processing tips and techniques, or any flaws you have picked up in our methodology will be helpful. Any advice will be greatly appreciated!
that sounds like a great research project your are working on! To understand wehter the data, the processing or the analyses itself is the culprit for your unsatisfying results we need more information.
Whats the research question?
What’s your dependent and what’s your independent variable, so what do you want to predict with what ?
How do you generate your variables from the orthomosaiks in GIS ?
Did you try OLS Regression to see if there are strong trends in your data ?
Also, depending on your goal, I just wanted to let you know that you can change the GSD inside Pix4Dfields https://support.pix4d.com/hc/en-us/articles/360035741052. Julius has been working on research as well. He will be able to give you further details.
We are assessing maize crop health and water-stress in smallholder farms using proxies of chlorophyll content and canopy temperature.
We are wanting the random forest model to predict chlorophyll content and temperature. We have done this by creating a spreadsheet of our field data measurements, which include the following columns: co-ordinates of maize crops we are sampling (longitude and latitude), chlorophyll SPAD reading measurements, and canopy temperature measurements. We insert this spreadsheet into the model to see if the random forest can predict similarly. However, the accuracies are low. We tried utilised other regression models such as SGB and SVM and the accuracies remain just as low.
Our variables are generated from our field measurements. We collect field measurements for about 60 points in a field (so for 60 different maize crops we take a GPS reading, measure chlorophyll and temperature). We then create a point map of the field data and overlay it with the orthomosaic and then extract the DN/reflectance values for the various points. Then we add those to the spreadsheet that we use in rattle.
No we have not, we have only been using the ones available on R-statistical (rattle) which include random forest, svm, linear, neural net etc.
Thanks. Yes, we have processed several orthomosaics at different GSD’s and tried that as well. The highest accuracies we achieve come from an orthomosaic of 30 cm GSD, but are still low at R-squared of 0.1. Yet, we would have thought the high GSD images e.g. 7cm would produce higher accuracies.
From my process mentioned above (in the first message) on how we have been processing the images in Pix4Dfields, is this more or less correct?
What vegetation indices do you use ? It would be interesting to see a scatterplot of VI vs. Chlorophyll and Temperature. We would expect a trend there.
With high resolution also comes a lot of noise. I would recommend to a) buffer around the 60 Points ~ 50cm in QGIS and then extract the average Index/reflectance Value for the buffered zone or b) Aggregate multiple points to zones where you take the averages from.
But I would look at the scatterplots first to see the relation ships in your data.
I did a quick correlation check. As you can see temperature and chlorophyll from your ground sampling do not have strong relations to any of your remote sensing data.
So this could be a problem of your data: 1) Collection 2) Processing Quality 3) Noise from extraction
We conducted field measurements with a Trimble Juno 3 series. We also georeferenced the orthomosaic using points from Google Earth Pro. Please see attached image. Georeference 5cm.tif
I compared your geotiff tiff to the google basemap and it looks like that your scene is still shifted. But I am also not sure how accurate the google basemap is.
Your Trimble Juno 3 has just an accuracy of 1 to 5m , but only with postprocessing. So your data points also have some jitter. You can see in the overlay, that they are also outside of the field. And if you do a point extraction you will sometimes extract soil values and not plant values.
So I would recommend to buffer around the points, only use points which overlap with the maize crop and also mask out soil values and plant shadows.
Thanks for your insightful input Julius! My colleagues and I will definitely test out your recommendations. I will get back to you as soon as we get some new results.
the micasense altum is supported in pix4dfields for processing and you are also able to calculate the NDWI index. You can check out a test dataset at the micasense page: Altum-PT - Drone Sensors | AgEagle Aerial Systems Inc.
Hello, there are some papers that indicate that you can generate this index with this formula:
NDW I = (G − NIR)/ (G + NIR) to monitor water content in leaves
Otherwise, you need the SWIR band but you do not get that with Altum:
NDWI = (NIR – SWIR) / (NIR + SWIR)
I recommend having a look at some research papers about the first formula I provided in case you can use it.
Hope this helps.
These cookies are necessary for the website to function and cannot be switched off in our systems.
They are usually only set in response to actions made by you which amount to a request for services, such as setting your privacy preferences,
logging in, or filling in forms. These cookies do not store any personally identifiable information.
These cookies allow us to count visits and traffic sources so we can measure and improve the performance of our site.
They help us to know which pages are the most and least popular and see how visitors move around the site.
All information these cookies collect is aggregated and therefore anonymous.
If you do not allow these cookies we will not know when you have visited our site, and will not be able to monitor its performance.
These cookies may be set through our site by our advertising partner (Google).
They may be used by Google to build a profile of your interests and show you relevant adverts on other sites.
They do not directly store personal information but are based on uniquely identifying your browser and internet device.
If you do not allow these cookies, you will experience less targeted advertising.
Georeference 5cm.tif
