Hi Lucy

As others have mentioned, camera trap temperature readouts are inaccurate, and you have the additional problem that the camera's temperature can rise 10C if the sun shines on it.

I would also agree with the suggestion of getting the moon phase data off the internet.

Do you need to do this for just one project?  And do you use the same camera make/model for every deployment?  Or at least a finite number of camera makes/models?  If the number of camera makes/models you need to worry about is finite, even if it's large, I wouldn't try to solve this for the general case, I would just hard-code the pixel ranges where the temperature/moon information appears in each camera model, so you can crop out the relevant pixels without any fancy processing.  From there it won't be trivial, exactly, but you won't need AI. 

You may need separate pixel ranges for night/day images for each camera; I've seen cameras that capture video with different aspect ratios at night/day (or, more specifically, different aspect ratios for with-flash and no-flash images).  If you need to determine whether an image is grayscale/color (i.e., flash/no-flash), I have a simple heuristic function for this that works pretty well.

Assuming you can manually define the relevant pixel ranges, which should just take a few minutes if it's less than a few dozen camera models, I would extract the first frame of each video to an image, then crop out the temperature/moon pixels.

Once you've cropped out the temperature/moon information, for the temperature, I would recommend using PyTesseract (an OCR library) to read the characters.  For the moon information... I would either have a small library of images for all the possible moon phases for each model, and match new images against those, or maybe - depending on the exact style they use - you could just, e.g., count the total number of white/dark pixels in that cropped moon image, and have a table that maps "percentage of white pixels" to a moon phase.  For all the cameras I've seen with a moon phase icon, this would work fine, and would be less work than a template matching approach.

FYI I recently wrote a function to do datetime extraction from camera trap images (it would work for video frames too), but there I was trying to handle the general case where I couldn't hard-code a pixel range.  That task was both easier and harder than what you're doing here: harder because I was trying to make it work for future, unknown cameras, but easier because datetimes are relatively predictable strings, so you know when you find one, compared to, e.g., moon phase icons.

In fact maybe - as others have suggested - extracting the moon phase from pixels is unnecessary if you can extract datetimes (either from pixels or from metadata, if your metadata is reliable).

camtrapR has a function that does what you want. i have not used it myself but it seems straightforward to use and it can run across directories of images:

https://jniedballa.github.io/camtrapR/reference/OCRdataFields.html

The folks at AB Entheos in Nairobi are also looking at wild life damage insurance https://ab-entheos.co.ke/

Hello to all, new to this group. This is very exciting technology. can it work for ID of individual animals? we are interested in Ai for identifying individual jaguars (spots) and andean Bears (face characteristics). Any recommendation? contact? thanks!

German

Have a look at kobocollect, very easy to use and reliable, can collect data and then send them to the cloud when you have internet.

Other solution:

QField - Efficient field work built for QGIS

SMART Conservation Software - Spatial Monitoring and Reporting Tool (smartconservationtools.org)

EarthRanger: Protecting Wildlife With Real-Time Data

I second Kobo/CyberTracker for tracks, but if you have money for an ArcGIS Online license, Survey123 is great (offline use - but can't record tracks)