Drone Meteorite Searching

Main Survey Data Collection

Drone Image Survey Requirements

• Ground resolution: 2 mm /pixel
• Overlap: 10%
• Minimum sun elevation: X degrees (to avoid long shadows creating too many false positives. Ignore if overcast). Use https://www.sunearthtools.com/dp/tools/pos_sun.php to calculate survey start/end times for your location.
• GPS tagging of the images (Exif)
• Timezone set to UTC

Note:

Survey strategy

When the data is processed, the oldest images are somewhat prioritised. Hence the higher priority/probability areas should be surveyed first.

The simplest way to achieve this is to cut the search area polygon into 5-20 smaller polygons, depending on the size of the survey. Try not to make the polygons too elongated.

Training Data Collection

Capturing quality training data is key to the whole process. Remember: Garbage in, garbage out. Hence take extra care in doing this step well.

True images (meteorites)

Pack list

• Meteorites OR meteoritic shaped rocks (meaning a rock between 2 and 10 cm diameter, with no elongated axis, or sharp edges)
• 2 people: one drone pilot, one pointer.
• black spray paint (matte or shiny).
• a drone
• a cardboard box and some gloves can be useful to avoid making a mess with the painted rocks

Capturing true training data

• Final resolution should be between 1.8 mm/pix and 2.2 mm/ pix resolution, know your drone, know your resolution. If using the M300 with the 50 mm lens, your altitude should be between 18 and 30 m.
• If you have real meteorites with fresh fusion crusts, place a meteorite and point to it at least 3 m away, then pick up the meteorite to re-place it, to then take the next picture
• If you have fake meteorites (i.e. spray-painted rocks) place them in a line near different background objects (limestone rocks, saltbush, grass tuft, hole, lichen colony, etc) at least 3 m apart from the others, then walk next to the line and point to each rock
• The pointer should walk slowly, pointing to the rock nearest to them, the drone flyer should take only one image per rock, and should call out to the pointer if the pointer moved too quickly and the flyer was unable to take an image, otherwise the flyer should confirm ‘good’ once they have taken an image fo the current rock in question
• Between images, rotate the drone by up to 90 degrees to change the angle of shadows in the image.
• Before uploading the data, remove non-useful images (duplicates of the same rock, unrelated random pictures). This clean up step will help limiting the amount of work downstream to label the data, and ultimately keep the training dataset clean.
• When labelling (drawing a box around the rocks) make sure to include the shadow the rock makes, that will help distinguish the rock from holes in the ground or other features that have no shadow. -> TODO move that bit to the webapp documentation.

Do we need to capture true training data at every site? We have built up a large dataset of meteorite and meteorite-looking rocks for training, hence the return on adding more training data tends to diminish. However, this step is a relatively low effort thing to do at each new site visited (assuming you are just doing 10-20 meteorites). Over time it does truly really help increase the variance in the training set, ultimately making model predictions more accurate. So avoid skipping it.

False Images (non meteorites)

• Fly in a perimeter around the fall zone and take images every 10-30 seconds.
• If the search area displays significant variations in ground features (rock type, vegetation), try to capture as much of these different backgrounds as possible.

Can I skip false training data collection at one site? No, the prediction model will perform better if trained on the local background (vegetation, rock types, animals...). So far we have not seen two fall sites that are similar enough to a point where that step could be skipped.

Low-resolution Survey

(optional) Needed to make a background map of the survey area (used as a layer on slippy map on Webapp).

Data Collection

• Ground resolution: 5 cm /pixel
• Overlap: 70%
• If available, use RTK.
• (optional), use Ground Control Points (https://docs.webodm.net/how-to/ground-control-points)

Data Processing

• Upload and process the data on WebODM (cloud version).
• Download the data product: Orthomosaic (large geoTiff file).
• Upload geoTiff to Mapbox: https://studio.mapbox.com/tilesets/
• Wait for processing, note the Tileset ID, make sure it is public, and put that ID somewhere... TBD

WebODM setup

• feature-type: orb
• use-fixed-camera-params: enable
• sfm-no-partial: disable
• skip-3dmodel: enable
• skip-report: enable
• fast-orthophoto: enable
• dsm: disable

Webapp Deployment

See deployment notes: https://github.com/desertfireballnetwork/dfn-meteorite-drone-webapp/tree/main/webapp

GPU machine Deployment

Heavy lifting jobs (Machine Learning training and inference) need a GPU desktop or beefy VM. This machine just needs to be connected to the internet, and will work as a slave wrt the webapp on the VM.

See install notes: https://github.com/desertfireballnetwork/dfn-meteorite-drone-webapp/tree/main/mldaemon

Checklists

Computer things

• Laptop + backup laptop. Linux distro, ~1TB free space, SD card reader. On DFN VPN, with relevant people's public keys authorized so remote debugging can happen.
• Spin-up Webapp VM.
• Start high-performance GPU computer (lab), and/or reserve a g2.xlarge GPU VM on Nectar (https://dashboard.rc.nectar.org.au/project/reservations/). Start the client (see #GPU_machine_Deployment).

Comms

• Starlink unit.
• Starlink ethernet adapter.
• Re-activate Starlink subscription that should be on pause (ask Hadrien).
• WiFi router with WAN port and swappable antenna (SMA).
• high-gain 2.4GHz WiFi antenna.
• hand-held UHF radios (1 pp) + charger.

Power

• Generator.
• Petrol and/or diesel. Usage: ~X litres / 12h
• Battery station + cables (notably XT90 to Anderson plug connector).
• Rhonda roof-rack solar system: ~300W panel + MC4 fuse + MC4 extension leads + MC4 to Anderson adapter (all of this should be mounted permanently on Rhonda and not removed).

Drone

• Drone.
• Drone camera + lenses.
• Drone controller.
• RTK base station.
• Batteries (charge them just before the trip).
• Battery charger.
• SD cards.
• Something to make a landing pad (e.g. tarp).
• stuff for making ground control points (for the low-res survey)