After more digging I searched for "software defined" "lora" to find that the Edge is a software defined radio method.  There are a LOT of really specific LoRa experimenters and ideas when you search that combination.

https://www.google.com/search?q=%22software+defined%22+%22lora%22

And I find there is a Hackaday LoRa project already called BUILDING A LORA PHY WITH SDR by Brian Benchoff

https://hackaday.com/2016/11/18/building-a-lora-phy-with-sdr/

When I search Hackaday,io for Lora, there are thousands of mentions.  Another of those active communities that you do not know about until you have reason to look. But they are not working together as a whole.  Not the group just on Hackaday.io, but globally.

site:hackaday,io "lora" gives 4,360 entry points on hackaday,io

Globally there are hundreds of thousands working on facets of this topic.

Richard Collins, Director, The Internet Foundation

Here is the link to the Wikipedia article -- https://en.wikipedia.org/wiki/LoRa

The big red flag is "proprietary spread spectrum modulation"  so better to go broader to Chirp Spread Spectrum at https://en.wikipedia.org/wiki/Chirp_spread_spectrum

The key components are the chirp transmitter and chirp detector.  Those are the packets that get sent.  Rather than simplistic sine waves that are ambiguous, the signals have several parameters to play with.  Random very short spikes should also work.  So you can design your own.

You might want to look more closely at https://en.wikipedia.org/wiki/Time_reversibility where a signal at the receiver is inverted and retransmitted.  This is  a positive feedback method to build and sustain a channel between two distant sites.  This is an old method.  I first read about it 50 years ago. Two receiver-transmitters characterized and strengthen the channel between them, so they could use lower power, and not have to guess.

Also the software defined radio part of global radio networks routinely send signals much farther, by sending pre-arranged signals. Knowing the precise time and pattern is a large part of long distance and low power methods.  https://en.wikipedia.org/wiki/Software-defined_radio

I want to use this for gravitational imaging arrays.  I have been looking at multi-channel global arrays to gather enough data to get unambiguous correlations.  Patterned signals are much easier than naturally occurring signals.  I suppose you are familiar with https://en.wikipedia.org/wiki/Very-long-baseline_interferometry and https://en.wikipedia.org/wiki/Pulsar.  The pulsar timing array can be thought of as a "single frequency array" approach - https://en.wikipedia.org/wiki/Pulsar_timing_array

For gravitational imaging arrays, a key part is three axis detectors, so each can add data about the direction of the source.  High sampling rates constrains the timing of the signals. And FFT and wavelet and chirp detection allows for characterizing the fine spatial and temporal patterns that are intrinsic to the source.  If you think of the pulsar signals, it is the tiny faster and slower variations that are unique to each source.  It is NOT random, just complex.  Enough memory and patience and it can be learned. The AI methods do it by brute force.  It is mainly a process of learning, storing, then predicting  the patterns of the signal source and channel pathways.  

Thanks for point out these low cost tools.  You should be able to adapt them to optical channels without too much difficulty

Richard Collins, Director, The Internet Foundation 

This is very interesting.  I would like to know more about the receiver and geolocation features, and the programming environment, and the costs.
Can the receiver sensitivity be increased? Does it have programmable gain by channel?
If you want people to create projects based on this, give them links and information. It may take time for a community to grow.  First everyone has to have the core information down in an easy-to-understand and apply form.  Don't leave any specialized word or source ambiguous.

Start by putting Long Range Low Power (LoRa LP)  right up front in your title.  Don't assume people have heard about it.  Invite collaboration, don't create minor secrets and mysteries that just get in the way.

Almost all the AI methods derive from basic statistics and correlation and fitting algorithms that are much faster to implement.  Rather than using floating point methods, you should be able to substitute many digital and nonparametric methods.  This is crude, but you can say that "finite automata methods" precedes "Markov chains".

I expect all the functions can by synthesized with software defined radios.  For specific environments the channel could be pre-scanned and then the appropriate hardware fabricated.  Model, optimize, fabricate.

Richard Collins, The Internet Foundation

I am interested in seeing projects built with this module!!!