Big Deac is a little bigger

Well, the title says it all. It came to my attention a couple weeks ago that Intel's Xeon Phi 31S1P coprocessor is on sale for less than $200. The Phi line is out to compete with nVidia Tesla in the HPC market but despite looking similar it has some critical differences. The biggest is that while a Tesla is a high-end graphics card with firmware designed for for HPC, the Phi is a highly-parallel Xeon processor (57 cores at 1.1 GHz, in our case). This means that while it isn't as parallelized as a GPU, it runs an (almost) full x86 architecture. While that should make it easier to program for, in our case it is unimportant. In fact, in our application it should be significantly inferior to a Tesla, or even a top of the line desktop GPU, but it is only $200. So, we got one to see how it goes.

The card, in all of its glory. As you can see, it is quite a behemoth. It is about the same size as our AMD Radeon 7970 and is, actually, a little thinner.

Another difference from the Tesla cards is that the Phi can (currently) only run on Linux. And, to make matters worse, it only natively supports RHEL and SUSE. Being the cheapskate that I am, I had no plans of investing in enterprise Linux. Instead I opted for Ubuntu 14.04. I found this excellent guide on how to get it up and running. I did have some difficulties getting it going but, for the most part, everything went as planned. Following the instructions closely you shouldn't have a problem. My only addition is that when you download the OpenCL drivers, get the ones that say they are for RHEL/SUSE. Even though they say Ubuntu isn't supported (and there is an Ubuntu version which lacks Phi support), the OpenCL 15.1 Runtime for Linux installs perfectly fine on Ubuntu.

The next (ongoing) challenge was cooling. The card is passively cooled and designed to be installed in a server with what amounts to a wind tunnel in it. So, to satisfy the cooling requirements of the card (a whopping 275W TDP, to be exact), I made my own wind tunnel. I put a 75 CFM intake fan in the front of our case and a 150 CFM 92 mm fan on the back with a, erm, custom ducting unit, to draw air through the heatsink of the card.

The exhaust fan. It is LOUD.

The card installed. Not shown is the 120mm fan now installed below the HDD bays to push air toward the intake side of the card's heatsink.

The exhaust fan is very, very loud (who knew pushing 150 CFM through a 92 mm diameter hole would be noisy), but it keeps the card running nice and cool (currently peaking at 52 C to be exact, which is well within spec).

So, with everything in place, it was time to fire up Photoscan and give it a whirl. To my relief the Phi and the 7970 both showed up in the OpenCL menu in Photoscan, so I selected them and got to work. I opened our 3000 image campus dataset, deleted the dense cloud, and set it loose reconstructing the whole thing on medium quality. It is looking like it will take about an hour and 45 minutes (a very respectable time) and isn't even able to fully utilize the Phi (not necessarily a good thing). I think chunking tasks and running them in separate instances of Photoscan will give further performance gains. Once it is done I will go in, disable the Phi, and see how long it takes then. But, for now, all is looking quite good.

Shortly after starting the reconstruction.

Pushing the limits

While I have, for some time now, hinted at our intentions to obtain a more capable aircraft which will provide longer endurance and higher payload capacity, I have kept it mostly in the shadows. Well, it is now time for me to [very excitedly] announce the arrival of the newest and most capable addition to our fleet of aircraft: the Mugin v3 aircraft.

This impressive aircraft will launch us into a whole new class of UAV remote sensing, giving us capabilities far beyond our current stable of electric aircraft. Here are some of the details:

Wingspan:	3m
Material:	Wood structure, fiberglass and film covering
Powerplant:	Saito FG-57TS 57cc, 4-stroke, twin-cylinder, gas engine
Propeller:	Xoar Beechwood 22x10
Flight controller:	3DR Pixhawk
Cruise speed:	120 km/h
Endurance (expected):	~5 hours
Range (radio limited):	~100 km

This aircraft is not widely used, but there are some notable exceptions. CanberraUAV used an earlier version of the aircraft in the 2012 Outback Challenge. There are a handful of others who are discussing using it with APM/Pixhawk or are in the process of adopting it into a program, but we are certainly spring-boarding ourselves to the forefront of this work.

Stay tuned for updates on the build and testing of the aircraft. At the moment I have done the little bit of assembly that is required on the fuselage, gotten the motor and fuel system set up, and installed all the electronics. The next steps are to break in and tune the engine and power up the electronics, configure them, and test them out. The it will be flying time. As I have been building I have been looking around and found that there is an utter dearth of information about this airframe, so I will do my best to document the process in painstaking detail.

Too much excitement

Alright, after long absence I have some updates to make. First of all, to follow up on my last post, the endurance quad got back-burnered but we're going to turn our attention back to it. We made some test flights and determined that the vibration levels were too high for reliable flight. I traced this issue back to the motors, so we're replacing them. We could change bearings and fix the vibration issue but I already wanted to swap the motors for some lower KV ones, so I decided to save the time on the bearing work and just get the right motors for the setup. They should arrive in the next week or so, so I'll update once we have some flight time on the aircraft.

2 Zephyrs heading to Peru soon

Second, I have been in a build frenzy around here, with 3 Zephyrs just getting finished up and another right in the middle of the build. I am also working on a rail launcher that should get the planes in the air more reliably. It looks like we will have some fantastic weather this weekend, so we should be able to get all those new aircraft up and see how they perform. I'll provide more info on that after the weekend.

The Deacon Zeph in its fantastic black and gold livery

A bit of bad news is that we did a test flight on one of the new Zephyrs a week ago and had a crash on takeoff. We had flown it before and it performed really well, so that caught us by surprise. The wind was quite gusty and swirling on the ground that day so we think there might have been a stall issue with a rogue gust on takeoff that put it in the ground. We should be able to determine if that was the issue this weekend. I review the logs and there was a suspect board voltage change, but it didn't happen until after impact. We'll keep an eye on that to see if we actually have a hardware issue.

Finally, on the same day we crashed the Zephyr, we also did our first ever successful fully automatic airplane flights. We had tried auto-takeoff before with a skywalker and it ended badly. We were hoping that new firmware had solved all the issues we had before, so I put together an auto takeoff, circuit, and auto landing mission on Marcus' Pixhawk-equipped Apprentice. We were able to carry out the flights with total success from hand launch and ground roll despite the heavy winds (really too windy for an Apprentice). We found the results to be repeatable, with identical performance on numerous flights.

Well, thats all for now, but hopefully on Monday I will have lots of exciting updates to report.

Endurance Quad

I have been talking with Jonathan Dandois (of Ecosynth and the Smithsonian Tropical Research Institute) about the work each of us has been doing lately, and he told me about a new aircraft they had. They have put together a long endurance quad using a modified MK Okto frame, MK motors, and 15" props and they got 53 minutes of flight time out of it. Hearing about that sort of impressive flight time pushed me over the edge, as I have wanted to put together an endurance quad for a while. It was made even sweeter by the fact that I had all the necessary parts in the hangar to make it a reality. So, 4 hours later, this is what we have:

Here she is, all ready to fly and only lacking props. I used the short arms from the Okto frame and added some strengthening to the frame where the other arms had been removed (they seemed to be structurally important to me).

In the interest of absolute minimum weight and maximum simplicity, I have simply attached a Canon SX260HS to the bottom of the battery cage.

The stack. I'm currently running a FrSky receiver but will probably switch to a Dragonlink after testing. An RFD900 sits on the back and HK F-20A ESCs drive the motors.

In the interest of minimizing weight I have used as few Deans and bullet connectors as possible, hard soldering most things in place. The motors have bullets but if things go well in testing I will probably go back and solder them.

As of right now it is ready for a test flight. I have configured the APM, gotten the transmitter set up, and have all the motors spinning the right way. Now all that is left is to get it airborne. Hopefully that will happen tomorrow. More soon.

The Final Countdown

Well, here it is, the home stretch. I'm currently back in Cusco after finishing up at CICRA and, sadly, retiring our old, trusty DAO. She had been through a lot and the foam at the rear of the fuselage was getting quite weak, which allowed the tail to move vertically quite a bit. I attribute it mainly to 3 weeks of daily rough landings, a 100m plunge into a stand of bamboo due to an in-air motor failure, and a failed takeoff which sent it tumbling over a cliff. So, all in all, it is pretty miraculous that she lasted as long as she did.

But anyway, it is time for the final chapter. On Monday I'll be heading back to Tres Cruces armed only with the hexacopter, my wits, and 2 compañeros from one of the universities here in Cusco. We will set of down the trail which is home to 8 of our tree plots, and we'll set up camp halfway down. We will use that as a base of operations to make daily trips out with the copter to image the plots and, potentially, some of the area surrounding the plots. When we're done the trail will spit us out at the other end, we'll catch a ride down the road a bit to San Pedro, and stay there for a night. The next day we'll head up to 2 more tree plots, image them, and hopefully head back to Cusco the same day.

That is if all goes well...

Needless to say I won't have internet on the trail, so check back in a week or so to see how it all went down.

Back to the grind

Well, as glamorous as seeking out illegal and environmentally destructive mining activity is, we have other missions which are just as important. Today we returned to our fundamental mission of mapping primary forest in order to be able to observe its current state and document change through time. Here is a look at this afternoon's flight path:

The 47.5 km flight lasted just under and hour and averaged 55 km/h flight speed. It covers just over 550 ha with an imaging height of 200-250 m AGL (depending on terrain). This flight will be combined with another, adjacent to it just to the north, to create a high resolution (<10 cm pixel size) mosaic of the majority of the area of the CICRA biological station. In addition to being useful in our own ecological research, this imagery will be available to others at CICRA to better inform them about the area in which they work.

ISR Part 2

We have more exciting news to report from the front line. As some of you may know, CLASlite is a deforestation detection software product developed by a team at the Carnegie Institution for Science at Stanford University. CLASlite uses sub-pixel analysis of Landsat imagery to detect small-scale deforestation such as artisanal gold mining and selective logging. The main limitation of CLASlite is that it is unable to identify the cause of deforestation, only that it is occurring. Thus, we need a system which allows us to classify the mode of deforestation at CLASlite hits and, specifically, to determine if it is natural or anthropogenic. While high-resolution satellite imagery can be used for this purpose, it is cost-prohibitive to obtain on a regular basis. We believe that drones are the perfect tool to address this need by flying to and investigating CLASlite deforestation hits. Today we tested that belief. Here is what we found:

Here you can see one of the mines we located using CLASlite guidance. At the top-center you can see active mining with the open pit, recently felled trees with leaves still on them, and a man operating a water cannon to wash away soil. Directly below that you can see the machine used to remove gold from the sediment and the tailings pile from the processing. At left you can see the pump used to provide the all important water floating atop a raft in a pond formed by an old mine.

It is not hard to fly over the forest here and find mining activity, but the way we found this site is special. We used the most recent CLASlite data, which identified deforestation which took place between June 23rd and August 10th, and used that to target our flights. Using this method we are able to locate new deforestation rapidly and readily identify its source. Here is a visualization of the process:

Red areas indicate deforestation detected by CLASlite in the 6/23/14-8/10/14 period

We then use the CLASlite hits to plan and execute a reconnaissance flight. This is the actual flight path from today.

In addition to the mining shown above, we also found other causes of deforestation. In one case we found agricultural activity and in another we found what appears to be a false CLASlite hit over a stand of young second-growth forest.

Using this method we can augment the impressive power of the CLASlite system by adding the ability to identify the cause of deforestation and provide continual monitoring of that area, if necessary. While our current aircraft has limited range, we intend to acquire larger, longer-range aircraft which can extend this powerful monitoring capability over vast portions of the Madre de Dios region under threat from ever-increasing gold mining activity.