Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Listening for illegal logging chainsaws using TensorFlow

blog.google

Post reply on HN
Print view View original

101–110 of 128 posts

Re: Listening for illegal logging chainsaws using TensorFlow

#101

Post by walrus01 »

aynsof wrote: post #30

Someone more familiar with satellites can correct me, but from my understanding you should be able to detect deforestation from real-time satellite data. You alert on changes in the level of greenery in the images. The indicator might trail behind detecting chainsaw noises, but it wouldn't require the mass distribution of detector devices.

one of the problems is that true near-realtime satellite imagery (90 degree angle direct overhead, visual spectrum or IR spectrum to detect tree cover) is quite costly, in terms of $$$ per square km, if you want images that were acquired recently. The older the data is, the cheaper it gets. This could be a huge recurring cost to buy new satellite imagery.

If you want almost-true-realtime satellite imagery for land-use law enforcement purposes, you're looking at a project that required the financial resources of a nation-stage IMINT apparatus (US NRO) for example, with dozens of polar-orbit, LEO satellites. The best you're going to get from commercial sources is a couple of days old, up to a few weeks old.

Re: Listening for illegal logging chainsaws using TensorFlow

#102

Post by notahacker »

XR0CSWV3h3kZWg wrote: post #91
zapfranklin wrote: post #58

Earlier quoted context omitted.

I assume he is talking about Planet Labs ( https://www.planet.com/ ) - a very cool service.

Do you happen to know what the pricing looks like?

Representative pricing for multiple sources including Planet's RapidEye: http://www.landinfo.com/satellite-imagery-pricing.html

Purchasing a before and after image of a specific logging site isn't cheap once you've factored in minimum order sizes, but it isn't unaffordable. Purchasing images to look at deforestation on a national scale (if you're looking at individual trees, not woods; open data will give you the latter) is millions of dollars per annum, but worked very well where tried to prevent illegal vegetation clearing in Australia.

Re: Listening for illegal logging chainsaws using TensorFlow

#103

Post by RainforestCx »

walrus01 wrote: post #100

> We’ve hidden modified smartphones powered with solar panels—called “Guardian” devices Question, how is the electrical PV power budget vs load budget calculated for these? Having implemented PV projects ranging in size from 300W to 30kW, I'm very interested to know what their predicted Wh per day generated from the solar cell is vs. what the load will consume (what's the actual load of the cellphone, how often does…

Great question, and one that’s of particular importance in our design. Many IoT-type systems make use of on-device data analysis in order to minimize the most energy intensive process: data transfer. Our hardware focuses instead on getting full audio streams up into the cloud in near-real time, continuously, 24 hrs a day. This requires an exceptional amount of energy, as well as the ability to pool energy to continue operating at night and during poor weather. That said, we’ve really overcompensated on energy capture and storage—we operate under the assumption that we should be able to support 1.5 watts of continuous power draw 24hrs a day (normally it’s closer to 50% of this). But that means we try to be able to generate 35-40 watt-hours of solar energy for each day (on average). This is further complicated by the shady/infrequently lit conditions of the tree canopy. So, when direct light shows up (sometimes for moments or minutes only) we want to be able today capture as much as possible (high charging amperage and ample solar panels). Our batteries (LiFePO4) must be able to support this, and not decline over 1-2 years of use being fully discharged and recharged each day). We also designed a type of solar array that is specifically intended for the tree canopies. If this is interesting I’d be more than happy to elaborate more!

Re: Listening for illegal logging chainsaws using TensorFlow

#104

Post by walrus01 »

RainforestCx wrote: post #103
walrus01 wrote: post #100

> We’ve hidden modified smartphones powered with solar panels—called “Guardian” devices Question, how is the electrical PV power budget vs load budget calculated for these? Having implemented PV projects ranging in size from 300W to 30kW, I'm very interested to know what their predicted Wh per day generated from the solar cell is vs. what the load will consume (what's the actual load of the cellphone, how often does…

Great question, and one that’s of particular importance in our design. Many IoT-type systems make use of on-device data analysis in order to minimize the most energy intensive process: data transfer. Our hardware focuses instead on getting full audio streams up into the cloud in near-real time, continuously, 24 hrs a day. This requires an exceptional amount of energy, as well as the ability to pool energy to continue…

That is a really good answer, and thanks for taking the time to write it up. The difference in instantaneous wattage produced by good quality 156mm monocrystalline Si cells is huge in direct sunlight vs shade. You can take a modern 60-cell, 300W rated panel (STC rating) and hold a piece of cardboard a meter away from it, to temporarily shade it, and see a huge difference. What size of PV panel are you using for each node, and how many/what type of cells?

I'm assuming that for cost reasons and economies of scale that you're using some sort of standard off the shelf PV module that can be acquired in bulk, at low cost, such as a 156mm poly x 54-cell module. Or maybe something a bit smaller than that.

1.5W load x 24 hours x 31 days = 1116 Wh per month

For a typical node what is your calculated PV production in Wh per month? Taking the 40Wh per day figure mentioned above, that's 40 * 31 = 1240Wh per month production goal from the PV. Will of course vary with mounting location and shading, but I'm sure you have a floor figure that you don't want to go below, or batteries will reach a LVD threshold overnight and everything will shut off.

I'm also making an assumption that per-node cost is somewhat of a concern in general. For very small PV setups a decent charge controller (that can handle up to the max of one medium sized panel) is not very expensive. Are you using an off-the-shelf $65-75 chinese small MPPT PV charge controller with voltage levels set up to match your LiFePo4 battery setup, or something you've designed yourselves? something like this? https://www.amazon.com/DROK-Controller-Converter-Regulator-C...

Re: Listening for illegal logging chainsaws using TensorFlow

#105

Post by rashkov »

pbhjpbhj wrote: post #95
ct0 wrote: post #81

Earlier quoted context omitted.

sign language will be everyone's primary language. Better yet a combination of sign and spoken.

Eyebrow semaphore? Hand-to-hand Morse Code?

I’ve been thinking a bit about wearable computing and pervasive encrypted communication. I think personal encryption could become more necessary in an age of deep learning enabled fake video and audio. Wearable computing could take readings from your body and environment as a sort of timestamp and fingerprint. It could also enable an easy way to type messages to others privately.

Re: Listening for illegal logging chainsaws using TensorFlow

#107

Post by John_KZ »

LifeLiverTransp wrote: post #90

Next step, mine the rainforrest with sound detecting landmines? Technical solutions to social problems - any examples where this worked?

Exactly. Not to mention that some other guy in this thread said that in Brazil now all forestry equipment is sold with GPS tracking, so there's literally zero need for such a device regardless.

If they want people to stop chopping down forests they need to provide them with a differnt, viable source of income. If I was living in deep poverty and there was a forest near me I'd chop it down, no question about it.

Re: Listening for illegal logging chainsaws using TensorFlow

#108

Post by telchar »

pbhjpbhj wrote: post #95
ct0 wrote: post #81

Earlier quoted context omitted.

sign language will be everyone's primary language. Better yet a combination of sign and spoken.

Eyebrow semaphore? Hand-to-hand Morse Code?

Atreides battle-language, better yet. The signal is encoded in subtle gestures to generate adversarial examples for the language detector.

Re: Listening for illegal logging chainsaws using TensorFlow

#109

Post by woweeeee »

yjftsjthsd-h wrote: post #92
woweeeee wrote: post #14

Logging Man, I must have computers on the brain. This isn't about noisy server logs at all.

...but now that you point it out, I would totally try AI log analytics.

Figure splunk and sumologic are weeks away from product releases, I’m sure.

Either that or it’s already one of their consultant up-sells, for professional service add-ons, before they GA something standard.

Re: Listening for illegal logging chainsaws using TensorFlow

#110

Post by Benofiniquity »

emaginniss wrote: post #65
osipov wrote: post #13

Earlier quoted context omitted.

lyrebirds are native to Australia

The swallow may fly south with the sun or the house martin or the plumber may seek warmer climes in winter yet these are not strangers to our land.

Pull the other one!
Post reply on HN