I don’t remember what I was trying to figure out. But I do recall getting the name of an expert in finding human remains. Somebody who’d gone to places like Chile, to look for the thousands of people that the dictatorship had murdered.
I called him, I’m sure, to ask a question about human remains detection dogs — HRD dogs, in the parlance.
I don’t remember my exact question, but I do remember he didn’t really answer it: instead, he said, “In my experience, dogs are best at detecting the urine of other dogs.”
Not very many responses you could make to a statement like that; I think he knew I was a dog handler, knew how provocative he was being. I don’t remember whether I offered a defense. In the first place, it really wasn’t my specialty, so I didn’t have a very knowledgeable defense to offer; in the second, it occurred to me that maybe an HRD handler had failed the man badly enough to earn that attitude.
Suffice it to say, we dog handlers sometimes struggle with a perception of our value that is, shall we say, rather deflated compared to what we think we deserve.
One major misconception that I think lurks in the artificial scent detection world is that so-called “detection dogs” do just that — detect a smell. And if you're screening suitcases for bombs, fish for spoilage, or whatever, that may be mainly true. But as any dog handler in search and rescue or human remains detection can tell you, detecting the smell is the easy part. Most of the work goes into locating the source, and getting the dog to tell you when she’s done it.
Problem One, the “getting them to tell us” part, comes from the balance between what researchers call type 1 errors and type 2 errors. A type 1 error is when you fail to detect something that's there. Type 2 is when you get a false detection of something that isn’t there. The issue is that anything you do to reduce one error type tends to increase your errors of the other type. Worried about missing a signal? You can increase your sensitivity, but that will also increase your rate of false detections. Want a “bomb-proof” indication, with utter certainty that when the dog (or device) says it’s there it really is? Inevitably, you’ll increase your number of total misses.
I don’t think a device is any more or less “reliable” than a dog on this count — it’s a phenomenon of the physics of detection, independent of the detector. But I will grant you can probably work out the optimal balance more easily on a device — and engineers are probably better at dispassionately working out “I’m willing to accept an X rate of type 1 errors if I can get a Y rate of type 2 errors” than dog handlers, who tend to ignore the issue and try for zero type 2 errors and assume it means zero type 1 errors as well, which it almost always doesn’t.
Problem Two is that localization issue. And ladies and gentleman, it is one bitch of a problem. The difficulty revolves around the nature of the signal rather than the detector: scent plumes consist of discontinuous clumps of scent that don’t simply or clearly point the way to the source in a snapshot. As often as not, the source lies in the direction of the weaker scent. Researchers have learned a lot about how animals negotiate intermittent scent plumes quickly and efficiently — and surprise, surprise, they don’t try to build an exact map of the plume. They follow these simple but effective rules of engagement:
if you detect the scent, dash into the wind
if you lose the scent, cast back and forth across the wind in until you detect it again
if the wind dies down, even if you’re detecting scent, stop and wait for it to pick up again so you have the directional cue to know which way to go*
Believe it or not, under conditions of any significant convection, this is all you need to do to find the source. It doesn’t make the job easy — there’s a lot of casting about to be done before you get close enough to the source to make any real headway — but it’s rapid and accurate. (Dog handlers’ search patterns, by the way, are just fancy versions of casting, to make sure that, while searching for the plume, we’re fully covering the area within the artificial boundaries dictated for us by the command staff’s planning.)
I think that part of the problem is that the engineers who work on artificial noses understand detection but they don’t know anything about search. Going back to our “pure” detection dog, checking your suitcase to see if you’ve smuggled fruit, there really is no search function. It’s just: is the target scent there, yes or no?
For the police dog handler searching a basement for a buried body or a warehouse for a bad guy, the search part is simple enough that the dog pretty much does it unguided. In these circumstances, I think this detector might be able to help because it may be a matter of just walking around until the signal maximizes. I think that’s how dogs search such small areas, and how they work “scent pools” — in other words, find the source when there’s no wind moving the scent and you essentially have a uniform murk, the center of which you need to pinpoint.
But we wilderness handlers think in terms of areas of 40 to 160 acres. Certainly, if you’ve got a body that’s truly missing — as opposed to believing it’s in a modest-sized back yard but needing to pinpoint it — the kind of painstaking mapping of intensities that the artificial approach would seem not to be the way to go. You need to find the scent plume in the first place, which brings us back to dashing and casting, at which point the device’s lesser sensitivity would probably put it at a disadvantage compared with the dog.
I haven’t even touched the issue — found with bodies under water or under collapsed structures — that the scent doesn’t always come up directly over the body, but can often take a winding route to the surface so it emerges some distance away. That’s an in-built limitation of detection by scent that isn’t going to go away no matter what detector we’re using.
I’ve said it before: We dog handlers will eventually be put out of business, but it won’t be by artificial noses. It’ll be by something like a Star Trek communicator that immediately calls for help when you’re in trouble and gives a sub-meter-accurate location to the rescue crews looking for you. Our mobile phones are already so close; it may be that what takes the longest is working out the privacy issues of automatically sharing that kind of data with the authorities.**
In the meantime, here’s my fear: that we’ll be replaced by something that isn’t as good, merely because we’ve failed to document our utility to the search effort. Or worse, because somebody screwed up, turning us all into suspected “urine detection dog handlers.” Either way, it’s a matter of the standards — operational standards, yes, but also standards of proof — to which we hold ourselves.
It may be more up to us than we realize.
*I’m simplifying a bit here. For one thing, different species do seem to riff differently depending on their specific scent mission. For another, even among the species that do employ dash-and-cast, there are different strategies for different atmospheric conditions. Dog handlers will have seen another pattern, called “weaving,” in which the dog works into the wind in a weaving pattern that gets narrower and narrower until you reach the source. There’s evidence that this behavior results from a superimposition of dashing and casting, a response to a relatively “clean,” low convection scent plume without large gaps of scent. Your major challenges under those conditions — usually at night or on cloudy, windy, or winter days — are to keep track of the edges of the plume so you don’t run out of it, and to keep your nose from desensitizing to the smell.
A new paper, interestingly, combines drone technology with the scent organs transplanted from live moths, to search robotically for a scent. And their search strategy consists of dash-and-cast. So some scientists are both aware of and working to crack the localization problem. You can read it here.
**Since I wrote this, drone technology has really advanced a lot, and we're looking at another way SAR dogs might be replaced: a swarm of intelligent drones you release into the woods, which are capable of avoiding the trees and each other as they search.