Analog computation of crop performance

The obvious way to do this is by looking on the other side of the Equator. They got your answer, and six months in advance! Well, maybe not quite. I feel certain that few Brazilian soybean growers are ever contending with frost-limited growing seasons.

But speaking of Latin America, may I assert that buses there can carry anything, and to prove it, on one trip I packed a William F. Buckley novel. It was not too incoherent, though it was about spies. At one point, two guys - James Jesus Angleton and Kim Philby? - are debating how to assess clandestine Soviet agriculture reports. You know how this goes. Is the source lying? Is he lying willfully? If he is lying and willfully, what does that mean? If he is lying but under duress, what does THAT mean? And who the heck cares? That's me, not James or Kim. And here is more of me: why not just do aerial photography?

It's been discussed. One presumes photo interpretation is a mix of human-eye and automated review. One further presumes that the latter is an exercise in digital computation: the rendering of photographic data as numbers, which are after all the things to compare to other numbers. This is part of the much broader fields of image recognition and computer vision...which are themselves hard to ignore in any debate on whether computers can even do such stuff. As with language translation, one might back up a bit and ask first if digital computers can do such stuff. I'd guess not; I'd guess that brains don't digitize words. I am also guessing, here, that brains don't digitize pictures either.

Here, I am thinking about analog computation. The analog computer would compare whole images, one of the target image (say, a vigorous beet plant) and something much bigger (a beet field in an unpredictable state of overall health), and find in the latter copies, in varying sizes and orientations, of the former. Digitization would require arduous adjustments for unexpectedly sized and/or nonuniformly twisted or tilted images. Analog techniques, if you can find an analog, ought to be much faster.

The issue has been addressed: this is an example, though its "analog" element has to do with "contour detection" and "image segmentation," not what happens after one has satisfactorily detected and segmented. I do not propose to review here what is clearly a busily investigated subject, but will merely touch on what I think is the way analog computation for image recognition must go, which is topological. If you can break up an image into a map - areas with borders, borders bespeaking connectivity - then the actual sizes of the areas, their shapes even, are irrelevant to the determination of similarity or equivalence. It's the connectivity that topology seeks to quantify and which an image-recognition program must either figure out (if it's digital) or apprehend (if it's analog).


I might add that in the case of that spy story, the reason crop reports were of interest was that Soviet agriculture's fortunes were supposed to depend heavily on Soviet military labor details, practically to the point of making foodgrowing and warfighting a zero-sum game. So perhaps what aerial photography might have been seeking here was army helmets, not beet crowns. And I wonder if it still is. Much can have happened between 1992 and now, but the fact that even at that time I could see at all a sovkhoz here and a kolkhoz there on the long cab ride from Odessa to Kiev suggested the Soviet imperative to badness had a lot of staying power.