Take the 2-minute tour ×
Geographic Information Systems Stack Exchange is a question and answer site for cartographers, geographers and GIS professionals. It's 100% free, no registration required.

I have a Boolean raster.

In the grey areas of the raster I would like to fit a given size polygon within a contiguous extent.

Basically, I have an irregular polygon, and i would like to "fit" a known polygon within the extent of the irregular polygon as many times as possible.

Direction of the polygon does not matter, and it could be a square. I would like for it to fit graphically, but if it just attached a number to the polygon (# that fit) that would work as well.

Using Arcgis 10.

share|improve this question
7  
This is a very hard problem. It takes a lot of work just to fit as many circles into a square as possible, for example. When the original polygon is complicated--as in the illustration--you need some powerful optimization procedures. The best method I have found for this problem is simulated annealing, but that won't be available in ArcGIS and it would take extremely crafty scripting to script it (ArcGIS is too slow). Could you perhaps relax your requirements a little, such as fitting the smaller polygon a sufficient number of times, rather than as many times as possible? –  whuber Jun 12 '12 at 17:36
1  
@whuber Thanks for editing my post. Yeah, sufficient number of times would work. Or, how about at a given angle orientation. ex. in the image above, I have fit the polygon as many times as i could have at that orientation, had I rotated them 90 degrees you could fit one more... –  Thad Jun 12 '12 at 17:41
1  
Yes, but it's also fraught with pitfalls. Some are elementary. For example, the ESRI-authored and -published text, "Getting to Know ArcView GIS" (for version 3) included an exercise in which a rectangle representing a soccer field was placed interactively within a polygon. The problem was, the exercise's answer was wrong because the author failed to project the data and the errors in using geographic coordinates were large enough to affect the result. The answer looked good in the GIS, but if anyone had attempted to build that field, they would have found there wasn't enough room for it :-). –  whuber Jun 12 '12 at 18:58
6  
@whuber I guess they thought a "ball park" figure was sufficient. –  Kirk Kuykendall Jun 14 '12 at 20:08
2  
In the general case of irregular polygon within irregular polygon, this is a computationally intractable problem: Finding an optimal solution is not a plausible goal in all cases, and it's likely NP-complete from a technical perspective: Which cases those are cannot be predetermined. If you constrain the problem significantly, some iterative random fitting algorithms are likely to give you reasonably high numbers. My feeling if this is an assignment is that they're not looking for the correct answer, they're looking for creative approaches. –  MappingTomorrow Jun 14 '12 at 23:28
show 9 more comments

3 Answers

up vote 16 down vote accepted
+150

There are many ways to approach this problem. The raster format of the data suggests a raster-based approach; in reviewing those approaches, a formulation of the problem as a binary integer linear program looks promising, because it is very much in the spirit of many GIS site-selection analyses and can readily be adapted to them.

In this formulation, we enumerate all possible positions and orientations of the filling polygon(s), which I will refer to as "tiles." Associated with each tile is a measure of its "goodness." The objective is to find a collection of non-overlapping tiles whose total goodness is as large as possible. Here, we can take the goodness of each tile to be the area it covers. (In more data-rich and sophisticated decision environments, we may be computing the goodness as a combination of properties of the cells included within each tile, properties perhaps related to visibility, proximity to other things, and so on.)

The constraints on this problem are simply that no two tiles within a solution may overlap.

This can be framed a little more abstractly, in a way conducive to efficient computation, by enumerating the cells in the polygon to be filled (the "region") 1, 2, ..., M. Any tile placement can be encoded with an indicator vector of zeros and ones, letting the ones correspond to cells covered by the tile and zeros elsewhere. In this encoding, all the information needed about a collection of tiles can be found by summing their indicator vectors (component by component, as usual): the sum will be nonzero exactly where at least one tile covers a cell and the sum will be greater than one anywhere two or more tiles overlap. (The sum effectively counts the amount of overlap.)

One more little abstraction: the set of possible tile placements can itself be enumerated, say 1, 2, ..., N. The selection of any set of tile placements itself corresponds to an indicator vector where the ones designate the tiles to be placed.

Here's a tiny illustration to fix the ideas. It is accompanied with the Mathematica code used to do the calculations, so that the programming difficulty (or lack thereof) can be evident.

First, we depict a region to be tiled:

region =  {{0, 0, 1, 1}, {1, 1, 1, 1}, {1, 1, 1, 1}, {1, 1, 1, 1}};

Figure 1: region

If we number its cells from left to right, starting at the top, the indicator vector for the region has 16 entries:

Flatten[region]

{0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}

Let's use the following tile, along with all rotations by multiples of 90 degrees:

tileSet = {{{1, 1}, {1, 0}}};

Figure 2: tile

Code to generate rotations (and reflections):

apply[s_List, alpha] := Reverse /@ s;
apply[s_List, beta] := Transpose[s];
apply[s_List, g_List] := Fold[apply, s, g];
group = FoldList[Append, {}, Riffle[ConstantArray[alpha, 4], beta]];
tiles = Union[Flatten[Outer[apply[#1, #2] &, tileSet, group, 1], 1]];

(This somewhat opaque computation is explained in a reply at http://math.stackexchange.com/a/159159, which shows it simply produces all possible rotations and reflections of a tile and then removes any duplicate results.)

Suppose we were to place the tile as shown here:

Figure 3: tile placement

Cells 3, 6, and 7 are covered in this placement. That is designated by the indicator vector

{0, 0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0}

If we shift this tile one column to the right, that indicator vector would instead be

{0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0}

The combination of trying to place tiles at both these positions simultaneously is determined by the sum of these indicators,

{0, 0, 1, 1, 0, 1, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0}

The 2 in the seventh position shows these overlap in one cell (second row down, third column from the left). Because we do not want overlap, we will require that the sum of the vectors in any valid solution must have no entries exceeding 1.

It turns out that for this problem, 29 combinations of orientation and position are possible for the tiles. (This was found with a simple bit of coding involving an exhaustive search.) We can depict all 29 possibilities by drawing their indicators as column vectors. (Using columns instead of rows is conventional.) Here's a picture of the resulting array, which will have 16 rows (one for each possible cell in the rectangle) and 29 columns:

makeAllTiles[tile_, {n_Integer, m_Integer}] := 
  With[{ m0 = Length[tile], n0 = Length[First[tile]]},
   Flatten[
    Table[ArrayPad[tile, {{i, m - m0 - i}, {j, n - n0 - j}}],  {i, 0, m - m0}, {j, 0, n - n0}], 1]];
allTiles = Flatten[ParallelMap[makeAllTiles[#, ImageDimensions[regionImage]] & , tiles], 1];
allTiles = Parallelize[
   Select[allTiles, (regionVector . Flatten[#]) >= (Plus @@ (Flatten[#])) &]];
options = Transpose[Flatten /@ allTiles];

Figure 4: options array

(The previous two indicator vectors appear as the first two columns at the left.) The sharp-eyed reader may have noticed several opportunities for parallel processing: these calculations can take a few seconds.

All the foregoing can be restated compactly using matrix notation:

  • F is this array of options, with M rows and N columns.

  • X is the indicator of a set of tile placements, of length N.

  • b is an N-vector of ones.

  • R is the indicator for the region; it is an M-vector.

The total "goodness" associated with any possible solution X, equals R.F.X, because F.X is the indicator of the cells covered by X and the product with R sums these values. (We could weight R if we wished the solutions to favor or avoid certain areas in the region.) This is to be maximized. Because we can write it as (R.F).X, it is a linear function of X: this is important. (In the code below, the variable c contains R.F.)

The constraints are that

  1. All elements of X must be non-negative;

  2. All elements of X must be less than 1 (which is the corresponding entry in b);

  3. All elements of X must be integral.

Constraints (1) and (2) make this a linear program, while the third requirement turns it into an integer linear program.

There exist many packages for solving integer linear programs expressed in exactly this form. They are capable of handling values of M and N into the tens or even hundreds of thousands. That's probably good enough for some real-world applications.


As our first illustration, I computed a solution for the preceding example using Mathematica 8's LinearProgramming command. (This will minimize a linear objective function. Minimization is easily turned to maximization by negating the objective function.) It returned a solution (as a list of tiles and their positions) in 0.011 seconds:

b = ConstantArray[-1, Length[options]];
c = -Flatten[region].options;
lu = ConstantArray[{0, 1}, Length[First[options]]];
x = LinearProgramming[c, -options, b, lu, Integers, Tolerance -> 0.05];
If[! ListQ[x] || Max[options.x] > 1, x = {}];
solution = allTiles[[Select[x Range[Length[x]], # > 0 &]]];

Figure 5: solution

The gray cells are not in the region at all; the white cells were not covered by this solution.

You can work out (by hand) many other tilings that are just as good as this one--but you cannot find any better ones. That's a potential limitation of this approach: it gives you one best solution, even when there is more than one. (There are some workarounds: if we reorder the columns of X, the problem remains unchanged, but the software often chooses a different solution as a result. However, this behavior is unpredictable.)

As a second illustration, to be more realistic, let's consider the region in the question. By importing the image and resampling it, I represented it with a 69 by 81 grid:

Figure 6: Region

The region comprises 2156 cells of this grid.

To make things interesting, and to illustrate the generality of the linear programming setup, let's try to cover as much of this region as possible with two kinds of rectangles:

Figure 7: tiles

One is 17 by 9 (153 cells) and the other is 15 by 11 (165 cells). We might prefer to use the second, because it is larger, but the first is skinnier and can fit in tighter places. Let's see!

The program now involves N = 5589 possible tile placements. It's fairly big! After 6.3 seconds of calculation, Mathematica came up with this ten-tile solution:

Figure 8: solution

Because of some of the slack (.e.g, we could shift the bottom left tile up to four columns to its left), there are obviously some other solutions differing slightly from this one.

share|improve this answer
    
An earlier version of this solution (but not quite as good) appears on the Mathematica site at mathematica.stackexchange.com/a/6888. It may be worth noting, too, that a minor variation of the formulation can be used to solve the problem of completely covering the region with as few tiles as possible (allowing some overlaps, of course): this would solve the "pothole patching" problem. –  whuber Jun 18 '12 at 18:09
1  
In the interests of space, this answer does not describe some potentially helpful improvements. For instance, after finding all possible tile positions (as indicator vectors), you can add them all up to find which cells can actually be covered by some tile. The set of such cells breaks into two separate connected components in the second example. This means the problem can independently be solved in the two components, substantially reducing its size (and therefore computing time). Such initial simplifications tend to be important for tackling real-world problems. –  whuber Jun 19 '12 at 0:58
    
Great effort and answer. Chris's answer was also helpful. Thanks everyone for the help! Works, and got me moving in the right direction again. –  Thad Jun 19 '12 at 12:57
    
Wow! I was interested in a similar problem and this post gave me new perspective. Thank you. What if R is bigger (e.g. 140x140≈20000), are there any ways to reduce computation cost? Do you know any papers related to this problem? My search keywords don't lead me in the right way (until now). –  nimcap Aug 29 '13 at 9:13
    
@nimcap This is an important class of problems, so much research goes on. Keywords to search on would begin with "mixed integer linear program" and branch out from there based on what you find. –  whuber Aug 29 '13 at 15:29
show 1 more comment

The link to On Genetic Algorithms for the Packing of Polygons, provided in my answer to a similar question, might be useful. It seems like the method could be generalized to work with arbitrary container shapes (and not just rectangles).

share|improve this answer
    
That paper has some nice ideas (+1), but all its algorithms focus, in a fundamental way, on packing polygons within rectangular regions. This is because it represents packings with a discrete data structure (a sequence of polygons along with their orientations) which represents a set of procedures in which the polygons are slid, parallel to the square's sides, towards a designated corner. It appears that such a simple discrete encoding would be less effective for more complicated regions. Maybe an initial simplification of the regions in the grid would help. –  whuber Jun 14 '12 at 21:04
add comment

For the highly constrained subset you mentioned (square/triangular tiling in a pothole), assuming the explicit optimizations above, this pseudocode should arrive at an approximate answer by simply taking you through the possibilities with a high resolution, brute forcing the problem. It won't work correctly for situations where individual tile rotation can see gains, like rectangle tiles or a highly irregular container. This is 1 million iterations, you can try more if necessary.

Assume a square with sides of length L

Create a checkerboard pattern of squares, which is at least of the dimensions of the extent of the container, plus at least 1L in each direction.

N = 0

DX = 0

DY = 0

DR = 0

Reset checkerboard position to original centroid

For (R=1:100)

For (Y=1:100)

For (X=1:100)

M = Count number of squares completely within container

If (M>N)

DR=R

DY=Y

DX=X

N=M

Move checkerboard east by L/100

Reset checkerboard easting

Move checkerboard north by L/100

Reset checkerboard northing

Rotate checkerboard by 3.6 degrees CW around its centroid

DY = DY * L

DX = DX * L

Reset checkerboard to original position and rotation

Print DR & ", " & DX & ", and " & DY & " are the final translation/rotation matrix"

Rotate checkerboard by DR

Translate checkerboard by DX, DY

Select squares which are completely within container

Export squares

share|improve this answer
    
If you try this procedure on a 2 by 5 region with a cell missing along the middle of one long edge, you will find you can put only one 2 by 2 square into it. However, two such squares easily fit. The problem is that they are not part of a regular "checkerboard" pattern. This difficulty is one of the things that makes this problem fairly hard. –  whuber Jun 18 '12 at 20:48
1  
Yup. If you have a container shape irregular enough that it can support multiple discontiguous regular patterns on the order of a few cells each, this ends up very far from optimal. Adding things like that to the possibility space increases processing time very rapidly, and requires a certain degree of planning for the particular case you're targetting. –  MappingTomorrow Jun 19 '12 at 0:15
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.