Crop-row detection of already known plant positions - How to?

Question

Through image recognition and segmentation I have already obtained an abstract representation of plants on a field (i. e. I exactly know all the coordinates of all plants in an image).

Now I want to detect the crop-rows in this abstract representation - and can't quite figure out how. My problems are:

the rows in the images may be a bit rotated and not exactly in north/south orientation (angles may vary between -10° to +10°)
The number of crop-rows per image can vary per image and is not fixed - also it is unknown before processing has taken place
The rotation of the crop-rows may be slightly different in each processed image
I have hundreds of images / representations to be processed (so doing it by hand is obviously not really feasible :-) ), so I would need an algorithm that I can later e.g. put into a loop

Can you perhaps help me with at least a strategy (or code snippets) to do such a crop row detection? Idealy in the end for each crop row I would have the parameters of a linear equation (y=m*x+t), so that abline() can be used, but I am open for anything. In the end it could look like something like this (here purely for illustration purposes done by hand):

Underlying data of the images is here:

structure(c(5278.072, 2632.564, 393.34, 4057.704, 3805.599, 611.269, 
1823.835, 3359.069, 3598.284, 5262.873, 2069.963, 1579.745, 4539.584, 
3579.977, 4296.46, 1831.153, 2333.835, 1126.639, 152.948, 4030.205, 
3368.738, 2066.733, 855.111, 2579.665, 3092.37, 1318.357, 1109.438, 
3578.606, 375.756, 3796.788, 4520.064, 1807.36, 5001.773, 87.272, 
4033.594, 836.708, 639.13, 3105.628, 1569.256, 2831.851, 826.444, 
3557.598, 1078.643, 576.266, 4789.585, 3091.929, 5239.658, 1099.954, 
1807.972, 2534.677, 4271.841, 5019.276, 2053.246, 1536.071, 3347.644, 
4019.766, 3793.392, 5257.628, 604.323, 2561.307, 1792.665, 884.25, 
109.456, 3066.108, 3750.833, 4511.819, 2815.08, 119.468, 4499.801, 
2582.512, 2822.354, 3773.842, 1054.719, 4251.171, 4002.476, 2018.277, 
1775.284, 4959.269, 2541.009, 4742.312, 2265.149, 3071.313, 1779.218, 
3972.64, 2822.409, 5217.848, 1265.449, 1522.899, 3057.732, 5364.729, 
346.341, 4226.012, 3287.299, 1767.18, 3991.963, 1811.498, 2785.251, 
4488.214, 822.509, 2016.435, 3022.344, 2528.079, 4470.315, 3017.716, 
572.771, 97.748, 5168.119, 4199.643, 2006.285, 3946.505, 2771.626, 
3495.94, 1745.531, 3734.241, 3265.819, 4963.116, 1058.788, 300.408, 
1252.845, 4453, 5411.107, 2768.93, 557.806, 2004.424, 2218.582, 
4214.073, 4698.292, 5149.238, 4953.886, 1238.343, 3502.518, 2753.044, 
5417.502, 1031.945, 2518.901, 1483.487, 4450.737, 2258.484, 289.261, 
2987.945, 5156.371, 4171.407, 1995.901, 781.96, 3918.94, 1974.667, 
316.758, 1470.993, 5160.868, 3237.828, 521.251, 787.228, 1039.416, 
1202.261, 3456.837, 4148.167, 2200.492, 2720.912, 4915.451, 3902.744, 
4435.419, 1209.418, 1471.057, 4641.269, 3913.51, 5412.672, 1953.878, 
2220.277, 4911.249, 1006.368, 2974.173, 4410.827, 1688.391, 293.729, 
1462.871, 4618.785, 5150.904, 2689.061, 1952.56, 5389.383, 2176.387, 
995.073, 4125.245, 498.978, 5137.266, 5358.118, 1444.34, 1674.431, 
2689.288, 2465.351, 4566.352, 765.125, 1196.984, 1687.859, 258.247, 
1914.911, 4575.408, 3421.147, 495.879, 979.079, 1922.943, 4097.704, 
737.439, 3410.562, 234.74, 2159.697, 471.983, 1418.991, 2440.575, 
1942.708, 1162.525, 5312.409, 2162.656, 5059.814, 1411.412, 4558.905, 
247.618, 4319.106, 3411.827, 1786.69, 1670.462, 1180.524, 1640.636, 
4715.993, 3576.548, 3566.57, 3589.872, 3565.564, 3531.571, 3415.178, 
3511.07, 3510.051, 3487.762, 3470.791, 3443.062, 3369.329, 3386.999, 
3387.786, 3277.473, 3376.266, 3421.932, 3387.869, 3367.994, 3346.403, 
3259.785, 3296.081, 3297.633, 3285.163, 3300.119, 2941.504, 3264.344, 
3277.9, 3235.499, 3198.869, 3235.508, 3156.907, 3221.313, 3123.96, 
3165.979, 3186.806, 3148.158, 3129.906, 3035.963, 2987.899, 3053.684, 
3050.107, 3052.643, 3037.767, 3037.525, 2994.456, 3006.454, 2960.606, 
2973.443, 2919.843, 2917.246, 2939.87, 2914.804, 2886.588, 2920.769, 
2906.616, 2908.866, 2868.052, 2885.769, 2860.088, 2801.168, 2853.439, 
2853.863, 2847.141, 2805.677, 2806.183, 2718.094, 2661.652, 2695.19, 
2656.518, 2612.372, 2603.286, 2602.449, 2591.63, 2595.714, 2593.287, 
2575.333, 2572.15, 2476.559, 2435.917, 2538.626, 2514.215, 2458.875, 
2477.5, 2385.366, 2421.47, 2220.899, 2397.842, 2396.848, 2393.501, 
2352.039, 2292.429, 2315.84, 2328.682, 2256.508, 2236.925, 2192.809, 
2241.279, 2144.107, 2195.016, 2185.86, 2112.28, 2098.085, 2020.843, 
1971.232, 1979.691, 1968.859, 1943.755, 1974.743, 1891.801, 1944.186, 
1951.423, 1872.022, 1928.441, 1880.504, 1912.82, 1893.822, 1878.889, 
1850.38, 1834.762, 1851.886, 1806.117, 1776.713, 1682.26, 1733.805, 
1714.941, 1700.778, 1686.258, 1703.367, 1549.601, 1682.525, 1563.277, 
1632.103, 1609.4, 1621.888, 1587.126, 1545.346, 1537.933, 1542.424, 
1366.974, 1494.822, 1498.618, 1494.055, 1450.098, 1407.89, 1345.613, 
1388.68, 1380.527, 1368.772, 1372.391, 1161.35, 1297.577, 1312.849, 
1304.972, 1286.721, 1292.485, 1257.53, 1241.146, 1263.164, 1217.146, 
1226.615, 993.046, 1166.837, 1112.254, 1072.249, 1117.723, 1061.758, 
1098.207, 1084.597, 1059.916, 1059.685, 1063.814, 1054.735, 944.2, 
982.653, 963.989, 969.55, 941.066, 907.014, 930.988, 776.849, 
877.918, 889.259, 805.872, 831.361, 803.752, 786.654, 791.649, 
814.271, 794.444, 776.833, 694.969, 664.718, 653.238, 661.703, 
652.696, 655.997, 637.118, 539.101, 555.694, 491.482, 459.712, 
453.73, 490.567, 391.441, 409.506, 319.697, 391.505, 390.46, 
308.658, 310.59, 285.799, 268.86, 245.89, 195.933, 243.418, 214.203, 
172.129, 173.754, 191.456, 194.795, 98.098, 99.4479999999999, 
62.1419999999998), .Dim = c(224L, 2L))

user_13223986 · Answer

Here is something that may help:

For each detected plant point, find the closest neighboring plant. Hopefully this finds a plant in the same crop row more often than not. If it's known a priori that images are roughly in north/south orientation, we should prefer looking more in the vertical direction to choose neighboring plants. One way to do that is to redefine "distance" for the nearest neighbor search as something anisotropic like

distance = 10 * (x0 - x1)² + (y0 - y1)²

Here is a plot of what this produces, making a line segment between each plant and its nearest neighbor:

It's not perfect, but could be a useful start. Most crop rows are lucky enough that a run of 4 or more plants are correctly chained together.

A thought on a possible strategy from here:

Identify the connected components, the "chains" of plants.
For each chain, regress a best fit line by least squares. Or better yet, use the RANSAC algorithm so that the fit robustly ignores a single stray plant in an otherwise colinear chain.
Again using the rough north/south orientation, consider the best fit line "valid" only if it's close enough to vertical. Supposing it is valid, find all plants that are close to the best fit line. If many plants are close, then the best fit line is likely a crop row.

Crop-row detection of already known plant positions - How to?

1 Answers