Evaluation of crop yield can be tedious because sampling methods requires the actual counting of fruits for a whole tree or canopy area. If we want to optimize this time demanding task we can use new and open source machine learning algorithms available. We have selected Scikit-Learn for this tutorial, a machine learning library in Python for it ease to use, the available documentation and the sort of available tools.

The tutorial covers the whole procedure of image representation, point of interest selection, template matching, cluster analysis and fruit counting. Python scripting was done in Jupyter Notebook, it is interactive and allows the user to add more points of interest or remove inaccurate points.

Tutorial

Code

This is the whole code in Python:

# coding: utf-8

# # Import the require libraries and images

get_ipython().magic('matplotlib notebook')
import numpy as np
import matplotlib.pyplot as plt
from skimage.feature import match_template
from PIL import Image

#ImagenTotal = np.asarray(Image.open('Input\OlivoTotal.png'))
ImagenTotal = np.asarray(Image.open('Input/Citrus.png'))
#ImagenTotal = np.asarray(Image.open('Input\BananoTotal.png'))

# # Interactive selection of points

#define empty cells 
puntosinteres = []

fig = plt.figure(figsize=(9, 6))
ax = fig.add_subplot(111)
ax.imshow(ImagenTotal, cmap=plt.cm.gray)
text=ax.text(0,0, "", va="bottom", ha="left")

#interactive function that stores points clicked on the image
def onclick(event):
    tx = 'button=%d, x=%d, y=%d, xdata=%f, ydata=%f' % (event.button, event.x, event.y, event.xdata, event.ydata)
    text.set_text(tx)
    puntosinteres.append([event.xdata, event.ydata])

cid = fig.canvas.mpl_connect('button_press_event', onclick)

# amount of points clicked
len(puntosinteres)

#plot points over the image and select more if you want
fig = plt.figure(figsize=(10, 6))
ax = fig.add_subplot(111)
ax.imshow(ImagenTotal, cmap=plt.cm.gray)
ax.scatter([x[0] for x in puntosinteres],[y[1] for y in puntosinteres],c='red', marker='+', s=8)
text=ax.text(0,0, "", va="bottom", ha="left")

def onclick(event):
    tx = 'button=%d, x=%d, y=%d, xdata=%f, ydata=%f' % (event.button, event.x, event.y, event.xdata, event.ydata)
    text.set_text(tx)
    puntosinteres.append([event.xdata, event.ydata])

cid = fig.canvas.mpl_connect('button_press_event', onclick)

#show all the points of interest, please be careful to have a complete image, otherwise the model wont run
fig, ax = plt.subplots(len(puntosinteres)//6+1, 6)
i = 0
for item in puntosinteres:
    xinteres = int(item[0])
    yinteres = int(item[1])
    radio = 20
    ax[i//6,i-i//6*6].imshow(ImagenTotal)
    ax[i//6,i-i//6*6].plot(xinteres,yinteres,color='red', linestyle='dashed', marker='+',
     markerfacecolor='blue', markersize=8)
    ax[i//6,i-i//6*6].set_xlim(xinteres-radio,xinteres+radio)
    ax[i//6,i-i//6*6].set_ylim(yinteres-radio,yinteres+radio)
    ax[i//6,i-i//6*6].axis('off')
    ax[i//6,i-i//6*6].set_title(i)
    i+=1

#in case you have a wrong point or a incomplete image please uncomment the following line with the point index to delete it
del puntosinteres[4]
len(puntosinteres)

# # Match the image to the template

listaresultados = []
for punto in puntosinteres:
    xinteres = int(punto[0])
    yinteres = int(punto[1])
    radio=10
    imagenband = ImagenTotal[:,:,0]
    templateband = ImagenTotal[yinteres-radio:yinteres+radio,xinteres-radio:xinteres+radio,0]
    result= match_template(imagenband, templateband)
    result = np.where(result>0.75)
    listaresultados.append(result)

# # Plot interpreted points over the image

from itertools import cycle
cycol = cycle('bgrcmk')

fig = plt.figure(figsize=(10, 8))
ax = fig.add_subplot(111)

i = 1
for lista in listaresultados:
    ax.plot(lista[1],lista[0], '.', linewidth=0, markerfacecolor=next(cycol), label=i)
    i+=1
ax.legend(loc='upper center', bbox_to_anchor=(0.5, -0.05),
          fancybox=True, shadow=True, ncol=5)
ax.imshow(ImagenTotal[radio:-radio,radio:-radio,:])

# # Cluster analisys with Birch algorithm

datalist = [np.asarray(pares).T for pares in listaresultados]
print(len(datalist))
datalist = np.vstack(datalist)
print(datalist)

from sklearn.cluster import Birch
brc = Birch(branching_factor=10000, n_clusters=None, threshold=9, compute_labels=True)
brc.fit(datalist)
puntosbirch = brc.subcluster_centers_
fig = plt.figure(figsize=(10, 8))
ax = fig.add_subplot(111)
ax.scatter(puntosbirch[:,[1]],puntosbirch[:,[0]], marker='+',color='red')
ax.imshow(ImagenTotal[radio:-radio,radio:-radio,:])

len(puntosbirch)