Using Remote Sensing to Identify Illegal Crop Plantation

Identification of Cannabis plantation

By Manuel Lukas

Results & Analysis


Key Results



Figure 1: NDVI transformation of (a) Image A on the left and (b) Image B on the right,

(c) The maximum difference between image A and B


 Supervised Classification

From the transformation image (Figure 1), it could be seen that there is definitely changes in spectral reflectance between image A and B. The intersection of red lines in Figure 1 shows the point on the location of the plantation from the given coordinates. Figure 1c shows the difference in reflectance value between image A and B, which have the highest difference of 0.1046. Other features such as water and soil have less than 0.0178 in difference.













Figure 2: Supervised classification results (L-R) Maximum Likelihood, Minimum Distance,

Parallelepiped, and Mahalanobis Distance


It is very clear as shown in Figure 2, that out of the 4 decisions rule of classification the best algorithm is the mahalanobis distance. The first reasons is because mahalanobis distance algorithm are able to differentiate the 5 classes assigned than the other algorithms. Most of the other algorithms are able to classified only 2 to 3 classes. In terms of the accuracy however, both maximum likelihood and mahalanobis distance performs similarly. Even though none of the algorithms are able to classified cannabis plantation only on the study area (in centre of the image).


 Spectral Angle Mapper

Image B results will be analysed using three different point (not in any particular order);

1. The number of pixel in the study area (close to 4 pixels in any directions from the line intersection),

2. The number of other pixels that is included as cannabis but outside the boundary specified (4 pixels),

3. The maximum angle value that the spectra reference used for image classification.

4. Equivalent results will be analysed further by looking at the maximum angle value.

5. The smaller maximum angle value will be concluded as the best spectra.


For image A analysis, a different approach is taken to determine the best spectra. Since image A is the image before the cannabis are planted, there should not be any pixels being classified as cannabis in the ROI; especially within 4 pixels of the study area in any direction. 4 pixels in each side of a square will made up of 16 pixels, which is equivalent to 1 hectare in area.


In comparison between the original image B results and the calibrated image B results it is found that;

1. The calibrated image able to use lower maximum angle value than the original image in classifying cannabis.

2. The original image have more pixels within the study area being classified as cannabis, however the pixels outside the study area that being classified as cannabis also greater than the calibrated image. The best results for calibrated image have only one pixel outside the study area being classified as cannabis.

3. The results that are labelled as the best classification, occurs more in the calibrated image.


In comparison between the original image A results and the calibrated image A results it is found that;

1. Results from the calibrated image require lower maximum angles value; there is only one pixel that is classified closest to cannabis reference spectra.

2. In the original image, the number pixels classified as cannabis is much greater than in the calibrated image. The number of results that is similar to the calibrated image results in the original image is the same.


Further analysis is done to determine which spectra is the best for classification. This is done by comparing the best classification results of image B with image A of the same results that uses the same spectra, in the same sample, and uses the same maximum angle value; not necessarily the best classification of image A. It is found that during the comparison of sample 3 spectra 4, this particular reference spectra provide the best results.


Sample 3 Sample 3

Spectra 4(0.20) Spectra 4(0.20)

Figure 3: (a) Image B(s) band 2-4 (left), (b) Image A(s) band 2-4 (right)


Figure 3 shows that at the same maximum angle value, the study area in image B have been classified as cannabis (Figure 3a); however image A did not have a pixels within the study area being classified as cannabis (Figure 3b). It also happens that spectra 4 from sample 3 are one of the best classification results in image A as well. This shows that by using spectra 4 from sample 3 on the original image, SAM classification are able to shows precise results in classifying cannabis in the time of plantation and no classification during the time before the plantation. No other spectra are able to perform this comparison.


Band 2, 4, and 5 are generally very good in producing better classification results than Band 1 and 4, however, it required higher maximum angle value to achieve these results. As have been mentioned, the higher the maximum angle value, the further it become for the classification closer to the cannabis reference spectra. When Band 2, 4, and 5 is used in the classification of white box tree, it does not produce as much classification of white box as band 2 and 4. It also requires higher value of maximum angle.


All the classification has been done based on a small region where the location is known, approximately 1.5 km square in area. Using larger area for classification will results other vegetation in massive quantity being classified as cannabis. The caused is probably not enough spectral resolution in the multi-spectral imagery to differentiate the spectra of cannabis with other vegetation in more details.