• Tidak ada hasil yang ditemukan

5.2 Results and Discussion

5.2.3 Panchromatic Image Texture Analyses

As was mentioned in section 5.1.3.4 above, although the multi-spectral resolution of 2.4 m is regarded as a high resolution (it is in fact that highest commercial resolution currently available from a satellite platform (Digital Globe, 2005)), it was still not sufficient to be able to separate crop from weeds on the basis of the crop rows.

However, when viewing the 0.6 m resolution panchromatic imagery the crop rows were very apparent, particularly as the crop age exceeded four months (see Figure 5.2.3A). Even in very young stands, less than three months old, crop rows were often discernable where there had been a "line-cleaning" operation. This is where a clear row is hoed down to bare soil along the planting line. This line is usually 1 t01.5 m wide, with the centre line of this cleared area being a distance of 3 m perpendicular to the adjacent rows on either side (see Figure 5.2.1).

Figure 5.2.1 Line-cleaning along crop row

Fig. a December2003

Fig. c December 2004

Fig.b May2004

Fig. d March2005

Fig. e April 2005

j

N

110000

Figure 5.2.2 Example of Four-Class Unsupervised Classification (Compt. E022)

Figure 5.2.3 Comparison of Texture Analyses Processes (part of Compt. E014)

The purpose of this hoed row is to remove any initial weed competition as the plant crop becomes established, following its planting (if a seedling) or germination (where seed is sown directly into planting line - this occurs only in wattle establishment). It was this visibility of the crop rows that was believed to be a key element in being able to separate crop from weed within the vegetation signal, provided a means could be found to merge the two differing resolution datasets. The most effective methodology to utilise this characteristic was to apply some form of textural analysis (Tso and Mather, 2001; Lillesand and Kiefer, 2000; Janssen, 2000) to enhance the lineation features of the crop rows. Three textural parameters and a frequency domain parameter, Fourier Analysis, were tested in order to determine an optimal methodology. The textural parameters were edge enhancement, variance and skewness (see Figure 5.2.3), and were applied using a convolution window, the size of which was determined using semivariograms, as detailed in section 5.1.3.4 above.

5.2.3.1 Texture Analysis - Variance

The Variance function produced images that were smoothed and tended to group areas that had a similar texture into clumps. As such it actually did the opposite of what was intended, in terms of trying to define the crop rows more sharply (see Figure 5.2.38). Only where the rows were already very sharply defined in the original image was there some definition of the rows, and then only at certain stages of the crop development. It was also noted that the convolution window size played a role in the amount of row definition that occurred. The only lineation that this methodology did highlight was the brushwood rows and roads within the compartments. However, this was due more to the fact that these features had sufficient uniformity within themselves to cause the smoothing function to group them together, rather than the lineation itself. What could be deduced from this test was that the crop rows themselves did not possess SUfficiently uniform texture such that a smoothing function could uniquely identify them. Although not tested in this study, this methodology could play a role in trying to identify concentrations of vegetation growth that might indicate potential weed patches. It might also have application if one wanted to obtain a measure of stand uniformity. An inspection the histograms of the variance texture images showed that the bulk of the values were concentrated close to the origin. Even when the values were stretched, however, one gained little additional information. As this methodology did not assist in identifying crop rows it was discarded as a suitable enhancement method.

5.2.3.2 Texture Analysis - Skewness

The Skewness function did define linear features much more effectively, except that it tended to define the inter-row areas, rather than the crop rows as was expected (see Figures 5.2.3C; 5.2.4). This was probably due to the inter-row signal being the more dominant signal, particularly in the younger stages of the study sites. With the planting rows being 3 m apart, and early canopy width along the rows being less than 1.5 m, this meant that the inter-row signal would be about 60-70% of the signal range. This would result in the Skewness statistic highlighting this. However, this factor caused it to also be discarded as a suitable enhancement technique for the aims of this study. It could however have useful applications in some situations. As with the Variance function, the histogram data were again concentrated close to the origin, but stretching the values did not add much clarity or information to the resultant image. Up to the age of 12 months the Skewness function tended to define the image data into clumps, similar to the way the Variance function defined the data. After this age threshold, the lineation patterns became much more defined, as the inter-rows areas became more definitive.

5.2.3.3 Texture Analysis - Edge Enhancement

In contrast to both the above mentioned procedures, the Edge Enhancement function was able to detect the lineation virtually as soon as it had occurred in the form of a line-cleaning operation. In some cases this was even prior to the crop emerging as a distinctive feature. While other strong linear features such as the brushwood rows (both unburnt and burnt) were also highlighted, this did not hide the crop rows, as was the case with the other two processes (see Figure 5.2.3D). It also tended to maintain the crop row definition well into canopy closure, with the oldest study site at 26 months still having some indication of the crop rows. This methodology was by far the most successful at defining the crop rows (see Figure 5.2.5), and so was selected as the textural analysis method of choice for this study, which was based on the premise that crop rows could be used to assist in separating crop from weed. The edge enhanced images were the input to the two- class classified "Row; No Row" images (see section 5.1.3.4).

As mentioned in the literature review, various studies have attempted to produce tree counts based on individual tree identification (e.g. Jacobs and Mthembu, 2001;

Wulder et al., 2000), with minimum crown dimensions of 1.5 m required for 1 m

Fig.A Row Delineation

.-\

N

1:1 000

Fig.BSkewness Image highlighting Inter-Rows

~N

1:1 000

Figure 5.2.4 Illustration of Skewness Image highlighting Inter-Rows (part of E014)

Skewness image overlain on Row Delineation image. Green is Crop Row; Black is Skewness image highlighting inter-row areas.

Fig a Edge Enhanced Panchromatic Image (Bright pixels= Rows)

Fig. b Two CLass Reclassified Image highlighting Crop Rows

Fig. c Classified Rows overlain on Enhanced Pan Image

6

N

1:1 000

Figure 5.2.5 Illustration of Classified Row Image highlighting Crop Rows (part of Compt E014)

Fig. c shows Classified Row Image (Fig.b) overlain on Edge Enhanced Pan Image (Fig.a)

spatial resolution imagery being reported (Wulder et al., 2000). However, because this study focussed on the tree row as a whole, rather than individual tree crowns, successful identification of tree rows was achieved where crown dimensions were much smaller than figures reported in these other studies. This was a one of the unique findings of this study.

5.2.3.4 Frequency Domain Analysis - Fourier Transform

The Fourier Transform analysis did not provide a distinctive result that could be used to delineate the crop rows, although there were indications that crop rows were being identified to some extent. However, a great deal of processing was required to separate the "noise" signal from the crop row signal and in the end the edge enhancement process provided a much quicker and simpler solution, which was in agreement with Jensen (1996) regarding the application of the Fourier Transform.

What was particularly noticeable in some of the Fourier images was the presence of one or more frequency components (shown up as white dots), usually in some form of a linear arrangement, and based on the theory of the Fourier Transform, could indicate the frequencies associated with repeating lines across the image (Jensen, 1996). It was thought that these frequency components could indicate the crop rows.

However, this was not investigated in this study, but could possibly warrant further investigation (see section 5.3.2, Recommendations,below).