• No se han encontrado resultados

Apart from the reference data, the input data sources are the most important factor in tree species classification. If the input data sources do not yield any information that is capable of separating the species that are contained in the classification scheme, then the best classification algorithm trained on ideal reference data will not be able to separate the species correctly.

Table 6.12 gives an overview over the input data sources and the achieved results. The mean user’s accuracy is calculated for 10 pairs of training and test sets and its mean is given as µuser with a standard deviation of σuser. The mean of the overall

accuracy on the 10 test sets is given as µ with a standard deviation given as σ. The table is divided into five sections. The first section uses only one data source at a time. This gives information on the ability of the individual data sources to separate the six tree species used for the comparison. The second part of the table describes combinations of input data that contain the airborne color (RGB) and color infrared (CIR) images at a resolution of 0.8 m per pixel as well as the LIDAR nDSM and varying additional data sources. The third section of the table excludes the airborne images and the fourth part excludes LIDAR nDSM and intensity (I) data. In each section the data sources are sorted by the mean overall classification accuracy

in ascending order. The last line shows the results obtained on the two satellite data sets Système Probatoire d’Observation de la Terre (SPOT) and RapidEye. The following additional abbreviations are used in the table: SWIR is the short wavelength infrared band of the SPOT satellite data set. RE is the red edge band of the RapidEye satellite data. RapidEye-RE contains all bands of the RapidEye data set except for the red edge band (B, G, R, NIR).

Table 6.12: Comparison of input data sources

µuser σuser µ σ

single input source types

RGB 46.54 % 5.60 pp 33.39 % 1.73 pp CIR 48.89 % 1.56 pp 44.12 % 1.97 pp SPOT 54.46 % 3.79 pp 48.46 % 2.32 pp RGB+CIR 59.00 % 1.38 pp 50.07 % 1.35 pp RapidEye-RE 58.01 % 2.98 pp 55.62 % 2.53 pp RapidEye 67.96 % 1.72 pp 64.43 % 2.26 pp

airborne images and LiDAR data

RGB+CIR+nDSM 63.62 % 1.08 pp 55.84 % 1.97 pp RGB+CIR+nDSM+I 66.87 % 1.45 pp 61.85 % 2.21 pp RGB+CIR+nDSM+SWIR 66.13 % 2.06 pp 62.97 % 1.53 pp RGB+CIR+nDSM+RE 66.94 % 1.21 pp 63.11 % 1.50 pp RGB+CIR+nDSM+SWIR+RE 67.99 % 1.51 pp 65.16 % 1.84 pp RGB+CIR+nDSM+SPOT 68.91 % 1.49 pp 66.32 % 1.81 pp RGB+CIR+nDSM+SWIR+I 69.49 % 1.18 pp 66.85 % 1.26 pp RGB+CIR+nDSM+RE+I 70.94 % 1.35 pp 68.17 % 1.76 pp RGB+CIR+nDSM+SPOT+I 72.19 % 1.68 pp 69.67 % 2.36 pp RGB+CIR+nDSM+SWIR+RE+I 72.83 % 1.33 pp 70.92 % 1.42 pp RGB+CIR+nDSM+RapidEye 76.29 % 1.62 pp 74.12 % 1.77 pp RGB+CIR+nDSM+RapidEye+I 77.22 % 1.54 pp 75.44 % 1.58 pp RGB+CIR+nDSM+SPOT+RapidEye 77.40 % 1.46 pp 75.73 % 1.79 pp RGB+CIR+nDSM+SPOT+RapidEye+I 78.01 % 2.00 pp 76.54 % 2.16 pp LiDAR data, no airborne images

nDSM+SPOT+I 62.44 % 2.59 pp 57.01 % 2.52 pp nDSM+RapidEye-RE 65.19 % 2.25 pp 63.13 % 1.99 pp nDSM+RapidEye 72.18 % 1.72 pp 69.51 % 1.71 pp nDSM+RapidEye+I 73.38 % 1.71 pp 71.04 % 2.31 pp nDSM+SPOT+RapidEye 75.85 % 1.39 pp 73.26 % 2.19 pp nDSM+SPOT+RapidEye+I 76.70 % 1.98 pp 74.49 % 2.42 pp

airborne images, no LiDAR data

RGB+CIR+SWIR 62.32 % 1.86 pp 58.35 % 1.79 pp RGB+CIR+RE 64.74 % 1.67 pp 59.40 % 1.64 pp RGB+CIR+SWIR+RE 65.81 % 1.78 pp 61.94 % 1.76 pp RGB+CIR+SPOT 66.68 % 1.70 pp 63.75 % 1.16 pp RGB+CIR+RapidEye 74.88 % 1.01 pp 72.89 % 1.53 pp RGB+CIR+SPOT+RapidEye 76.34 % 1.44 pp 74.49 % 1.48 pp

only satellite data

RapidEye+SPOT 75.82 % 1.51 pp 73.92 % 2.13 pp

The first part of table 6.12 shows that the RapidEye data set yields substantially higher user’s and overall accuracies than the other data set. Without the red edge (RE) band, it achieves a higher overall accuracy than the airborne color and color infrared data set (RGB+CIR) but has a lower user’s accuracy. When using only the color or the color infrared images instead of both, the user’s accuracy declines further and lies far beneath the accuracy achieved on the 10 m and 20 m resolution SPOT data set. The classification on the individual airborne images does not achieve 50 % accuracy for either of the two accuracies. For the SPOT satellite data, the overall accuracy is slightly below 50 % and the user’s accuracy is a bit higher than 50 %. Adding the LIDAR nDSM leads to an improvement in overall classification accuracy of approximately 5.8 percentage points for the airborne images, 3.4 percentage points for the SPOT data set and 5 percentage points for the RapidEye data set. The single bands that were extracted from the SPOT and RapidEye satellite data sources and added individually yield less increase in classification accuracy than adding all available satellite bands. This indicates that, it is not only the additional spectral region that improves the classification accuracy for the combined data sources. The

additional R, G and NIR bands that were recorded at a different time and with another sensor, also improve the classification accuracy. The information gain by using the LIDAR I band is lower than the gain by using the SPOT short wavelength infrared (SWIR) band, although they are recorded at similar spectral regions. One possible explanation thereof is the previously described low quality of the I band due to variations and discontinuous coverage. The highest accuracy is achieved for the combination of all available bands, which leads to an overall classification accuracy of 76.54 % and a mean user’s accuracy of 78.01 %. The overall accuracy achieved on the combination of both satellite data sets with the two LIDAR data sets is approximately equal to the overall accuracy achieved on the two satellite data sets combined with the two airborne image sets. Both lead to an overall accuracy of 74.5 % but the user’s accuracy of the data set containing the LIDAR data sources is slightly higher. It is also interesting to note that the two satellite data sets combined yield a mean overall accuracy of approximately 74 %, which is only 1.62 percentage points below the highest achieved mean overall accuracy. This combination is especially interesting, if the airborne data sets are not available and not needed for additional calculations, as they are much more expensive to obtain than satellite data sets. Therefore, it could be a viable solution to use only two satellite data sets for tree species classification. However, at this point no information can be given concerning which satellite combination performs best — whether the combination of one RapidEye and one SPOT satellite is better than e.g. two RapidEye data sets at two different times within one growing period.

The comparison of accuracies achieved on different data sources is especially interesting for the planning phase of future projects. The approach needs to be affordable and therefore, the data sources need to be affordable. It is also important to weight the data source costs and the expected gain in classification accuracy. The costs of the data sources are very different and a very rough approximation of the prices in January 2012 is shown in table 6.13. The costs are given for an area of 3000 km2, as it is hard to estimate realistic costs on smaller areas. For one, RapidEye

has a minimum order size of 500 km2 as shown in [198]. On the other hand, SPOT

images are available in four different tile sizes with a fixed price for each size as given in [199]. It is apparent, that the airborne data sources are far more expensive than satellite based data sources. If the airborne data sources cannot be obtained from a regional surveyor’s offices, an aerial survey may need to be commissioned.

Table 6.13: Comparison of input data source costs Data Source Costs∗ per 3000 km2

airborne RGB + CIR + nDSM 507 000 e†

SPOT 3 300 e

RapidEye 2 850 e

For many applications it may therefore be worth considering only satellite data. It is worth noting that some satellite data providers offer a discount for more than one data set of a specific area within a limited amount of time (e.g. 6 months), thereby allowing the customers to obtain a multitemporal data set at a reduced cost.