In keltoskytoi/detectATE: Detecting Alpine Treeline Ecotones - an Automated Remote Sensing Approach

Discussion

The points examined during the development of this proposed combined workflow were detailed in Chapter 2 - Overview and they have been touched upon in Chapter 3 - Methodology and Chapter 4 - Results. Here they shall be discussed in depth.

The aim of the 'Segmentation' was to achieve a good 'Segmentation' of only the trees and it was tried to include young trees - hence the experiments with h = 3.5 m. A minimum height had to be defined for the 'Segmentation' of trees, which is region, area and species dependent. The ATE is situated below the Alpine Tundra Domain and it is very common, that here trees occur with krummholz/shurbs. This poses a problem in respect of how to distinguish young trees from krummholz/shrubs which might have the same height. Because the aim was to accurately segment trees, it was chosen to sacrifice young trees - even the h = 4 m did not include all young trees. Such a distinction cannot be made based on height, only based on spectral signatures of the tree species (that is color).

The questions if and how different training shapes influence the result of the 'Classification' and if and how they influence the selection process of 'Forward Feature Selection' with IKARUS::BestPredFFS are connected to the question if and how different spectral raster stacks influence the choice of 'Forward Feature Selection' with IKARUS::BestPredFFS and have to be answered more in length and depth. \ If we think about the selection process of the predictors for 'Classification' in general, then the first selection steps are merely statistical analyses - whether the Indices and RGB and IR raster show homogeneity (this step was implemented to exclude the indices VARI, RI and HI) and if the resulting not homogeneous scpectral raster correlate between each other. \ In the case of test area 1 the homogeneity threshold was set to THvalue = 0.9 and it indeed dropped the indices VARI, RI and HI. \ In the case of the ROI the indices VARI, RI and HI presented similarly ample homogeneous areas and the THvalue had to be raised to 0.93, otherwise with 0.9 and 0.92 VVI too was dropped and with 0.95 only VARI and RI were dropped.\ It is generally interesting to see, that with the addition of four spectral raster (R, G, B & IR) the composition of the remaining raster stack after the test on correlation changed. \ In the case test area 1 the differences are: \ In the raster stack clean_ALL_ind_stack_1_homog09_filt: GLI_modal3, GLI_sobel_v3, MSR_min3 and MSR_sobel_v3 is different. \ In the raster stack clean_ALL_RGBMS_stack_1_homog09_filt: GLI_max3, NGRDI_sobel3, NGRDI_sobel_v3, Blue_min3 and Blue_sobel_v3 is different. \ In the case of the ROI the difference is smaller: \ clean_ALL_ind_stack_ROI_homog093_filt: MSR_min3 and TDVI_sobel3 \ clean_ALL_RGBMS_stack_ROI_homog093_filt: Blue_min3 and IR_ROI

When comparing the test area 1 and the ROI the differing spectral raster are to 50% different between the two data levels (small are vd. big area). So far the selection of the spectral raster was only on the basis of cell statistics. It is visible, that adding spectral raster changes the result of the test on correlation.

Before creating the actual original predictor stacks the correlated raster are combined with the training shape files: with train_1 accurate, 3x3 pixel training polygons (classes 1-tree, 4-tree_in shadow and 5-shadow 10 instances each, classes 3-grass and 2-shrubs present 5 instances each and class 6-stone presents 3 instances) are used and with train_2 minimal number of arbitrary sized and shaped coarse training polygons (1 polygon/class).

The input to the 'Forward Feature Selection' is a data frame, which consists of values extracted from the two spectral raster stacks using the two training shape files. Each polygon in the shape files (and the information extracted using this polygon) supposed to be of a certain class. Because the size of the polygons in the two shape files is very different, the actual class of the information extracted by using the different shape files will be different - the original class if the pixels extracted by train_2 will be inhomogeneous. This is important to understand when interpreting the results and looking for answers to the question asked at the end of the Chapter 2 - Overview.

Comparing the results of the 'Forward Feature Selection', FFS_1_ROI - FFS_4_ROI with FFS_1 - FFS_4 we can see the following investigating the predictors (order based on frequency) ((Table 8)):

| FFS |PREDICTOR1|PREDICTOR2|PREDICTOR3| PREDICTOR4 | PREDICTOR5 | |----------:|---------:|---------:|---------:|-----------:|-----------:| |FFS_1 |TGI_max3|VVI_min3| | | | |FFS_1_ROI|TGI_max3|VVI_min3| | | | |FFS_2 |TGI_max3|VVI_min3| | | | |FFS_2_ROI|TGI_max3| |GLI_min3|VVI_sobel3|VVI_modal3| |FFS_3 |TGI_max3| | | | | |FFS_3_ROI|TGI_max3| |GLI_min3|VVI_sobel3|VVI_modal3| |FFS_4 |TGI_max3| | | | | |FFS_4_ROI|TGI_max3|VVI_min3|GLI_min3| | |

| FFS |PREDICTOR6 |PREDICTOR7|PREDICTOR8 |PREDICTOR9|PREDICTOR10| |----------:|----------:|---------:|----------:|---------:|----------:| |FFS1 | |MSR_min3| | | | |FFS_1_ROI| |MSR_min3| | | | |FFS_2 | | |Blue_min3|GLI_max3| | |FFS_2_ROI| | | | |SI_max3 | |FFS_3 |NDVI_min3| | | | | |FFS_3_ROI| | | | | | |FFS_4 |NDVI_min3| | | | | |FFS_4_ROI| | |Blue_min3| | |

Table: The frequency distribution of the 10 predictors selected by 'Forward Feature Selection' in all eight predictor stacks of test area 1 and the ROI.

10 different index predictors were chosen by 'Forward Feature Selection' from the correlated indices in general for 'Classification', out of the 18/19 of test area 1 and the 17 of the ROI. In the case of test area 1 this means five different index predictors, compared to the ROI where eight different ones were found meaningful. FFS2_ROI and FSS3_ROI consist of almost the same predictors save SI_max3 in favor of the latter.

FFS_1 and FFS2 were created using the accurate, 3x3 pixel training polygons of the training shape file train_1, while FFS_3 and FFS_4 were created using the training shape file train_2, which consists of minimal number of arbitrary sized and shaped coarse training polygons. If we group the predictors according to the training shapes of which the were created, we can see the following ((Table 9)):

| PREDICTORS | train_1 (FFS1&2) | train_2 (FFS3&4) | |-----------:|--------------------:|--------------------:| |TGI_max3 |ta1, ta2 & ROI1, ROI2|ta3, ta4 & ROI3, ROI4|
|VVI_min3 |ta1, ta2 & ROI1 | ROI4|
|GLI_min3 | ROI2| ROI3, ROI4|
|VVI_sobel3| ROI2| ROI3 |
|VVI_modal3| ROI2| ROI3 |
|NDVI_min3 | | ROI3, ROI4| |MSR_min3 |ta1, ROI1 | |
|Blue_min3 | ta2 | ROI4| |GLI_max3 | ta2 | | |SI_max3 | ROI2| |

Table: The occurrence of the 10 predictors selected by 'Forward Feature Selection' in the predictor stacks of test area 1 and the ROI.

It is safe to say, that TGI_max3 contains in each four combinations meaningful information and thus seems robust enough for the Classification. Predictors GLI_max3, SI_max3 and MSR_min3 contain only in combination with train_1 meaningful information to build a model for 'Classification'. In the case of train_2 this predictor is NDVI_min3. VVI_min3 is more often in combination of training shape train_1 and GLI_min3 is more often in combination of train_2. VVI_sobel3, VVI_modal3 and Blue_min3 occur equally in combination with training shapes train_1 and train_2.

Looking to the relation of the areas to the predictor indices we can see that four out of the 10 predictors chosen by 'Forward Feature Selection' are used in the prediction of both areas and six only in one of them ((Table 10)):

| AREA | PREDICTOR1 | PREDICTOR2 | PREDICTOR7 | PREDICTOR8 | |----------:|--------------:|--------------------:|----------------:|---------------:| |test area 1|TGI_max3(4x) |VVI_min3(FFS1&2) |Blue_min3(FFS2)|MSR_min3(FFS1)|
|ROI |TGI_max3(4x) |VVI_min3(FFS1&4) |Blue_min3(FFS4)|MSR_min3(FFS4)|

| AREA | PREDICTOR3 | PREDICTOR4 | PREDICTOR5 | |----------:|-------------------:|-------------------:|-------------------:| |test area 1| | | |
|ROI |GLI_min3(FSS2,3&4)|VVI_sobel3(FSS2&3)|VVI_modal3(FFS2&3)|

| AREA | PREDICTOR6 | PREDICTOR9 | PREDICTOR10 | |----------:|------------------:|---------------:|--------------:| |test area 1|NDVI_min3(FFS3&4)|GLI_max3(FFS1)| | |ROI | | |SI_max3(FFS2)|

Table: Tie-in of the predictors selected by 'Forward Feature Selection' in the predictor stacks of the test area 1 and the ROI.

It can be observed that GLI_min3, VVI_sobel3, VVI_modal3 and SI_max3 are only in index predictor stacks for the ROI and NDVI_min3 and GLI_max3 only for those of test area 1. TGI_max3,VVI_min3, Blue_min3 and MSR_min3 are associated with both areas.

As a last step let's look into the models on which the 'Classifications' are based, to understand which predictors chosen by the 'Forward Feature Selection' have contributed the most to the 'Classification' of the respective classes in the different 'Classifications' (plot(varImp(pred_1$model_cv)) was used for this) ((Table 11)):

|test area 1| grass | shadow | shrubs | stone | tree | tree in shadow| |--------------:|-------------:|--------------:|-------------:|-------------:|--------------:|--------------:| |PRED1 |MSR_min3|MSR_min3 |TGI_max3|MSR_min3| MSR_min3| MSR_min3| |PRED2 | BLUE_min3 | GLI_max3 |TGI_max3| BLUE_min3 | GLI_max3 | GLI_max3 | |PRED3 | NDVI_min3 |TGI_max3 |TGI_max3| TGI_max3 |TGI_max3 |TGI_max3 | |PRED4 | NDVI_min3 | TGI_max3 | TGI_max3 | TGI_max3 |TGI_max3 |TGI_max3 |

| ROI | grass | shadow | shrubs | stone | tree | tree in shadow| |---------:|-------------:|-------------:|-------------:|-------------:|-------------:|--------------:| |PRED1 |MSR_min3|MSR_min3|TGI_max3|MSR_min3|MSR_min3| MSR_min3| |PRED2 | VVI_modal3 | GLI_min3 |TGI_max3| GLI_min3 | GLI_min3 | GLI_min3 | |PRED3 |TGI_max3 |TGI_max3|TGI_max3| GLI_max3 |TGI_max3|TGI_max3 | |PRED4 |TGI_max3 | BLUE_min3 | GLI_min3 | VVI_min3 |TGI_max3|TGI_max3 |

Table: The predictors which contributed the most to the classification of the respective classes in the test area 1 and the ROI. The predictors which were used in the respective predictions of the two different area sizes are bold.

The first thing that visible is, that the predictors used for Prediction 1 for both the test area 1 and the ROI are the same. The predictors used for Prediction 3 and Prediction 4 are also the same, but this is due to the fact that the training shape train_2 was used in the generation of the predictors stack and because of it's traits only the indices TGI_max3 and NDVI_min3contained meaningful information for model building and 'Classification' (see also Chapter 4 - Results). \ Comparing the predictors which have contributed the most to the detection of a certain class chosen by the 'Forward Feature Selection' in the respective predictions in both area sizes we can say, that TGI_max3 and MSR_min3 are those predictors, which are the most frequently used for the 'Classification' of pixel classes in general. It can be also be concluded, that MSR_min3 works better with more accurate training shapes. In case that the original raster stack before the correlation contains also R, G and B bands, BLUE_min3 and GLI_max3 are chosen in the case of test area 1, and VVI_modal3 and GLI_min3 in the case of ROI, based on the correlation of the pixel values before even given over to the 'Forward Feature Selection'. \ This means, that the choice of the most meaningful predictor for the 'Classification' of the respective classes of the prediction depends on choices on multiple levels. It depends on the composition of the original raster stack and the possible correlation of the different raster as well as the nature of the training shapes used for the generation of the training data set. On this basis will the 'Forward Feature Selection' choose the most meaningful predictors for model building.

Based on the detailed examination of the predictor stacks of test area 1 and the ROI, conclusively it can be said, that different training shapes and different input spectral raster effect the 'Forward Feature Selection', based on the explanations detailed above. This also results in the fact that different training shapes influence 'Classification'.

keltoskytoi/detectATE documentation built on June 7, 2022, 11:41 p.m.