NEON NIST data science evaluation challenge: methods and results of team FEM

An international data science challenge, called NEON NIST data science evaluation, was set up in autumn 2017 with the goal to improve the use of remote sensing data in ecological applications. The competition was divided into three tasks: 1) segmentation of tree crowns; 2) data alignment; and 3) tree species classification. In this paper the methods and results of team FEM in the NEON NIST data science evaluation challenge are presented. The individual tree crown (ITC) segmentation (Task 1 of the challenge) was done using a region growing method applied to a near-infrared band of the hyperspectral images. The optimization of the parameters of the segmentation algorithm was done in a supervised way on the basis of the Jaccard score using the training set provided by the organizers. The alignment (Task 2) between the segmented ITCs and the ground measured trees was done using an Euclidean distance among the position, the height, and the crown radius of the ITCs and the ground trees. The classification (Task 3) was performed using a Support Vector Machine classifier applied to a selection of the hyperspectral bands. The selection of the bands was done using a Sequential Forward Floating Selection method and the Jeffries Matusita distance. The results in the three tasks were very promising: team FEM ranked first in Task 1 and 2, and second in Task 3. The segmentation results showed that the proposed approach segmented both small and large crowns. The alignment was correctly done for all the test samples. The classification results were good, even if the accuracy was biased towards the most represented species.

where , and ; ℎ ℎ ∈ (0;1) > 0 c. this procedure was iterated over all pixels that have , and was repeated until , ≠ 0 95 no pixels were added to any region; 96 6. from each region in the central coordinates of each pixel were extracted, and a 2D convex 97 hull was applied to these points; 98 7. the resulting polygons were the final ITCs.

99
The raster image used in this paper was the hyperspectral band at 810 nm, already used in

117
The height of the ITCs, for which this attribute was missing, was predicted using a 118 relationship linking the ITCs height ( ) and the ITCs crown radius ( ): = × (6) 119 Eqn. 6 was fitted using the function nls of the package stats of the R software (R Development 120 Core Team, 2008).

121
Each ITC was linked to the closest ground measured tree according to the Euclidean 122 distance between their position and their attributes (height, and crown radius): After the linking, a visual inspection of the results on a GIS software was done and some 124 trees were manually realigned.
125 Task 3: classification 126 The classification of the tree species was done with a four step procedure: 1) data normalization; 127 2) feature selection; 3) classification; and 4) aggregation.

128
Data normalization was done to ensure that the pixel values were uniformly distributed 129 across all the crowns. Each pixel value was divided by the sum of the values of that pixel in all 130 the bands (Yu et al., 1999). In this way, we reduced the difference in radiance due to the fact that 132 The feature selection is necessary in order to select only the bands that are useful to 133 separate the analysed species. A feature selection method is made up of a searching strategy and 134 a separability criterion. In this study, the search strategy we used was the Sequential Forward 135 Floating Selection (SFFS) (Pudil, Novovičová & Kittler, 1994), and the separability criterion The Jaccard score for the delineated ITCs over all the plots was 0.3402. The overall 148 confusion matrix (OCM) is showed in Table 1. To analyze the performance over each plot, the 149 confusion matrix for each plot was visualized as a bar chart (see Figure 1). The Jaccard score by 150 crown area is shown in Figure 2. Variability in the crown size did not change the Jaccard score, 151 showing that the method used is behaving in the same way for all crown sizes. The top-6 best 153 4.2 Task 2: alignment 154 All the test ITCs were aligned with the respective ground measured tree.  (Table 3) it can be seen that for some classes the performance metrics are really low 161 (e.g. ACRU), while others are really good (e.g. PIPA).

163
Team FEM ranked first for Task 1. As explained in the methods, we chose to segment a 164 hyperspectral band instead of the LiDAR point cloud. This choice was motivated by the fact that 165 looking at the training ITCs provided by the organizers, the hyperspectral data seemed more 166 suitable for this task. The comparison of results across teams showed that the FEM approach 167 outperforms the other approaches in the delineation of the small trees, while it was less efficient 168 for the large trees. This is due to the fact that we decided to use a small moving window (3x3). 169 The use of a variable size moving window, like the one that is implemented for LiDAR data in In Task 2 FEM team ranked again in the first place with all the trees correctly aligned.
178 Surely the choice to consider not only the position, but also the tree characteristics (i.e. height, 179 and crown radius) was the winning choice. Moreover, after the automatic matching a visual 180 inspection of the results helped make the final improvements, as two trees were reassigned after 181 this inspection. A visual inspection of the alignment is not doable over large datasets, even if, in 182 our experience, it is always suggested as it helps in finding macroscopic errors. As mentioned in 183 the introduction, the choice of alignment strategy can depend also on the type of data that can be 184 used for this purpose. The fact that each crown delineation paper uses a different alignment 185 method specific to the dataset is not a good approach. Indeed, there is the need to have a 186 reference alignment method that could be used in every crown segmentation paper that allows a 187 fair comparison among delineation results.

188
The classification task (Task 3) had the most participants and team FEM ranked at the 189 second place. In this case the architecture that we used was effective, even if the results showed a 190 serious problem in distinguishing minority species. This is a limitation of many other works 191 proposed in the literature as many classifiers tend to give priority to highly represented species.