LiDAR for Individual Tree Detection

Approximate time invested (from research to implementation): 45 hours

 

This data science pet project aims to utilize LiDAR data to identify and count individual trees in the ANA_A01 region of the Amazonian rainforest. The Popescu Local Filtering with Fixed Window Size algorithm will be implemented using Python.

Pet Project Objectives:

1. "Hello world" for LiDAR data and forestry projects.
2. Learn LiDAR/Raster data handling and the use of coordinate systems.
3. Use GeoPandas and Rasterio for data manipulation

Data Description:

The data used for this project was provided by the Sustainable Landscapes GeoNetwork Catalog. The contained files are the following:

1. 2018 LiDAR data collection for ANA_A01 region.

   • Filename: ANAA01_L0002C0002.las / ANA_A01_2018_LAS_6.laz
   • Latitude range: 721917.4 - 721917.4
   • Longitude range: 9627871 - 9628871
   • CRS: EPSG-32721
   • Density: 28.85 points/units²

2. 2018 Tree inventory for ANA_A01 region (Ground Truth).

   • CSV file with individual information for the trees in inventory (years 2015 and 2018).
   • Variables of interest:
     • htot - Total height of trees in inventory.
     • RN, RS, RE, RW - Crown radius measured from the tree log to North, South, East and West.
     • 2015-18.dead - Trees labeled as dead in 2015 and 2018 inventory.
     • UTM. Easting & Northing - UTM Latitude and Longitude coordinates of each tree in inventory.

3. GeoJSON file containing the locations where tree data was gathered - The tree inventory was NOT conducted across the entire ANA_A01 region. Instead, data was collected from 32 sampled locations. The corresponding polygons are contained within this file.

 

Procedure Description (summary):

1. Academic literature research:

   As a self taught project, the first step was to conduct a thorough academic literature and software tools research on Individual Tree Detection (ITD) algorithms.

2. Data Understanding:

   • Inventory data: Count the number of inventoried trees and understand the recorded properties. Look for proportion correlations between height and crown width.
   • LAS data: Understand the contained information besides the datapoint cloud (metadata), and comprehend the coordinate system used.

3. Canopy height Model (CHM) generation with LidR:

   • Load .las and .shp files to clip the coordinates of the designated locations, with the function clip_roi.
   • Height normalization: Normalize the digital terrain height by applying the tin algorithm.
   • CHM generation: Apply the raster_canopy over the nomalized digital terrain model, using the Point2Raster algorithm.

4. Local Maxima with Fixed Window Size application:

   • Prepare data by loading the .tif file and clipped Gound Truth data for further evaluation.
   • Apply functions (listed here some just of the main):
     1. CHM smoothening with ndimage.median_filter
     2. Kernel generation with NumPy np.ogrid or np.ones
     3. Local maxima filter with ndimage.maximum_filter
   • Fine tune parameters, specifically window sizes for smoothening and filtering to obtain better results.

5. Individual Tree Detection Evaluation:

   • Try to evaluate LM performance by applying the GeoPandas .buffer() over Ground Truth trees, then using the .within() function to check if detected trees are inside the Ground Truth buffered area.
   • Manual evaluation: Plot the results of Detected Trees against Ground Truth Trees and manually evaluate the accuracy.

Relevant Inisghts & Considerations:

1. No linear correlation between tree height and crown width to be taken: Despite looking for a linear model fitted for a variable window size model, the Amazon Rainforest has such a great species variety that it was not possible to use the coefficients without oversimplifying, therefore facilitating misleading results.

   The lack of a strong correlation is portrayed in the next graph:

2. R made CHM's were generated with a 0.5 resolution, giving .tif file similar to the next one.

3. Within the ANA_A01 region there were multiple sampling locations for inventory data; still, for the purposes of this pet project, just 3 sampling areas were selected to work with, as shown below:

Results:

After applying the Local Maxima with Fixed Window Size method to detect trees across the three different Canopy Height Models (CHMs), the recorded accuracy results were not the best. This may be explained by the following factors:

1. In academic research, this approach tends to have an accuracy between 0.6 to 0.8. However, this is often applied in environments with a lower density of trees, which allows for a clearer distinction between local maxima points and the rest.
2. The high species diversity and morphology variety, combined with the high density of trees, makes it less accurate to distinguish individual trees using the Local Maxima approach.

Zone	Ground Truth # trees	LM correct trees	Overall Accuracy
8	130	33	25.38%
18	110	19	17.27%
22	84	20	23.81%

Nevertheless, here are the manually checked results.

Results for Zone #8:

33 trees correctly identified out of 130

Results for Zone #18:

19 trees correctly identified out of 110

Results for Zone #22:

20 trees correctly identified out of 84

Conclusions & Next Steps:

Applying the Local Maxima with Fixed Window Size to such a dense and biodiverse environment generates some serious challenges that, given the approach limitations, did not deliver good model performance.

In order to increase the individual tree detection accuracy in this Amazonian region, it will be necessary to implement a different approach. Given the different challenges posed by the extremely biodiverse and dense enviorment, the Layer Stacking approach, proposed by the Canadian Journal of Remote Sensing, could be an interesting next step.