Home

Awesome

DSAIL-TreeVision: A software tool for extracting tree biophysical parameters of trees from stereoscopic images

Background

Forest cover reduction poses a threat to environmental and ecological sustainability in Kenya. Various efforts to combat it, including reafforestation, face numerous challenges including prominently, lack of efficient, quick and rapid monitoring to inform timely policy changes by concerned stakeholders. At Dedan Kimathi University of Technology's Centre for Data Science and Artificial Intelligence (DSAIL), we aim to explore various technologies that can be used to improve forest monitoring.

About DSAIL-TreeVision

Stereoscopic vision is a computer vision technique that can be used retrieve the 3D information of a scene based on the concept of Multiple-View Geometry. DSAIL-TreeVision was developed to extend the principles of Multiple-View Geometry to facilitate fast and accurate estimation of tree attributes.

DSAIL-TreeVision can accurately estimate the values of the diameter at breast height (DBH), crown diameter (CD), and tree height (CD). The DBH is usually measured at 1.3 m above the trunk base and the algorithms presented in DSAIL-TreeVision can estimate this location with impressive accuracy. The algorithms for estimating these parameters are found inside the Controller/algorithms.py file

DSAIL-TreeVision is built using the Kivy and KivyMD Python packages. You can learn how to use it by reading the user guide

The User Interface

DSAIL-TreeVision comprises three modules i.e., the Extract module for extracting tree biophysical parameters, the Calibrate module for calibrating both single and stereo cameras, and the Capture module for capturing single and stereoscopic images. Each of these modules are accessed in different screens built in to the user interface. The Extract module is the default module displayed when the software is first run.

1. Extract Module Interface

Below is a screenshot of the Extraction interface. The UI conntains widgets for displaying the left and right images, configuration settings that the user can tweak as desired, and a section for displaying logs such as extraction progress and values of the extracted parameters.

extraction interface

This module begins by deriving depth maps from the stereoscopic images using the semi-global block matching (SGBM) technique, OpenCV's implmentation of Hirschmuller's algorithm. The rest of the algorithms for extracting the individual tree attributes were developed at DSAIL DeKUT and can be found inside the Controller/algorithms.py module.

2. Calibrate Module Interface

Below is a screenshot of the Calibration interface. The UI conntains widgets for displaying the calibration images, calibration settings, and logs. The calibration performed by this module is implemented using OpenCV's calibration technique which uses a checkherboard pattern. The method is based on Zhang's algorithm.

calibration interface

3. Capture Module Interface

Below is a screenshot of the Image acquisition interface. The UI conntains widgets for displaying the left and right and camera streams, or single camera streams. To capture an image, there is a capture button on the right hand side of the screen. Both camera streams are shown in stereo mode while a single stream is shown in single camera mode.

Capture interface

One can switch between the stereo and single camera modes using the tab switch buttons at the bottom of the screen. In stereo mode, the user can the swap camera streams (make the right camera left, and the left right) using the camera swap button on the right hand side.

Installing Python

Skip this section if you already have python installed.

Windows

  1. Download the Miniconda3 setup and double click on it and follow the prompts to install. In the installation prompt, make sure to enable conda to be added to the PATH environmental variable as show in the screenshot below.

add path

  1. To confirm that Python was installed successfully, open command prompt and run the following command. The output should be the python version 3.8.
python --version
  1. If the installation is verified, you can delete the installation file by running the following command on command prompt.
del Miniconda3-py38_23.11.0-2-Windows-x86_64.exe

  1. Download the Git Bash setup and install it. Leave the defaults when following the installation prompts. Once installation is complete, proceed to launch Git Bash.

  2. To initialize Git Bash to recognize the conda command, running the following command.

mv ~/.bash_profile ~/.bashrc
conda init
  1. Close Git Bash and reopen it. Run the following command and confirm that there is no error raised.
conda activate

Linux

  1. Download the Miniconda3 installation file by running the following command in your terminal.
wget https://repo.anaconda.com/miniconda/Miniconda3-py38_23.11.0-2-Linux-x86_64.sh
  1. Without closing the terminal, run the following command and follow the prompts to install Miniconda3.
bash Miniconda3-py38_23.11.0-2-Linux-x86_64.sh
  1. Confirm that python is installed by running this command in terminal. The output should be python version 3.8.
python --version
  1. Once Python is installed, you can delete the installation file by running the following command.
rm Miniconda3-py38_23.11.0-2-Linux-x86_64.sh

Other platforms

Installation instructions for Miniconda3 on all platforms can be found here.

Note on Image Segmentation

An important step in the process of extracting tree biophysical parameters from stereoscopic images is image segmentation. It is necessary in order to retain only the foreground pixels (full tree or tree trunks) in the depth maps. The algorithms implemented in DSAIL-TreeVision extract the tree parameters from the segmented depth map.

We have incorporated two deep learning segmentation models in this application: Meta's Segment Anything Model (SAM) and the PerceptTreeV1 Mask R-CNN model finetuned on real images.You can read about them here:

Down below, we have provided instructions on how to download and use their pretrained models in the application.

Setting Up DSAIL-TreeVision

Note for Windows Users: Install Git Bash before proceeding

  1. Open your your file manager/explorer and navigate to where you would like to save DSAIL-TreeVision files.

  2. Right click on this location and select Git Bash Here if you are on Windows, or open in terminal if you are on Linux. This will open Git Bash or Bash in this location.

  3. Now go ahead and clone this github repository. To do this, copy and paste the following command on Git Bash or Bash and pressing enter. Once the process is complete, the root directory will be saved as TreeVision

git clone https://github.com/DeKUT-DSAIL/DSAIL-TreeVision.git
  1. Download the pretrained segmentation models:

  1. Copy the models downloaded in step 4 above to the assets/models path found in the application's root directory.

  2. On command line, open the application's root folder (TreeVision) by running the following command:

cd TreeVision
  1. On command line, run the setup.sh script from the DSAIL-TreeVision's root directory, using the command:
bash setup.sh
  1. Activate the virtual environment by running this command:
conda activate treevision
  1. DSAIL-TreeVision runs inferences on pretrained image segmentation models which require PyTorch and Detectron2 packages. Follow the instructions for installing PyTorch on your platform and installing detectron2 on Linux or MacOS. Windows users see below.
  1. With the virutal environment still active, check that PyTorch and detectron2 were correctly installed by verifying that they import successfully.
python
import torch
print(torch.__version__)
import detectron2
print(detectron2.__version__)
  1. If you have commpleted steps 1-8 successfully, you can start DSAIL-TreeVision by running the start.sh script using the command:
source start.sh

Notes for Linux Users

A. Install xclip and xsel

Run these commands to install the 2 packages:

sudo apt-get install xclip xsel
B. on libGL errors

Running Kivy applications on Linux requires having the right libGL drivers. You might find that on executing step no.6 above, you run into libGL errors that make DSAIL-TreeVision fail to start. If this happens, it is because the libGL drivers sought by the system to run the application are missing, or a specific version of the driver is missing. The driver is always installed in the virtual environment during the setup steps, and this setup process sometimes installs the wrong version of the driver. Here are some helpful links to help you go about troubleshooting and solving the problem:

  1. libGL error: failed to load drivers iris and swrast in Ubuntu 20.04
  2. Kivy issues with libGL errors
  3. libGL error: MESA-LOADER: failed to open iris

Test Usage: Parameter Extraction

A folder called test is included in DSAIL-TreeVision's root directory. It contains a full_trees folder of 20 image pairs of full trees from which you can extract CDs and THs, and a trunks folder of 10 image pairs of tree trunks and their masks from which you can extract the DBHs.

By default, the file manager in DSAIL-TreeVision opens in the host system's home directory. On Windows, this will be the C:\Users\<USER> directory (replace <USER> with the username you used in your PC). On Linux, this will be /home directory. By contrast, the test images you will be using in this test phase are found inside the root directory of the application. You will need to navigate form your home directory to application's root directory to select your files or folders. You can perform this navigation by clicking on the back button on the top left of the file manager user interface.

To test tree parameter extraction on DSAIL-TreeVision, follow these steps:

  1. On the user interface, click on the green Select project images button to open the file manager. Navigate to the TreeVision/test/full_trees folder and select this folder by pressing the checkmark (tick) button on the bottom right of the screen. This folder has now been selected as the location of the images to be processed.

  2. Click on the Calibrate File (YML) button and navigate to the TreeVision/configs/test folder and click on the stereo_full_trees.yml file to select it. This file has now been selected as camera calibration to be used in processing the images uploaded in step 1.

  3. Click on the Select ref values file button and the scroll down until you find the TreeVision/test_cd_th.csv file and click on it. This file has now been selected as the file containing the ground truth values of crown diameter (CD) and tree height (TH) for the trees in the images selected in step 1.

  4. Click on the orange Verify button. If you followed steps 1-3 correctly and did not modify any of the default settings on the interface, the Extract and Batch Extract buttons adjacent to the Verify button you clicked on should have turned green.

  5. Click on the Batch Extract button and observe the bottom right (we call this the logging section) of the user interface. You should notice that some information is being added continuously. Further, the images displayed on main image widgets on the screen are changing frequently. This is because DSAIL-TreeVision is extracting the crown diameters and tree heights of all the trees in the folder you selected in step 1. Once the parameters of one tree are estimated, the software moves to the next tree and displays a new pair of images.

  6. Step 5 is complete is complete when you see the message Batch extraction complete on the logging section. Notice right image of the pair displayed on the screen. It has annotations added to it, like the one in the image shown below. These annotations are arrows showing the extents and measurements of the tree height and crown diameter.

Annotated image

  1. Notice also that the Analyse button adjacent to the Batch Extract button has turned orange. Click on it and wait for about 5 seconds. You should now see two plots displayed where the original images of trees were displayed. In the logging section, a statistical summary of the analysis performed is also shown. It is a comparison of the estimated values with the ground truth values provided in step 3.

Congratulations !!!

You have successfully extracted tree parameters of 20 trees using DSAIL-TreeVision.