Detection of thinning necessity

Detection of the necessity of thinnings We apply deep convolutional neural networks for semantic sementation to detect the need for thinning in the forest from remmote sensing data.

---

1. If you don't have it - Install conda

1. Check you don't already have conda installed!
   1. which conda
   2. if you already have it installed, skip ahead to Create an Environment
   3. It doesn't matter if you have miniconda3, or anaconda3 installed (it does not even matter if it is version 2).
2. If you don't have conda, download the latest version of miniconda3
   1. cd ~/Downloads (you can make a Downloads folder if you don't have one)
   2. Download the installer (we prefer to use miniconda since it carries less baggage), depending on your system (you can check links here):
      • Linux: wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
      • Mac: wget https://repo.continuum.io/miniconda/Miniconda3-latest-MacOSX-x86_64.sh or curl -LOk https://repo.continuum.io/miniconda/Miniconda3-latest-MacOSX-x86_64.sh
      • Or just simply download from the site
3. Install miniconda3 with default settings
   1. bash Miniconda3-latest-Linux-x86_64.sh
   2. Follow the prompt - type yes and hit enter to accept all default settings when asked
4. Close Terminal and reopen
5. Try executing conda -h. If it works, you can delete the installer rm ~/Downloads/Miniconda3-latest-Linux-x86_64.sh

3a. Create an environment for geo

1. Update conda: conda update conda
2. Create the environment. Call it geo and install python 3 (hence the name): conda create -n geo
3. Activate enviroment: conda activate geo
4. Install packages: conda install -c conda-forge gdal conda install -c conda-forge geopandas

• numpy
• h5py

1. Create Jupyter kernel: python -m ipykernel install --user --name geo
2. Deactivate enviroment: conda deactivate

3b. Create an environment for pytorch

1. Create the environment. Call it pytorch and install python 3 (hence the name): conda create -n pytorch
2. Install packages: conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch conda install pandas conda install h5py

---

4. create_dataset

data preperation and concatination into data set

4.1 Create input data tiles

1. • prepare_data_ortho.ipynb
2. • prepare_data_dsm.ipynb
3. • prepare_data_dtm.ipynb
4. • prepare_data_slope.ipynb

preperation of the data for concationation into one data set (see Create data set). in particular the raster files are clipt into tiles defined by the vector file (.shp). in the case of dsm, dtm and slope the created files need to be aigned since they have diffrent spatial resolutions.

4.2 Create ground truth tiles

1. create_ground_truth.ipynb

Created ground truth raster files (.tif) out of a vector file (.shp)

add ground truth types defines the ground truths for Base, UR12 and UR1 (see disertation).

4.3 Create data set

1. create_dataset_mult_gt.ipynb

data set creation (hdf5-file: .h5)

Creates the data set as a single hdf5 file. The data must by prepared in order to allow the script run smothly.

• set input data directory containing the preprocessed data
• inside this directoy the directories must have the exact names as defined in the dictionary ´data_dict
• all files in this dictonaries must have a specific prefix

directory: xxx, tile number: nnnnnn, file name: tile_xxx_nnnnnn.tif

directory: dtm, tile number: 122277, file name: tile_dtm_122277.tif

calculation of stats calculation of mean and standard deviation and addition to data set (.h5)

4.4 Create extended data set (fliped) (optinal)

1.dataset_flip.ipynb indices of train, validation and test sets must be prepared as numpy arrays (.npy)

---

5 deep_learing

the three folders contain the code for training and evaluation of the models. each of the folder has the same structure.

• deeplabv3plus - DeepLabv3+ (DCNN impemented by: https://github.com/jfzhang95/pytorch-deeplab-xception)

• fcdensenet - FC-FCDenseNet

• unet - UNet

• datasets - loading and preparing data for the DCNNs

• log - for logging the progress of the training

• models - contains the implementation of the DCNNs

• utils - additional code, e.g. saving weighs, visulising data, ...

• weights - for storing the trained weights (every epoch)

train.ipynb - for training models test.ipynb - for testing models

execution

it is just necessary to run train.ipynb for training and test.ipynb for testing

---

prediction

for prediction of the necessity of thinnings. creation of verctor (.shp) data out of the predicted raster tiles from the DCNNs.

the folder structure is the same as in # deep_learing

create predictions

predict_data_512.ipynb

creates predictions as raster tiles.

create vector data

predict_data_rasterize.ipynb

concatonates the tiles created by predict_data_512.ipynb and vectorises the data.