Tree Recognition Process Based on ArcGIS Pro Deep Learning Module
Data preparation: aerial photography oblique visible light image data
Environment construction: download ArcGIS Pro, register an account (student user), use it to log in (login needs to be done in a networked environment),
install ArcGIS Pro, load the crack file, and install the Chinese language pack , clone the python environment (using the first step, an environment called deeplearning will be installed)
first step:
Install the deep learning framework.
Open the Python Command Prompt in the installation directory and enter the following command:
a.conda create --name deeplearning --clone arcgispro-py3
b. activate deeplearning
c. conda install tensorflow-gpu=1.14.0
d. conda install keras-gpu=2.2.4
e. conda install scikit-image=0.15.0
f. conda install Pillow=6.1.0
g. conda install fastai=1.0.54
h. conda install pytorch=1.1.0
i. conda install libtiff=4.0.10 --no-deps
j. proswap deeplearning
The installation is complete
Step 2:
Open ArcGIS Pro, click New Project, select the map for the project template, enter the name of the project and the folder where the project is located, and create a project file
third step:
Click Add Data to load the required data into the map window. The
map displays the results:
the fourth step:
Since the loaded image lacks projection information, the projected coordinate system is defined in this step. Click Geoprocessing, search for Define Projection, select the layer you want to add projection to, and select the appropriate projected coordinate system.
If the selected dataset layer is prompted, it indicates that there is a coordinate system, and no need to add a coordinate system.
the fifth step:
Select the layer, click on the Image toolbar, the classification tool in Image Classification, select the labeling object for deep learning, create a classification scheme, modify the scheme name, and add a new class.
select hand drawn
Step 6:
*** The most critical step is to draw the sample data strictly according to the tree boundary, otherwise the classification result will be less accurate. The number of samples should not be less than 50. The selection principle is: uniform selection of the whole width, representative, multiple types of trees should be selected, and the boundaries should be distinguishable *** The
selection results are saved
Step 7:
Click Geoprocessing, import and export training data for deep learning, select the input raster, Shuxi Lake folder, select the previous step to classify the sample data, select the value field, and the metadata format PASCAL visualization object class. Click Run.
Step 8:
Click Geoprocessing, input the training deep learning model, select the raster layer, set the output folder, and select the model definition emd file generated in the previous step.
operation result:
Step 9:
Click Geoprocessing, enter features to points, and run.
Running result:
Final extraction result:
summary:
(In this experiment, only 22 samples were selected, so the accuracy is low, and the accuracy is improved compared with the first experiment. The key point is the sample drawing in the sixth step, and there are many misidentified data. I hope readers will study it carefully. , which can improve accuracy)