The C:\LearnArcGIS\DL folder is where you will store all the data and outputs for the lesson.

• smalldata.zip—Contains a small training dataset and the training boundary, a small validation dataset and the validation boundary, and a DEM raster.
• testdata.zip—Contains a test dataset, the processing boundary, and a DEM raster.
• results.zip—Contains the training results from the large dataset that you can use instead of training the model for the large dataset.

The folders contain various information that you will use for preparing and training data. Some of the main files in the folder are LAS files. A LAS file is an industry-standard binary format for storing airborne lidar data. The LAS dataset allows you to examine LAS files, in their native format, quickly and easily, providing detailed statistics and area coverage of the lidar data contained in the LAS files. The results folder contains a trained model using the large dataset that was performed on a computer with 24 GB GPU. In the final section of this lesson, you will use this trained model later in the lesson to classify a LAS dataset.

Next, you will review and familiarize yourself with the downloaded data in ArcGIS Pro.

You can see the data folders that you downloaded and extracted to this location.

First, you will explore the files in the smalldata folder and update the statistics on the dataset to ensure that the proper class codes display.

The dataset contains several class codes:

• 1- Unassigned represents unclassified points, mainly low vegetation or objects above the ground, but lower than high vegetation and buildings.
• 2- Ground represents the ground.
• 5- High Vegetation represents high vegetation, such as trees.
• 6- Building represents buildings and other structures.
• 7- Noise represents low points.
• 14- Wire-Conductor represents actual power wires.
• 15- Transmission Tower represents the power line towers.

You are interested in the class code 14 for Wire - Conductor. Your goal is to train a model to detect lidar points that are the power lines so you can better assess wildfire risks for power lines near trees.

At the bottom of the window, there is a message that states that one file has out-of-date or no statistics. You will fix this issue now so that you can display the correct class codes in the scene.

A message appears confirming that the statistics are up to date.

Now that the statistics are up to date, you can style the layer according to class codes.

The LAS dataset layer appears in the scene.

Now that you have updated the statistics for the small dataset LAS file, you will do the same for the LAS dataset that will be used to validate the model, val_small.lasd.

You will use train_small.lasd to train the model and val_small.lasd for validating the trained model to prevent overfitting during the training process.

Next, you will update the symbology for each of the LAS dataset layers to show different colors by the class codes. You'll start with the train_small.lasd layer.

The Symbology pane appears and the train_small.lasd layer is symbolized by its class codes.

The val_small.lasd layer is now symbolized by its class codes.

The Geoprocessing pane appears.

• For Input Point Cloud, choose train_small.lasd.
• For Training Boundary Features, click Browse, browse to C:\LearnArcGIS\DL\smalldata\boundaries.gdb, choose bnd_train_small, and click OK.
• For Validation Point Cloud, choose val_small.lasd.
• For Validation Boundary Features, click Browse, browse to C:\LearnArcGIS\DL\smalldata\boundaries.gdb, choose bnd_val_small, and click OK.
• For Reference Surface, click Browse and browse to C:\LearnArcGIS\DL\smalldata. Click dem_small.tif and click OK.

The DEM raster will be used to calculate points' relative height from the ground. You have entered all the training data, validation data, and processing boundaries.

Next, you will specify which class codes to exclude.

Ground (class 2) and noise (class 7) points are excluded from the training data. Ground points typically account for a large portion of the total points so excluding ground points will make the training process more quickly.

• For Block Size, type 82. For Unknown, click the drop-down menu and choose US Survey Feet.
• For Block Point Limit, keep the default of 8192.

Block Size and Block Point Limit control how many points are in one block. To determine this value, it is important to consider the average point spacing, the objects of interest, available dedicated GPU memories, and the batch size when setting these two parameters. A general rule is that the block size should be large enough to capture the objects of interest with the least amount of subsampling.

You will start with the Block Size of 82 feet (about 25 meters) and the Block Point Limit of 8192. 82 feet is a proper block size to capture the power line’s geometry. You will examine the output to see if 8,192 is an appropriate block point limit. If most blocks have more than 8,192 points, you will need to increase the block point limit to reduce subsampling.

A confirmation message appears at the bottom of the tool pane when the tool completes.

Under Messages, you can see two histograms of the block point count, one for the training data and one for the validation data.

Both histograms show that almost all the blocks have a point count of less than 8,000, which confirms that a Block Size of 82 feet and the default Block Point Limit of 8,192 are proper for these datasets.

The output file contains two subfolders, train and val, which contain the exported training and validation data, respectively.

In each folder, you will see a Statistics.json file, a ListTable.h5, a BlockPointCountHistogram.png file, and a 0 folder containing Data_x.h5 files.

The ListTable.h5 and Data_x.h5 files include the information of points (xyz and attributes such as return number, intensity, and so on), which are organized into blocks.

Now that you have prepared the training data, you can use the blocks to train a classification model with deep learning.

• For Input Training Data, click Browse. Browse to C:\LearnArcGIS\DL\results, and double-click training_data_small.pctd.
• For Pre-trained Model, leave it blank.
• For Model Architecture keep the default of RandLA-Net.

  Note:

  You will use RandLA-Net architecture to train the classification model. RandLA-Net uses random sampling to reduce memory usage and computational cost and use local feature aggregation to preserve useful features from a wide neighborhood.

• For Attribute Selection, click the check box for Intensity and Relative Height.

• For Output Model Location, browse to the results folder. Click it and click OK.
• For Output Model Name, type Powerline_classification_model_small_data.
• For Minimum Points Per Block, type 1000.

By setting the Minimum Points Per Block to 1000 during the training, blocks whose points are less than 1,000 will be skipped. Blocks with a small number of points are likely to be located at the boundaries where there are no power line points. Skipping these blocks in the training will make the training faster and make the model's learning more efficient. This parameter is only applied on the training data blocks, not the validation data blocks.

Class 14 represents the wire conductors that you want to classify and locate in the point cloud.

By specifying the Class Remapping, classification code 14 will remain unchanged, meaning points in the LAS dataset that already classify as power lines will remain power lines. All other class codes (1, 5, and 15) will be remapped to the class code of 1. The resulting training data will only have two classes, 1 and 14, thus making the power lines easier to distinguish in the scene.

Model Selection Criteria specifies the statistical basis that will be used to determine the final model. The default of Recall will select the model that achieves the best macro average of the recall for all class codes. Each class code's recall value is determined by the ratio of correctly classified points (true positives) over all the points that should have been classified with this value (expected positives). For each class, its recall value is the ratio of correctly predicted points of this class to all the reference points of this class in the validation data. For example, the recall value of class code 14 is the ratio of correctly predicted power line points to all the reference power line points in the validation data.

An epoch is the complete cycle of the entire training data learned by the neural network (in other words, the entire training data is passed forward and backward through the neural network one time). You will train the model for 10 epochs to save time.

Leaving the Iteration Per Epoch (%) parameter at 100 ensures that all the training data will be passed per epoch.

Batch Size specifies how many blocks are processed at a time. The training data is split into batches. For example, if the Batch Size value is set to 20, 1,000 blocks are split into 50 batches, and each of the 50 batches are processed in one epoch. By splitting the blocks into batches, the process will consume less GPU memory.

Unchecking this option will allow the training to run for 10 epochs. If checked, the training will stop when the model is no longer improving after several epochs, regardless of the maximum number of epochs specified.

The Train Point Cloud Classification Model (3D Analyst Tools) window appears, displaying the Parameters tab, which shows the parameters that were used for running the tool.

The tool reports the following information:

• GPU used in the training.
• Count of the training data blocks used in the training (only training data block containingmore than 1,000 points are used in the training).
• Count of the validation data blocks used in the validation—all the validation data blocks are used in the validation.
• Iteration per epoch—the count of the training data blocks divided by the batch size.

The tool reports Training Loss, Validation Loss, Accuracy, Precision, Recall, F1-Score, and Time spent for each epoch.

As each epoch progresses, you can see the Training Loss and Validation Loss values decrease, which indicates that the model is learning.

After 10 epochs, the highest Recall value is over .98.

This model is not an accurate model, which is expected because the model was trained with a small dataset. These results highlight the need to have more sample points in separate datasets so that you can achieve better results.

In this folder, there are two subfolders. One is the model folder, and another is the checkpoints folder.

The Powerline_classification_model_small_data folder includes the saved model at the epoch of 9. There are several files inside the model folder. Among them, the .pth file is the model file and the .emd file is Esri Model Definition file (a configuration file in .JSON format). The model_metrics.htm includes a learning loss graph. The .dlpk file is the deep learning model package. It is a compressed file and can be shared at ArcGIS Online.

The Model Characteristics folder contains a loss graph and ground truth and predictions results graph.

The Powerline_classification_model_small_data.checkpoints folder includes the models folder, which contains a model for each of the epoch checkpoints and two .csv files. When the training tool was running, one checkpoint was created after each epoch. Each checkpoint contains a .pth file and an .emd file.

The CSV file opens in Excel.

It is the same table that you saw in the tool's message.

The Powerline_classification_model_small_data_Statistics file includes the precision, recall, and f1-score values for each class after each epoch. In certain cases, the model with the best overall metrics may not be the model that performed the best in classifying a specific class code. If you are only interested in classifying certain class codes, you may consider using the checkpoint model associated with the best metrics for that class code. Viewing the statistics in Excel allows you to sort the columns and easily locate the epoch with the highest recall value. You have trained a model using a smaller sampling of points. Next, you will use a model that was trained using a larger sampling of points to classify a LAS dataset.

The test.lasd LAS dataset appears in a local scene.

The test.lasd LAS dataset is now rendered using class codes.

Ground points (class 2) and low noise points (class 7) are already classified. The left points are unassigned (class 1).

By selecting a Processing Boundary, the tool will use the trained model you provide to classify only the points within the boundary.

The Target Classification parameter appears after selecting a model definition file.

Leave the Batch Size blank and let the tool to calculate an optimal batch size for you based on the available GPU memory.

Since class code 2 and 7 are excluded from the classification. The model only classifies the left points (class 1). When the model runs, the points are predicted as power lines and will have their class codes assigned as 14; otherwise, their class codes are kept as 1.

The tool runs. Next, you will update the symbology to show the results of the classification model.

The symbology updates and the points in the point cloud that are wire conductors display in yellow.

The classification results are accurate using the trained model. Almost all the power line points are correctly classified. Notice that the power poles are still Unassigned.