Field Classification Research

Related Papers and Presentations

Applying Convolutional Neural Networks to Per-pixel Orthoimagery Land Use Classification (2017)

Presented at the Midwest Instruction and Computing Symposium (MICS) 2017

• Paper
• Presentation
• Accolades: Placed 1st in the Undergraduate Research Paper Competition

Exploring the Usefulness of Adding Auxiliary Preprocessed Image Layers With Convolutional

Presented at the Midwest Instruction and Computing Symposium (MICS) 2018

• Paper
• Presentation

0. Requirements

To prepare the image dataset and train the model you will need the following requirements:

• Docker - used for easily installing a reproducable environment by using container technology
• nvidia-docker - Allows docker containers to access the nvidia cuda API.
• imageinfo - used for getting the dimensions of images while converting from .tif to .png

1. Preparing Dataset

Dataset Download instructions:

1. Navigate to Earth Explorer
2. Create an account and log in. Account creation is required in order to download images above a certian size.
3. TODO finish download instructions

2. Reprojecting

1. Copy all image zip files into unprocessed_images/ and cd into unprocessed_images
2. Unzip all files with unzip \*.(ZIP|zip)
3. Run ./generate_images.sh. This will reproject the downloaded images.
4. Edit ./generate_nlcd_layer.sh to reflect where you downloaded nlcd_2011_landcover_2011_edition_2014_10_10.img to.
5. Run ./generate_nlcd_layer.sh This will use the rerojected images to cut and reproject a file containing label data for a corresponding region.
6. Ensure FieldClassification/images exists. If it does not create it wil mkdir ../images (this command is reletive to unprocessed_images/).
7. Run ./copy_images.sh to copy the generated PNG files to ../images

3. Cut images into swatches

1. Run mkdir raw_images
2. ./docker_run.sh "python generate_dataset.py"

4. Calculate Test/Train Datasets

Calculates test/train datasets and stores their corresponding id's in a .npy file (numpy file). This makes it easy to run multiple versions of the model on the same test and training sets.

1. ./docker_run.sh "python calc_splits.py"

5. Build the Docker Container

This step should only need to be done once initially, and then periodically to pull updated versions of the Tensorflow container.

1. ./docker_build.sh - this may take a while.

6. Run the Model

1. Edit the second line of ./docker_tensorflow.sh to set training parameters.
2. Run ./docker_tensorflow.sh - This will take a very long while as the model runs. While running, move on to the next step in a new terminal to view graphs.

7. View Model Progress/Statistics

1. Edit the absolute path to FieldClassification/ on the first line of ./docker_tensorboard.sh
2. Run ./docker_tensorboard.sh [path to summary dir] By default the summary dir is located in FieldClassification/output/summaries. To view one run instead of all runs of the model pass the path to the specific model summary dir, e.g. output/summaries/my_model_name_here/.