Skip to content

Latest commit

 

History

History
93 lines (70 loc) · 7.15 KB

README.md

File metadata and controls

93 lines (70 loc) · 7.15 KB
Raw

Subjective Image Processing

The main goal of this project was to quantify the influence of various hyperspectral image processing steps on final model accuracy for predicting crop nitrogen uptake in maize.

This repository aims to track project scripts, methodology, and data over time by leveraging git version control tools. Anytime a "run" is processed on the Minnesota Supercomputer, a "release" is published in the repository indicating a particular version of the code used to generate that particular data.

The results from this analysis are organized such that the meta results files have "msi_X" in there filename, indicating the MSI run number. For example, msi_1_hs_settings.csv is a table of the image processing settings evaluated in MSI run #1. There is also a folder titled "msi_X_results" (e.g., msi_1_results) that contains a directory for each image processing scenario (e.g., "mis_1_results/msi_1_005" contains the results for MSI run #1, image processing scenario #5).

Setup and Installation

There is an environment.yml file that can be used to create the environment and install the dependencies. After cloning from Github, create the environment: Pip dependencies should be installed separately.

conda env create -n sip_run_2 -f sip/environment.yml
pip install git+git://github.com/maroba/findiff.git@master

Minnesota Supercomputer Institute (MSI)

The Minnnesota Supercoputer Institute high performance computing (HPC) and data storage resources were used to complete this project. Without access to resources like this, this project would not be feasible.

Run a job on MSI

Login to MSI via PuTTY

ssh mangi
cd ../public/hs_process
conda activate msi_sip_37
qsub <job_file.pbs>

More information at MSI Job Submission and Scheduling (PBS Scripts)

MSI notes

  • Total disk space required for the naive approach (process all files without deleting/transferring any of the data/results) is 12.1 TB. Thus, I try to split the processing into smaller chunks (say, 72 scenarios at a time), then allow all transfer to 2nd tier to complete before starting the next chunk.
  • The bulk of image/spec data files can be transferred automatically using transfer_data_level.py at the end of tune_train.py jobs. It is important to be sure data files are not transferred to 2nd tier until all tuning/training is completely finished for those processing scenarios.
  • After all MSI batch jobs of tune_train.py finish for a "chunk", transfer_data_level.py can be run at the "clip" level (for each "clip" scenario) to transfer and delete the spent data without doing so manually.

jobs

The jobs directory contains MSI shell scripts (can be read by PBS or Slurm schedulers) that carry out all image processing tasks, all model tuning, training, and testing, creation of figures/plots of final data, and transfer to 2nd tier storage and subsequent deletion from high performance storage.

Image processing

Before images were uploaded to MSI high performance directory, the following pre- and post-processing steps were performed:

  1. [pre-] Images were converted from raw (digital number) to radiance using calibration file provided by Resonon.
  2. [pre-] Images were georectified (with GPS and IMU data to project each image line to a spatial reference system)
  3. [pre-] Images were converted from radiance to reflectance using measured spectra from reference panesl and the radiance from the reference panels (from images).
  4. [post-] Images were cropped (either to plot boundaries or with an added negative buffer to remove boundary plants).

hs_process library is being used to achieve all image post-processing steps for this project (carried out by the process_img.py script). The specific version of hs-process used for a given MSI run ID should be indicated by the pip install commit hash ID in the Environment section.

Hyperspectral image naming convention

Hyperspectral reflectance images that have undergone the cropping step were uploaded to MSI high performance storage. These images are named with the following unique identifiers in their filename:

  • study name (aerfsmall, aerfwhole, or wells)
  • acquisition date
  • plot_id

Thus, there is a hyperspectral image for every study-date-plot combination. The ground truth observations that exist for any of the available images were used for model training/testing.

Image file format

  • .bip: hyperspectral datacubes and band math images (single band)
  • .bip.hdr: header file (see the ENVI .hdr specification for more information)
  • .hdr "history" tag shows all operation done to an image up to this point (follows the convention of Spectranon (Resonon software)).
  • .spec: BIP file with only a single "pixel"; .spec files contain data across the full spctral domain and do not contain any spatial information (they do have an accomanying .hdr file)

MSI Folder structure

At the highest level, there are two directories:

  1. data: contains the ground truth data, as well as all the image/spectral data before and after MSI processing.
  2. results: contains all the tuning, training, and testing results of the supervised regression models (Lasso and partial least squares regression were used).

Hyperparameter tuning loop

Tuning and all subsequent steps were carried out separately for each ground truth measurement (biomass, tissue nitrogen concentration, and nitrogen uptake), as well as each set of available features to be evaluated.

Ground truth measurements:

  • above-ground biomass (kg ha-1)
  • total nitrogen uptake (kg ha-1)
  • tissue nitrogen concenteration (%)

Available features to consider:

  • spectral features only
  • spectral features plus the 10th percentile MCARI2 value
  • spectral derivative features only

Supervised regression models evaluated:

  • Lasso regression
  • Partial least squares regression

Hyperparameter tuning was carried out by splitting the training dataset (60% of samples) using a repeated stratified k-fold cross validation (4 splits and 3 replications). Thus, each tuning fold uses 75% of the training samples, which use 60% of the total samples. The results of hyperparameter tuning are saved to the "tuning" folder in the appropriate directory

Model training and testing

Following hyperparameter tuning, each model is trained on the full training set (60% of all samples) using the optimal hyperparameters deterimined from the tuning step. The trained model is then used to predict each ground truth measurement using the test set (40% of samples). Test predictions and scores are saved to a .csv for each ground truth measurement and model. Figures are also created that show measured vs. predicted values for each number of features, as well as a figure that shows error as a function of feature number.

Other tasks

  • The time it takes to execute each step in the loop is recorded and saved to a .csv file in the base "results" folder ("msi_0_runtime.csv")
  • A .csv is created for each ground truth sample and each error metric (MAE, MSE, and R2) and is also saved in the base "results" folder (e.g., "msi_0_biomass_kgha_MAE.csv")