hyperspectral-imaging-ml

Reproducing HybridSN

• Create a conda environment with your OS using env-mac.yml or env-ubuntu.yml:

conda env create -f env-ubuntu.yml
conda activate hyperspec

• Optional update the gin.config with desired hyper-parameters. Current configuration matches the paper.
• Run the training script

python train.py

• View training results in Tensorboard

tensorboard --logdir runs

Note: data will be downloaded to ~/.hyperspec/

reporting

• Reproducibility Report
• Tips and Tricks

Papers:

• Deep Learning for Classification of Hyperspectral Data: A Comparative Review
  • An overview of the field relating to deep learning
  • code base
• HybridSN: Exploring 3D-2D CNN Feature Hierarchy for Hyperspectral Image Classification
  • Current state-of-the-art on the Indian Pines, Pavia University and Salinas Scene datasets
  • code base
• Hyperspectral Image Classification with Deep Metric Learning and Conditional Random Field
  • State of the art without additional data on the Indian Pines data set
  • None :(

Datasets

• Overview
• Indian Pine
• Data Fusion Contest 2018

Why

Hyper-spectral imaging is a upcoming field that has potential in the agriculture industry with many benefits including crop yield and carbon monitoring.

Paper Review

• Rigor vs. Empirical - Balanced?
• Readability - Excellent
• Algorithm Difficulty - Low
• Pseudo Code - None / Step-Code?
• Hyperparameters Specified - Yes
• Compute Needed - GPU
• Number of Equations - 2
• Number of Tables - 5

Paper Notes

• Proposes a hybrid 3d and 2d model for general hyperspectral image(HSI) classification
• 3-D CNN: Employs principal component analysis on input data to reduce spatio-spectral images by its spectral bands(depth) in order to remove spatial redundancy
  • 3D convolution → 3D kernel convolves on 3D-data(spatio-spectral image)
  • Uses 3d patches to determine image classification
  • 3D patches: overlapping spatio-spectral convolutions where the centered pixel is used for classification
  • Computationally expensive
  • Papers recommend 3 layered model to extract spectral features
    • One paper dubs this the Deep Metric Learning followed by a Conditional Random Field layer to make predictions
• 2-D CNN: Input data is convolved with 2d kernels(normal)
• Hybrid of both 3D and 2D Kernels are used for learning
  • Use of 3D convolutions to capture spatial data and 2D convolutions to decrease computational expense and learn non-spectral information (features of images for classification)
• Utilizes both spatio-spectral imaging in the form of 3-d convulsions and non spatio-spectral imaging in the form 2d convolutions
• This model also shows great performance with little data

Conclusion: We believe the paper is highly reproducible and very well documented. The only potential issue we foresee is within the preprocessing phase.