HSI-Dataset-API 1.5.1

API for accessing HSI datasets

Install

pip install HSI-Dataset-API

Links to the available HSI datasets

• Nextcloud: HSI Dataset v1.21.zip
• Google Drive: HSI Dataset v1.21.zip

Dataset structure

Dataset should be stored in the following structure:

Plain structure (#1)

{dataset_name}
├── hsi
│   ├── 1.npy
│   └── 1.yml
├── masks
│   └── 1.png
└── meta.yaml

Or in structure like this (such structure was created while using data cropping)

Cropped data structure (#2)

{dataset_name}
├── hsi
│   ├── specter_1
│   │   ├── 1.npy
│   │   ├── 1.yml
│   │   ├── 2.npy
│   │   └── 2.yml
│   └── specter_2
│       ├── 1.npy
│       └── 1.yml
├── masks
│   ├── specter_1
│   │   ├── 1.png
│   │   └── 2.png
│   └── specter_2
│       └── 1.png
└── meta.yaml

Meta.yml

In this file you should provide classes description (it's name and label). Also, you can store any helpful information that describes the dataset.

For example:

name: HSI Dataset example
description: Some additional info about dataset
classes:
  cat: 1
  dog: 2
  car: 3
wave_lengths:
- 420.0
- 640.0
- 780.0 

{number}.yml

In this file you can store HSI specific information such as date, name of humidity.

For example:

classes:
  - potato
height: 512
width: 512
layersCount: 237
original_filename: '210730_134940_'
top_left:
  - 0
  - 0

Python API

Via API presented in this repo you can access the dataset.

Importing

from hsi_dataset_api import HsiDataset, HsiDataCropper

Cropping the data

base_path = '/mnt/data/corrected_hsi_data'
output_path = '/mnt/data/cropped_hsi_data'
classes = ['potato', 'tomato']
selected_folders = ['HSI_1', 'HSI_2']  # Completely optional

cropper = HsiDataCropper(side_size=512, step=8, objects_ratio=0.20, min_class_ratio=0.01)
cropper.crop(base_path, output_path, classes, selected_folders)

Plot cropped data statistics

cropper.draw_statistics()

Using the data

Create Data Access Object

dataset = HsiDataset('../example/dataset_example', cropped_dataset=False)

Parameter cropped_dataset controls type of the dataset structure. If the dataset persist in the memory in the structure like second (#2) - set this parameter to True

Getting the dataset meta information

dataset.get_dataset_description()

Getting the shuffled train data using python generator

for data_point in dataset.data_iterator(opened=True, shuffle=True):
    hyperspecter = data_point.hsi
    mask = data_point.mask
    meta = data_point.meta

Examples

See jupyter notebook example by the following link:

https://nbviewer.org/github/Banayaki/hsi_dataset_api/blob/master/examples/ClassificationMLP.ipynb

Source code

Source code is available:

https://github.com/Banayaki/hsi_dataset_api