Hyperspectral-Image-Classification-using-Deep-Learning

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

Hyperspectral-Image-Classification-using-Deep-Learning

(★ 11)

Hyperspectral Image Classification using Deep Neural Network Architectures with Transfer Learning

Hyperspectral-Image-Classification-using-Deep-Learning 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Hyperspectral-Image-Classification-using-Deep-Learning
Hyperspectral Image Classification using Deep Neural Network Architectures with Transfer Learning

This is an attempt to implement SGCNN-X (Shuffled Group Convolutional Neural Network) models from the paper https://www.mdpi.com/2072-4292/12/11/1780 where X represents the number of convolution layers.

To classify Hypersectral Images using transfer learning, following steps are executed
(Source : Indian Pines, Target : Botswana)

1. Samples of size **S X S X 64** (S - sample size,) from the image and labels are assigned to those samples using ground truth image. Samples are extracted using a 
   variable overlap_ratio which leads to generation of several datasets. An overlap ratio of 25% implies that a sample belonging to a class will be selected if and 
   only if the next sample from same class does not overlap more than 25% for this sample.
   
   Sample size (cube size) chosen is 20 (19 in paper)
   
   Center pixel from the ground truth image is used to assign label to a sample.
   
   An overlap ratio of 1 generates maximum number of samples.

2. Once the samples are extracted, two methods below have been used to split the extracted samples into training & test sets.
   
   **V1** : Specific number of samples from each class are put into the training set for pretraining (source) as well as fine tuning (target). 
            This methods follows the train test split tables mentioned in the paper.
        
   **Other** : Some percentage of samples are put into the training set and the rest are added to the test set.
               Models in folder SGCNN_7, SGCNN_8 & SGCNN_12 use this train test split method.
   
## Folder Details

Datasets : mat files of image and ground truth for Indian Pines, Botswana & Pavia datasets.

V1 : Implementation of SGCNN with traning and test set split as mentioned in the paper.

The following folders implement a training test split where a percentage of samples from each class are added to the training set while the rest are
added to the test set. This split is done at both the pretraining stage and the fine tuning stage.

## The python files in all folders implement the sample extraction, train test split, model training, transfer learning and model architecture functions.

SGCNN_7 : Implementation of SGCNN-7 architecture from the paper. Model is trained on Indian Pines and fine tuned on Botswana.

SGCNN_8 : Implementation of SGCNN-8 architecture from the paper. Model is trained on Indian Pines and fine tuned on Botswana.

SGCNN_12 : Implementation of SGCNN-12 architecture from the paper. Model is trained on Indian Pines and fine tuned on Botswana.

Hyperspectral-Image-Classification-using-Deep-Learning

Hyperspectral-Image-Classification-using-Deep-Learning 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Hyperspectral-Image-Classification-using-Deep-Learning

Hyperspectral-Image-Classification-using-Deep-Learning 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README