Resnet based hybrid convolution LSTM for hyperspectral image classification

What is it about?

There are many spectral bands of different wavelengths present in Hyperspectral Image containing a huge amount of information that helps to detect and identify various objects. Many challenges are faced at the time of analyzing a hyperspectral image like information loss, hindrances posed by redundant information lingering on input data and the presence of high dimensions, etc. In this paper, we proposed a Resnet ConvLSTM model which is composed of a 2D Convolution Neural Network together with Batch Normalization and it helps to minimize the computational complexity and to extract features from Hyperspectral Image. At the same time, we added skip connections to eliminate the vanishing gradient problem, being followed by the Long Short Term Memory layer to remove redundant information from an input image. We implemented our model on three different types of hyperspectral data sets and also on three different types of time series data sets. Our model produced better accuracy than others’ proposed models reaching the levels of 0.07%, 0.01%, and 0.56% more in the "Indian Pines", "Pavia University", and "Botswana" data sets respectively. The commitment of our errors decreased in time series datasets by 0.44, 0.08, and 0.5 in "Electricity production", "International Airline Passenger" and "Production of shampoo over three years" respectively.

Featured Image

Why is it important?

This paper introduces a Resnet ConvLSTM model for hyperspectral image analysis, addressing challenges like information loss and high dimensionality, achieving superior accuracy across diverse datasets. By effectively extracting features and mitigating the vanishing gradient problem, it demonstrates significant advancements in hyperspectral and time series data analysis, showcasing improved accuracy and error reduction.

Perspectives