A Semantic Segmentation Framework for Hyperspectral Imagery Based on Tucker Decomposition and 3DCNN Tested with Simulated Noisy Scenarios

Abstract

The present work, unlike others, does not try to reduce the noise in hyperspectral images to increase the semantic segmentation performance metrics; rather, we present a classification framework for noisy Hyperspectral Images (HSI), studying the classification performance metrics for different SNR levels and where the inputs are compressed. This framework consists of a 3D Convolutional Neural Network (3DCNN) that uses as input data a spectrally compressed version of the HSI, obtained from the Tucker Decomposition (TKD). The advantage of this classifier is the ability to handle spatial and spectral features from the core tensor, exploiting the spatial correlation of remotely sensed images of the earth surface. To test the performance of this framework, signal-independent thermal noise and signal-dependent photonic noise generators are implemented to simulate an extensive collection of tests, from 60 dB to −20 dB of Signal-to-Noise Ratio (SNR) over three datasets: Indian Pines (IP), University of Pavia (UP), and Salinas (SAL). For comparison purposes, we have included tests with Support Vector Machine (SVM), Random Forest (RF), 1DCNN, and 2DCNN. For the test cases, the datasets were compressed to only 40 tensor bands for a relative reconstruction error less than 1%. This framework allows us to classify the noisy data with better accuracy and significantly reduces the computational complexity of the Deep Learning (DL) model. The framework exhibits an excellent performance from 60 dB to 0 dB of SNR for 2DCNN and 3DCNN, achieving a Kappa coefficient from 0.90 to 1.0 in all the noisy data scenarios for a representative set of labeled samples of each class for training, from 5% to 10% for the datasets used in this work. The source code and log files of the experiments used for this paper are publicly available for research purposes.