Inference in Supervised Spectral Classifiers for On-Board Hyperspectral Imaging: An Overview

Abstract

Machine learning techniques are widely used for pixel-wise classification of hyperspectral images. These methods can achieve high accuracy, but most of them are computationally intensive models. This poses a problem for their implementation in low-power and embedded systems intended for on-board processing, in which energy consumption and model size are as important as accuracy. With a focus on embedded and on-board systems (in which only the inference step is performed after an off-line training process), in this paper we provide a comprehensive overview of the inference properties of the most relevant techniques for hyperspectral image classification. For this purpose, we compare the size of the trained models and the operations required during the inference step (which are directly related to the hardware and energy requirements). Our goal is to search for appropriate trade-offs between on-board implementation (such as model size and energy consumption) and classification accuracy.

FPGA Accelerator for Gradient Boosting Decision Trees

“…We have decided to focus on GBDT instead of Random Forest since recently GBDT have demonstrated an enormous potential [5]. Moreover, [10] compared the results of Random Forest and GBDT and the results show that GBDT provided better accuracy while using smaller models, hence we believe that it is a better approach for embedded systems. Another difference is that we have designed a custom register-transfer level (RTL) architecture instead of using a high-level synthesis that will automatically generate the RTL design from a C-code.…”

“…For instance, they have been successfully used to produce accurate forecasts for the COVID-19 evolution, and to identify factors that influence its transmission rate [6]; to detect fraud from customer transactions [7]; to estimate major air pollutants risks to human health [8] using satellite-based aerosol optical depth; or to classify the GPS signal reception in order to improve its accuracy [9]. Moreover, a recent publication [10] analyzed different machine learning methods for image processing in remote systems, focusing on on-board processing, which requires both performance and low-power. In this work the authors identified that GBDTs present a very interesting tradeoff between the use of computational and hardware resources and the obtained accuracy.…”

“…Hence, our primary goal is efficiency. The register-transfer level (RTL) design of our accelerator has been written in VHDL , and, to demonstrate its potential, we have implemented it in a low-cost FPGA evaluation board (ZedBoard) [12], which includes an FPGA for embedded systems, and we have used our implementation for the relevant case study analyzed in [10]: pixel classification of hyperspectral images with the objective of processing the data at run-time. We have measured the execution time and power consumption of our accelerator and we have identified that our design can be used to process complex GBDT models even when using a small FPGA.…”

Multiclass Non-Randomized Spectral–Spatial Active Learning for Hyperspectral Image Classification

“…Some widely used supervised classification approaches for HSI analysis are Multinomial Logistic Regression (MLR) [10], Support Vector Machine (SVM) [11], Maximum Likelihood [10,12], Ensemble Learning (EL) [13,14], Random Forests (RF) [15,16], Deep Learning (DL) [17][18][19], Transfer Learning [20,21], k-Nearest Neighbors (KNN) [22] and Extreme Learning Machine (ELM) [23]. The major limitation of supervised HSI classification is the poor performance due to the Hughes phenomena [24].…”

Towards the Concurrent Execution of Multiple Hyperspectral Imaging Applications by Means of Computationally Simple Operations

“…Unfortunately, the data transmission from the remote-sensing platforms to the Earth surface introduces important delays related to the communication of a large amount of data between the source and the final target, producing a bottleneck that can seriously reduce the effectiveness of real-time applications or applications that demand prompt replies [21,22]. Consequently, real-time on-board processing has become a very interesting topic within the remote-sensing field to provide a solution to this type of applications [14,15,22,23]. In the space domain, this is due to the fact that both the acquisition rates of the next-generation hyperspectral sensors and the computational capabilities of the latest-generation space-grade hardware devices are growing [24].…”