Estudio Comparativo de Técnicas de Clasificación de Imágenes Hiperespectrales

Abstract

ResumenLas imágenes hiperespectrales constituyen el núcleo de varios programas de observación remota de la Tierra. La cantidad de información que contienen estas imágenes, formadas por cientos de canales espectrales estrechos y casi continuos, resulta de gran utilidad en aplicaciones en las que la caracterización de los materiales observados en la superficie terrestre resulta de gran relevancia. Esto se debe a la posibilidad de caracterizar de forma inequívoca cada material a través de su firma espectral. Algunas de estas aplicaciones son la agricultura de precisión, la planificación de espacios urbanos, o la prevención y seguimiento de desastres naturales. Sin embargo, la gran dimensión de las imágenes hiperespectrales supone un reto en su tratamiento, almacenamiento y procesamiento, debido a la gran variabilidad espectral y la correlación existente en los datos. En la literatura se han desarrollado múltiples algoritmos de análisis de imágenes hiperespectrales. En este artículo revisamos los algoritmos más utilizados para la clasificación de este tipo de imágenes, realizando experimentos con tres imágenes públicas y presentando una comparativa entre los métodos más ampliamente utilizados en este campo. AbstractHyperspectral images are very important in many Earth Observation programs. The large amount of information is contained in hyperspectral images (hundreds of narrow and continuous spectral channels) is very useful for applications in which the characterization of the Earth surface materials is relevant. This is due to the fact that each observed element can be uniquely characterized by its spectral signature, for instance in precision agriculture, urban planning or detection/prevention of natural disasters, among others. However, the large dimensionality of hyperspectral images represents a challenge for analysis algorithms, both from the storage and processing viewpoints, resulting from data variability and correlation. Several algorithms have been proposed in the literature for the analysis of hyperspectral images. In this paper, we review the most popular techniques for hyperspectral classification. These techniques are inter-compared using three publicly available hyperspectral data sets.

Trend Technologies for Robotic Fertilization Process in Row Crops

“…They are measuring green vegetation through a normalized ratio ranging from -1 to 1. ( Paoletti et al, 2019 ; Moriarty et al, 2019 ).…”

“…The main developments carried out with multispectral cameras are based on the vegetative analysis of plant growth, using NDVI indices, which have been shown to provide relevant information for making decisions in fertilization applications. ( Paoletti et al, 2019 ; Cardim Ferreira Lima et al, 2020 ; Lu et al, 2020 ; Zhou et al, 2021 ).…”

Rotation Equivariant Convolutional Neural Networks for Hyperspectral Image Classification

“…Traditionally, HSI information has been exploited in machine learning (ML) by pixel-wise methods which consider HSI data as a list of spectral vectors, assuming that each pixel is pure and typically labeled as a single land cover type [32]- [34]. In the current literature, there are abundant pixelwise methods, such as the popular support vector machines (SVMs) [35], [36], K-nearest neighbor (KNN) [37], [38], multinomial logistic regression (MLR) [39], [40], random forests (RFs) [41], [42] and standard artificial neural networks (ANNs) [43], among others.…”

A Deep Learning Approach for Fusing Sensor Data from Screw Compressors

“…We propose a one-dimensional convolutional neural network (1D CNN) to merge data from several sensors. A 1D convolutional layer is used as an input to our model in order to reduce the number of network connections and to process all information together [30]. This layer receives an input data array with a shape of C channels and a sliding window length of L time steps.…”