We propose a novel Deep learning approach using autoencoders to map spectral bands to a space of lower dimensionality while preserving the information that makes it possible to discriminate different materials. Deep learning is a relatively new pattern recognition approach which has given promising result in many applications. In Deep learning a hierarchical representation of increasing level of abstraction of the features is learned. Autoencoder is an important unsupervised technique frequently used in Deep learning for extracting important properties of the data. The learned latent representation is a non-linear mapping of the original data which potentially preserve the discrimination capacity.
We have performed a field trial with an airborne push-broom hyperspectral sensor, making several flights over the
same area and with known changes (e.g., moved vehicles) between the flights. Each flight results in a sequence
of scan lines forming an image strip, and in order to detect changes between two flights, the two resulting image
strips must be geometrically aligned and radiometrically corrected. The focus of this paper is the geometrical
alignment, and we propose an image- and gyro-based method for geometric co-alignment (registration) of two
image strips. The method is particularly useful when the sensor is not stabilized, thus reducing the need for
expensive mechanical stabilization. The method works in several steps, including gyro-based rectification, global
alignment using SIFT matching, and a local alignment using KLT tracking. Experimental results are shown but
not quantified, as ground truth is, by the nature of the trial, lacking.
We present a new hyperspectral data set that FOI will keep publicly available. The hyperspectral data set was collected in
an airborne measurement over the countryside. The spectral resolution was about 10 nm which allowed registrations in 60
spectral bands in the visual and near infrared range (390-960 nm). Objects with various signature properties were placed
in three areas: the edge of a wood, an open field and a rough open terrain. Several overflights were performed over the
areas. Between the overflights some of the objects were moved, representing different scenarios. Our interest is primarily
in anomaly detection of man-made objects placed in nature where no such objects are expected. The objects in the trial
were military and civilian vehicles, boards of different size and a camouflage net. The size of the boards range from multipixel
to subpixel size. Due to wind and cloud conditions the stability and the flight height of the airplane vary between
the overflights, which makes the analysis extra challenging.
The ROC curve is the most frequently used performance measure for detection methods and the underlying
sensor configuration. Common problems are that the ROC curve does not present a single number that can be
compared to other systems and that no discrimination between sensor performance and algorithm performance
is done. To address the first problem, a number of measures are used in practice, like detection rate at a specific
false alarm rate, or area-under-curve. For the second problem, we proposed in a previous paper<sup>1</sup> an information
theoretic method for measuring sensor performance. We now relate the method to the ROC curve, show that it
is equivalent to selecting a certain point on the ROC curve, and that this point is easily determined. Our scope
is hyperspectral data, studying discrimination between single pixels.
The multi-optical mine detection system (MOMS) is a research project focused on the detection of surface laid mines. In
the sensor suite, both passive and active sensors are included, such as IR as well as hyper- and multispectral cameras,
and 3-D laser radar.
Extensive field experiments have been conducted to collect data under various environmental conditions. Three seasons
have been covered during the field campaigns: Spring, summer, and autumn. Furthermore, the mines have been arranged
in three different types of vegetation scenarios. Also, a long term data collection effort has been conducted to collect
diurnal and seasonal signature variations.
Among the signal processing techniques considered, anomaly detection emerges as a key component in a system
concept. The method is based on detecting small differences between the mine-sized object and a local background. The
spectral features of the detected anomalies are further analyzed with respect to general commonalities in the scene and
known spectral properties of mine-like objects. In this paper we present some of the results from the project.
State of the art and coming hyperspectral optical sensors generate large amounts of data and automatic analysis is
necessary. One example is Automatic Target Recognition (ATR), frequently used in military applications and a coming
technique for civilian surveillance applications. When sensors communicate in networks, the capacity of the
communication channel defines the limit of data transferred without compression. Automated analysis may have
different demands on data quality than a human observer, and thus standard compression methods may not be optimal.
This paper presents results from testing how the performance of detection methods are affected by compressing input
data with COTS coders. A standard video coder has been used to compress hyperspectral data. A video is a sequence of
still images, a hybrid video coder use the correlation in time by doing block based motion compensated prediction
between images. In principle only the differences are transmitted. This method of coding can be used on hyperspectral
data if we consider one of the three dimensions as the time axis. Spectral anomaly detection is used as detection method
on mine data. This method finds every pixel in the image that is abnormal, an anomaly compared to the surroundings.
The purpose of anomaly detection is to identify objects (samples, pixels) that differ significantly from the background,
without any a priori explicit knowledge about the signature of the sought-after targets. Thus the role of the anomaly
detector is to identify "hot spots" on which subsequent analysis can be performed. We have used data from Imspec, a
hyperspectral sensor. The hyperspectral image, or the spectral cube, consists of consecutive frames of spatial-spectral
images. Each pixel contains a spectrum with 240 measure points. Hyperspectral sensor data was coded with hybrid
coding using a variant of MPEG2. Only I- and P- frames was used. Every 10th frame was coded as I frame. 14
hyperspectral images was coded in 3 different directions using x, y, or z direction as time. 4 different quantization steps
were used. Coding was done with and without initial quantization of data to 8 bbp. Results are presented from applying
spectral anomaly detection on the coded data set.
We address the problem of target discrimination using hyperspektral and multisensor data. The problem is of significance
to detection and classification of low signature targets such as landmines. The problem will be described with stochastic
models and by using information theoretic concepts we will derive limits to system performance in terms of probability of
false alarm and probablility of detection.
The stochastic model is suitable to evaluate and optimize sensor parameters and sensor configurations. We will for example
investigate how much information different combinations of spectral and spatial data will give. With the stochastic
model sensors with different types of characteristics can be compared and the contribution of the different sensors and
configurations can be evaluated.
Besides the optimization of different sensor configurations with respect to specific applications, the stochastic model will
be used to evaluate different anomaly detectors.
The strength of our approach is shown by examples from an analysis of measurements on a natural scene with various
objects using an electro-optical hyperspectral sensor and several other sensors. We expect that our approach will give significant
indications on how to choose and configure sensors for efficient and reliable target discrimination.