An in-depth analysis of hyperspectral target detection with shadow compensation via LiDAR☆
Introduction
Target detection lies at the very core of hyperspectral imaging (HSI) research. However, shadowy areas present a big hindrance in target detection from HSI, as the reflectance data received from target materials in the near infrared (NIR) and short wave infrared (SWIR) bands is significantly diminished when measured from shadowy areas [1], [2]. Since hyperspectral sensors use the sun as a source of radiation, the same material in a hyperspectral dataset can show a spectral variation based on the lighting conditions, which lead to decreased target detection rates in shadows. With the use of LiDAR, shadow finding algorithms based solely on HSI [3], [4], [5] were complemented with more accurate information [6], and the researchers gained the ability to use radiance models [7], [8], [9], [10]. The studies in [11], [12] make a detailed comparison of shadow pixels detected from HSI and from LiDAR. Apparent from these studies is the critical importance of the position of the sun in data collection, the effect of shadows on target detection and the advantages of using LiDAR.
In this work, our goal is to get a clear understanding of the effects of targets and illumination on detection. The first step in dealing with shadows is to locate the shadow regions, which can be efficiently done using LiDAR data [13], [14]. Therefore, hyperspectral and LiDAR data, complete with a set of atmospheric measurements and the information on the angles and altitudes of data collection, present a very unique medium for HSI research. Unfortunately, such data with good ground truths is very limited, and even if it exists, researchers have to first deal with shadow extraction from LiDAR. Further, if one wants to incorporate this to the hyperspectral radiance model, atmospheric modeling tools such as MODTRAN [15], [16] take ample time, and may not be available to everyone.
In this paper, we use the Share2012 dataset [17], deal with all of the issues listed above for this dataset, and share all the necessary codes and outputs on the web to facilitate rapid and comparable research. Further, we propose to perform shadow correction on hyperspectral data using LiDAR and explain it on an elementary guide to the radiance model. Then, we show that shadow correction using LiDAR data significantly increases HSI target detection rates. For this purpose, we provide the results of the spectral angle mapper (SAM), adaptive coherence estimator (ACE) and matched filter (MF) detectors and compare their outputs on both the reflectance dataset of Share2012 (provided with the dataset based on atmospheric correction with ATCOR), and on our reflectance dataset which is obtained through shadow and atmospheric correction. Finally, we investigate if data collection in the morning or in the evening makes a difference; and how much the color or the background of the target affect the results. With these results, we provide a benchmark for HSI-LiDAR exploration.
The rest of the paper is organized as follows: we present the related work in Section 2, and discuss the details of the proposed method in Section 3 with an elementary guide to the radiance model in Section 3.1. We explain the dataset and the targets detection methods used to evaluate our results in Section 4. Then we give the experimental results in Section 5 and conclude in Section 6.
Section snippets
Related work
Shadows do not only lower the overall intensity of the radiance spectrum, they also cause a change in the spectral shape. Ashton et al. [18] explained this phenomena with the color of light in a shadowed region being different from that in a sunny area. In one of the earlier studies, Nischan et al. [19] suggested using an active illumination source to improve the target-detection performance. Later, Ibrahim et al. [20] tested the use of an absorptive polarization filter placed at the front of
Our approach
The flowchart of our shadow correction work is given in Fig. 1. We first match the hyperspectral and LiDAR data based on their coordinates. Then, we detect the shadows using a line-of-sight analysis from LiDAR, using the sun angles at the time of HSI collection. Third, we use the parameters in Table 2 and find the atmospheric values from MODTRAN. And finally, we put all of these into the radiance model and compute the shadow-compensated reflectance values. These reflectances are then used in
Dataset and evaluation
Although one dataset may not be enough to come to a concrete understanding of an underlying phenomena, there are not very many multi-source accessible datasets that have the properties that Share2012 data has. Share2012 data presents very unique features by the availability of the atmospheric and flight parameters. In this section, we first describe the dataset and list all the necessary parameters to run MODTRAN. Then, we briefly list the three target detectors, namely the SAM, ACE and MF that
Experimental results
For the experiments, first the ground truth files of the targets provided with the Share2012 data are matched with the radiance dataset, then atmospheric and shadow correction is performed. Afterwards, a target signature from the blue and red targets from the fully illuminated region is extracted. These target signatures are compared with all the other pixels using the ACE, SAM and MF algorithms. For covariance computations, background is taken as the entire data as shown in Fig. 10, of size
Conclusion
In this paper, we analyze the target detection based on the illumination conditions, based on the background on which the target is placed, and based on the target color. While it is obviously clear that illumination or the lack of plays an important role in target detection, this phenomena is reversible by incorporating the shadow information, which can be easily obtained via LiDAR, into the physical radiance model. The improvement in target detection due to shadow correction can be seen
CRediT authorship contribution statement
Emrah Oduncu: Methodology, Software, Writing – original draft. Seniha Esen Yuksel: Methodology, Conceptualization, Supervision, Writing – reviewing and editing, Funding acquisition.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (50)
- et al.
Operational and performance considerations of radiative-transfer modeling in hyperspectral target detection
IEEE Trans. Geosci. Rem. Sens.
(2011) - A. Karakaya, S.E. Yuksel, Target detection in hyperspectral images, in: IEEE Signal Processing and Communications...
- U. Sakarya, C. Demirkesen, M. Teke, Unsharp masking filter based shadow-invariant feature extraction for hyperspectral...
- et al.
Shadow removal from vnir hyperspectral remote sensing imagery with endmember signature analysis
- et al.
Shadow-insensitive material detection/classification with atmospherically corrected hyperspectral imagery
- et al.
Illumination and shadow compensation of hyperspectral images using a digital surface model and non-linear least squares estimation
Leveraging lidar data to aid in hyperspectral image target detection in the radiance domain
- K.O. Niemann, G. Frazer, R. Loos, F. Visintini, Lidar-guided analysis of airborne hyperspectral data, in: Hyperspectral...
- et al.
Fusion of hyperspectral and LIDAR remote sensing data for classification of complex forest areas
IEEE Trans. Geosci. Remote Sens.
(2008) - et al.
Hyperspectral images and lidar based DEM fusion: A multi-modal landuse classification strategy
IEEE Geosci. Remote Sens. Symp.
(2014)
The effect of the LiDAR sensor on the success of shadow detection from hyperspectral data
Pamukkale Uni. J. Eng. Sci.
Locating the shadow regions in LiDAR data: results on the SHARE 2012 dataset
Active spectral imaging
Linc. Lab. J.
Illumination invariance and shadow compensation via spectro-polarimetry technique
Opt. Eng.
Hyperspectral analysis of objects under shadow conditions based on field reflectance measurements
Appl. Opt.
Shadow modelling and correction techniques in hyperspectral imaging
Electron. Lett.
Shadow detection and restoration for hyperspectral images based on nonlinear spectral unmixing
Remote Sens.
Hyperspectral shadow removal via nonlinear unmixing
IEEE Geosci. Remote Sens. Lett.
Review of shadow detection and de-shadowing methods in remote sensing
Chin. Geogr. Sci.
Cited by (4)
Triple shadow multilinear unmixing for near-ground hyperspectral vegetation canopy shadow removal
2024, Computers and Electronics in AgricultureStudy on a risk model for prediction and avoidance of unmanned environmental hazard
2022, Scientific ReportsHyperspectral Image Classification Based on Convolutional Neural Network Embedded with Attention Mechanism and Shadow Enhancement by Dynamic Stochastic Resonance
2022, 2022 7th International Conference on Image, Vision and Computing, ICIVC 2022