Spectral characteristics and feature selection of hyperspectral remote sensing data

Hyperspectral remote sensing data with bandwidth of nanometre (nm) level have tens or even several hundreds of channels and contain abundant spectral information. Different channels have their own properties and show the spectral characteristics of various objects in image. Rational feature selection from the varieties of channels is very important for effective analysis and information extraction of hyperspectral data. This paper, taking Shunyi region of Beijing as a study area, comprehensively analysed the spectral characteristics of hyperspectral data. On the basis of analysing the information quantity of bands, correlation between different bands, spectral absorption characteristics of objects and object separability in bands, a fundamental method of optimum band selection and feature extraction from hyperspectral remote sensing data was proposed.     

RLQ and fourth-corner analysis of plant species traits and spectral indices derived from HyMap and CHRIS-PROBA imagery

Databases on plant traits as well as the availability of global coverage of high spatial and spectral resolution remote-sensing data are constantly growing. However, little effort has been made to analyse the relationship between plant traits and remote-sensing data while simultaneously taking species identity and abundance into consideration. We correlated quantitative and qualitative plant traits from a dwarf shrub savanna in Namibia, with spectral indices derived from two hyperspectral sensors, HyMap and the Compact High Resolution Imaging Spectrometer Project for On-Board Autonomy (CHRIS-PROBA), which differ in their spatial and spectral resolution. We used RLQ analysis and the fourth-corner statistic, which are two three-table ordination approaches that circumvent the so-called fourth-corner problem. A higher spatial resolution helped to identify trait–index correlations linked to vegetation structure, while a lower spatial resolution pointed at traits linked to vegetation cover. A higher spectral resolution did not improve the relationships between spectral indices and plant traits. However, continuous hyperspectral signatures allowed for the calculation of spectral indices that make use of the detailed spectra allowing for more sophisticated spectral indices. We propose RLQ and the fourth-corner statistic as suitable tools for the remote sensing and Earth observation community that allow the direct correlation of trait databases with remotely sensed information.     

Hyperspectral data spectrum and texture band selection based on the subspace-rough set method

In order to improve the utilization rate of spectroscopic data and texture information, this study proposes a method for optimal selection of spectrum and texture features based on automatic subspace division and rough set theory. This method takes advantage of rough set reduct ideology in order to realize the reduction of different types of ground object spectral features on the basis of the conventional subspace division method. In using this method, the primary spectral band based on spectral information can be determined. Then, the grey-level co-occurrence matrix method can be used to calculate the texture information of the primary spectral band and determine the reduction and optimization in order to obtain the final band based on the spectrum and texture information. Verification of this method is made by using CASI data of Heihe Region, China, and AVIRIS data of the Indiana Region, USA, and also using Support Vector Machine (SVM) classification of the original spectral, primary spectral, and final bands. The results indicate the following. (1) The method for optimal selection of the critical spectral band and texture band, based on the rough set theory, can efficiently improve the classification accuracy of high-spatial resolution remote-sensing images. However, the effects for the low-spatial resolution images are minimal. (2) For high-spatial-resolution remote-sensing images, such as roads, trenches, buildings, and other types of object with obvious textural features, the addition of image texture information can increase the degree of distinction of these different types and thereby improve the classification accuracy. However, the addition of the textural information for some objects with similar texture features will cause misclassification and reduce the classification accuracy for these types of images. (3) This method can realize the optimal selection of spectrum and texture bands of a hyperspectral image and has a certain universality. Also, the texture information will be richer and this method will be more practical through increasing the spatial resolution of images.     

Influence of acquisition time and resolution on wheat yield estimation at the field scale from canopy biophysical variables retrieved from SPOT satellite data

Detailed information about the prediction of within-field potential in terms of yield at the field scale is an attractive goal that would allow useful applications in precision agriculture. Biophysical variables characterizing crop canopies, such as the leaf area index (LAI), fraction of absorbed photosynthetically active radiation (FAPAR), fractional ground cover (Fcover) and the concentration of chlorophyll-a and -b (Cab), can be estimated from satellite remote-sensing data through the application of a neural network inversion of a radiative transfer model, such as PROSAIL. The knowledge of the temporal and spatial variability of these variables can enhance the possibilities of estimating yield at the field scale. The aim of this study is to investigate the influence of acquisition time and spatial resolution of biophysical variables estimated from satellite data on the grain yield estimation of wheat crops. We used SPOT 4 (spatial resolution: 20 m) and SPOT 5 (spatial resolution: 10 m) images, acquired at six different dates during the wheat growing season in 2012, to obtain LAI, Fcover, FAPAR, and Cab on five fields in Maccarese (Central Italy). A preliminary survey was carried out to correlate spatially biophysical variables with the final grain yield at each acquisition date. Biophysical variables estimated at a spatial resolution of 10 m during the stem elongation stage showed the best simple and spatial correlation with yield. At this stage, all the biophysical variables showed the highest correlation values as compared to the other dates. Subsequently, we used the variables estimated from SPOT data at each growth stage to calibrate multiple linear regression (MLR) and cubist regression (CR) models for two fields, which were then validated on five independent fields. Although the CR calibration models provided better accuracy than MLR, the best validation statistics were gained from MLR models, obtaining a root mean square error (RMSE) of about 1 t ha^?1 for three of these fields, using remote data having a spatial resolution of 10 metres and acquired between steam elongation and booting stage. The optimal acquisition time is affected, ceteris paribus, by the agricultural management and in particular by the variety that can influence the trend of crop growth. However, the optimal growth stage for yield estimation seems to be quite similar over the study area during a growth season. The validation of models on field data collected in another growing season is mainly affected by the climate conditions. These results highlight the importance of spatial resolution and the influence of acquisition time of satellite images on the estimation of yield at the field scale by remote-sensing data.     

Phase correlation based redundancy removal in feature weighting band selection for hyperspectral images

Feature weighting based band selection provides a computationally undemanding approach to reduce the number of hyperspectral bands in order to decrease the computational requirements for processing large hyperspectral data sets. In a recent feature weighting based band selection method, a pair‐wise separability criterion and matrix coefficients analysis are used to assign weights to original bands, after which bands identified to be redundant using cross correlation are removed, as it is noted that feature weighting itself does not consider spectral correlation. In the present work, it is proposed to use phase correlation instead of conventional cross correlation to remove redundant bands in the last step of feature weighting based hyperspectral band selection. Support Vector Machine (SVM) based classification of hyperspectral data with a reduced number of bands is used to evaluate the classification accuracy obtained with the proposed approach, and it is shown that feature weighting band selection with the proposed phase correlation based redundant band removal method provides increased classification accuracy compared to feature weighting band selection with conventional cross correlation based redundant band removal.     

高光谱数据的光谱信息特点及面向对象的特征参数选择——以北京顺义区为例

波段宽度为纳米级的高光谱数据,具有几十乃至几百个光谱通道,它们各有不同的特点。如何根据具体的应用目的,在这众多的波段中选择出最佳波段和特征参数,对于有效地进行高光谱数据的处理、分析及信息提取至关重要。以北京顺义区高光谱数据为例,首先分析了通道间的相关性,根据通道的相关性大小和分组块状结构特点,将其分为若干组;然后全面分析了高光谱数据的光谱信息特征,在综合考虑各波段的信息含量、波段间的相关性以及地物光谱的吸收特性和可分性等因素
的基础上,提出了面向对象的分层多次选择高光谱数据最佳波段和提取特征参数的基本思路和方法;最后用其它地区的成像光谱数据对此方法进行了验证。     

Spectral discrimination of healthy and Ganoderma-infected oil palms from hyperspectral data

Although hyperspectral remote sensing has been used to study many agricultural phenomena such as crop stress and diseases, the potential use of this technique for detecting Ganoderma disease infestations and damage to oil palms under field conditions has not been explored to date. This research was conducted to investigate the feasibility of using a portable hyperspectral remote-sensing instrument to identify spectral differences between oil-palm leaves with and without Ganoderma infections. Reflectance spectra of samples representative of three classes of disease severity were collected. The most significant bands for spectral discrimination were selected from reflectance spectra and first derivatives of reflectance spectra. The significant wavelengths were identified using one-way analysis of variance. Then, a Jeffries–Matusita (JM) distance measurement was used to determine spectral separability between the classes. A maximum likelihood classifier method was used to classify the three classes based on the most significant wavelength spectral responses, and an error matrix was finally used to assess the accuracy of the classification.     

面向分类应用的高光谱谱段选择方法

高光谱数据在物质分类识别领域得到了广泛应用,但存在数据量大、波段间相关性高等问题,严重影响分类精度及应用。针对以上问题分析了已有的波段选择方法,提出了基于波段聚类及监督分类的遗传算法,对高光谱数据进行波段选择：采用[K]均值聚类算法对波段数据进行聚类分析,构造波段子集合;利用分类器族分类精度构造适应度函数,采用遗传算法对波段子集合进行优化选择。最后用阔叶林高光谱数据对提出的算法进行对比实验,实验结果表明针对分类应用,提出的算法能够非常有效地选择高光谱谱段。     

A non-negative low-rank representation for hyperspectral band selection

Hyperspectral images are widely used in real applications due to their rich spectral information. However, the large volume brings a lot of inconvenience, such as storage and transmission. Hyperspectral band selection is an important technique to cope with this issue by selecting a few spectral bands to replace the original image. This article proposes a novel band selection algorithm that first estimates the redundancy through analysing relationships among spectral bands. After that, spectral bands are ranked according to their relative importance. Subsequently, in order to remove redundant spectral bands and preserve the original information, a maximal linearly independent subset is constructed as the optimal band combination. Contributions of this article are listed as follows: (1) A new strategy for band selection is proposed to preserve the original information mostly; (2) A non-negative low-rank representation algorithm is developed to discover intrinsic relationships among spectral bands; (3) A smart strategy is put forward to adaptively determine the optimal combination of spectral bands. To verify the effectiveness, experiments have been conducted on both hyperspectral unmixing and classification. For unmixing, the proposed algorithm decreases the average root mean square errors (RMSEs) by 0.05, 0.03, and 0.05 for the Urban, Cuprite, and Indian Pines data sets, respectively. With regard to classification, our algorithm achieves the overall accuracies of 77.07% and 89.19% for the Indian Pines and Pavia University data sets, respectively. These results are close to the performance with original images. Thus, comparative experiments not only illustrate the superiority of the proposed algorithm, but also prove the validity of band selection on hyperspectral image processing.     

Multidimensional local spatial autocorrelation measure for integrating spatial and spectral information in hyperspectral image band selection

Hyperspectral band selection aims at the determination of an optimal subset of spectral bands for dimensionality reduction without loss of discriminability. Many conventional band selection approaches depend on the concept of “statistical distance” measure between the probability distributions characterizing sample classes. However, the maximization of separability does not necessarily guarantee that a classification process results in the best classification accuracies. This paper presents a multidimensional local spatial autocorrelation (MLSA) measure that quantifies the spatial autocorrelation of the hyperspectral image data. Based on the proposed spatial measure, a collaborative band selection strategy is developed that combines both spectral separability measure and spatial homogeneity measure for hyperspectral band selection without losing the spectral details useful in classification processes. The selected band subset by the proposed method shows both larger separability between classes and stronger spatial similarity within class. Case studies in biomedical and remote sensing applications demonstrate that the MLSA-based band selection approach improves object classification accuracies in hyperspectral imaging compared with conventional approaches.     

Band Selection of Hyperspectral Images for Automatic Detection of Poultry Skin Tumors

This paper presents a spectral band selection method for feature dimensionality reduction in hyperspectral image analysis for detecting skin tumors on poultry carcasses. A hyperspectral image contains spatial information measured as a sequence of individual wavelength across broad spectral bands. Despite the useful information for skin tumor detection, real-time processing of hyperspectral images is often a challenging task due to the large amount of data. Band selection finds a subset of significant spectral bands in terms of information content for dimensionality reduction. This paper presents a band selection method of hyperspectral images based on the recursive divergence for the automatic detection of poultry carcasses. For this, we derive a set of recursive equations for the fast calculation of divergence with an additional band to overcome the computational restrictions in real-time processing. A support vector machine is used as a classifier for tumor detection. From our experiments, the proposed band selection method shows high detection accuracy with low false positive rates compared to the canonical analysis at a small number of spectral bands. Also, compared with the enumeration approach of 93.75% detection rate, our proposed recursive divergence approach gives 90.6% detection rate, which is within the industry-accepted accuracy of 90-95%, while achieving the computational saving for real-time processing.     

Non-photosynthetic vegetation biomass estimation in semiarid Canadian mixed grasslands using ground hyperspectral data,Landsat 8 OLI,and Sentinel-2 images

ABSTRACT

Research on quantifying non-photosynthetic vegetation (NPV) with optical remote-sensing approaches has been focusing on optically distinguishing NPV from green vegetation and bare soil. With a very similar spectral response curve to NPV, dry moss is a significant component in semiarid mixed grasslands and plays a large role in NPV estimation. However, limited attention has been paid to this role. We investigated the potential of optical remote sensing to distinguish NPV biomass in semiarid grasslands characterized by NPV, biological soil crust dominated by moss and lichen, and bare soil. First, hyperspectral spectral indices were examined to determine the most useful spectral wavelength regions for NPV biomass estimation. Second, multispectral red-edge indices and shortwave infrared (SWIR) indices were simulated based on Landsat 8 Operational Land Imager (OLI) and Sentinel-2A MultiSpectral Instrument band reflectance, respectively, to determine the most suitable multispectral indices for NPV estimation. The potential multispectral indices were then applied to Landsat 8 OLI images and Sentinel-2A images acquired in early, middle, peak, and early senescence growing seasons to investigate the potential of satellite images for quantifying NPV biomass. Our results indicated that hyperspectral red-edge indices, modified simple ratio, modified red-edge normalized difference vegetation index (mNDVI₇₀₅), and normalized difference vegetation index (NDVI₇₀₅) are better than the SWIR hyperspectral indices, including cellulose absorption index for quantifying NPV biomass. The simulated multispectral red-edge spectral indices (NDVI_red-edge and mNDVI_red-edge) demonstrate good and comparable performance on quantifying NPV biomass with SWIR multispectral indices (normalized difference index [NDI5 and NDI7] and soil-adjusted corn residue index). Nevertheless, the multispectral indices derived from Landsat 8 OLI and Sentinel-2 images have limited potential for NPV biomass estimation.     

A semi-supervised spatially aware wrapper method for hyperspectral band selection

Band selection is widely used to identify relevant bands for land-cover classification of hyperspectral images. The combination of spectral and spatial information can improve the classification performance of hyperspectral images dramatically. Similarly, the fusion of spectral–spatial information should also improve the performance of band selection. In this article, two semi-supervised wrapper-based spectral–spatial band selection algorithms are proposed. The local spatial smoothness of hyperspectral imagery is used to improve the performance of band selection when limited labelled samples available. With superpixel segmentation, the first algorithm uses the statistical characteristics of classification map to predict the classification quality of all samples. Based on the Markov random field model, the second algorithm incorporates the spatial information by the minimization of spectral–spatial energy function. Four widely used real hyperspectral data sets are used to demonstrate the effectiveness of the proposed methods, when compared to cross-validation-based wrapper method, the accuracy is improved by 2% for different data sets.     

Patch-based active learning (PTAL) for spectral-spatial classification on hyperspectral data

Active learning (AL) has been shown to be a useful approach to improving the efficiency of the classification process for remote-sensing imagery. Current AL methods are essentially based on pixel-wise classification. In this paper, a new patch-based active learning (PTAL) framework is proposed for spectral-spatial classification on hyperspectral remote-sensing data. The method consists of two major steps. In the initialization stage, the original hyperspectral images are partitioned into overlapping patches. Then, for each patch, the spectral and spatial information as well as the label are extracted. A small set of patches is randomly selected from the data set for annotation, then a patch-based support vector machine (PTSVM) classifier is initially trained with these patches. In the second stage (close-loop stage of query and retraining), the trained PTSVM classifier is combined with one of three query methods, which are margin sampling (MS), entropy query-by-bagging (EQB), and multi-class level uncertainty (MCLU), and is subsequently employed to query the most informative samples from the candidate pool comprising the rest of the patches from the data set. The query selection cycle enables the PTSVM model to select the most informative queries for human annotation. Then, these informative queries are added to the training set. This process runs iteratively until a stopping criterion is met. Finally, the trained PTSVM is employed to patch classification. In order to compare this to pixel-based active learning (PXAL) models, the prediction label of a patch by PTSVM is transformed into a pixel-wise label of a pixel predictor to get the classification maps. Experimental results show better performance of the proposed PTAL methods on classification accuracy and computational time on three different hyperspectral data sets as compared with PXAL methods.     

Selection of informative hyperspectral band subsets based on entropy and correlation

ABSTRACT

The article proposes two novel and relatively simple unsupervised procedures for the selection of informative small subsets of spectral bands in hyperspectral images. To ensure the informativeness of the subsets, bands featuring higher entropy are included. The correlation of band images is restricted to avoid redundancy of the subsets. The entropy multiple correlation ratio procedure employs the entropy-correlation ratio for the selection of spectral bands. The entropy-based correlated band grouping (ECBG) procedure divides the spectrum into groups of bands featuring highly correlated images. The subsets obtained were characterized by the performance of classifiers using only data from included bands. The ECBG procedure provided better results than the alternatives if the number of selected bands was low. Another advantage of this procedure is the possibility of averaging the images obtained for spectral bands within the groups found. It is shown that classification results are significantly improved if such an averaging is used. In the data acquisition practice, it can be used for a purposeful merging of spectral bands in the configuration of hyperspectral imagers, which allows one to reduce the amount of data to be saved in real time and thus helps one to improve the achievable spatial resolution.     

A Two-Step Filtering approach for detecting maize and soybean phenology with time-series MODIS data

The crop developmental stage represents essential information for irrigation scheduling/fertilizer management, understanding seasonal ecosystem carbon dioxide (CO₂) exchange, and evaluating crop productivity. In this study, we devised an approach called the Two-Step Filtering (TSF) for detecting the phenological stages of maize and soybean from time-series Wide Dynamic Range Vegetation Index (WDRVI) data derived from Moderate Resolution Imaging Spectroradiometer (MODIS) 250-m observations. The TSF method consists of a Two-Step Filtering scheme that includes: (i) smoothing the temporal WDRVI data with a wavelet-based filter and (ii) deriving the optimum scaling parameters from shape-model fitting procedure. The date of key crop development stages are then estimated by using the optimum scaling parameters and an initial value of the specific phenological date on the shape model, which are preliminary defined in reference to ground-based crop growth stage observations. The shape model is a crop-specific WDRVI curve with typical seasonal features, which were defined by averaging smoothed, multi-year WDRVI profiles from MODIS 250-m data collected over irrigated maize and soybean study sites.In this study, the TSF method was applied to MODIS-derived WDRVI data over a 6-year period (2003 to 2008) for two irrigated sites and one rainfed site planted to either maize or soybean as part of the Carbon Sequestration Program (CSP) at the University of Nebraska-Lincoln. A comparison of satellite-based retrievals with ground-based crop growth stage observations collected by the CSP over the six growing seasons for these three sites showed that the TSF method can accurately estimate the date of four key phenological stages of maize (V2.5: early vegetative stage, R1: silking stage, R5: dent stage and R6: maturity) and soybean (V1: early vegetative stage, R5: beginning seed, R6: full seed and R7: beginning maturity). The root mean square error (RMSE) of phenological-stage estimation for maize ranged from 2.9 [R1] to 7.0 [R5] days and from 3.2 [R6] to 6.9 [R7] days for soybean, respectively. In addition, the TSF method was also applied for two years (2001 and 2002) over eastern Nebraska to test its ability to characterize the spatio-temporal patterns of these key phenological stages over a larger geographic area. The MODIS-derived crop phenological stage dates agreed well with the statistical crop progress data reported by the United State Department of Agriculture (USDA) National Agricultural Statistics Service (NASS) for eastern Nebraska's three crop agricultural statistic districts (ASDs). At the ASD-level, the RMSE of phenological-stage estimation ranged from 1.6 [R1] to 5.6 [R5] days for maize and from 2.5 [R7] to 5.3 [R5] days for soybean.     

分段2维主成分分析的超光谱图像波段选择

目的 超光谱图像具有极高的谱间分辨率,巨大的数据量给分类识别等后续处理带来很大压力。为了有效降低图像数据维数,提出基于分段2DPCA的超光谱图像波段选择算法。方法 首先根据谱间相关性对原始图像进行波段分组,然后根据主成分反映每个光谱波段的信息比重分别对每组图像进行波段选择,从而实现超光谱图像的谱间降维。结果 该算法有效降低了超光谱图像的光谱维数,选择的波段明显反映出不同地物像元矢量的区别。结论 实验结果表明,该波段选择算法相对传统算法速度更快,并且较好地保留了原始图像的局部重要信息,对后续处理有积极意义。     

基于蚁群优化算法的遥感器波段设置探究

卫星载荷研制发射后其光谱和空间观测模式固定,无法根据复杂地表的多样化需求进行实时灵活调整,且目前遥感器波段设置尚不完善还存在优化空间.引进基于蚁群优化算法的波段选择方法(AntColonyOptimization basedBandSelection,ACOBS),结合北美区域３３景AVIRIS航空高光谱图像,开展了不同区域、不同地表覆盖类型的高光谱波段优选研究,发现各地表类型优选波段组合存在一定差异,其中４波段组合中红光、近红外波段为２个共同入选波段,６波段组合中绿光、红光、短波红外波段为３个共有波段,８波段组合中紫光、绿光、红光、红边、近红外１、近红外２、短波红外１、短波红外２为８个共有入选波段,其他入选波段与地表覆盖类型有关.在此基础上,进一步开展了多光谱卫星波段设置评价研究,发现:４波段优化方案中,绿光、红光、近红外波段１ (７７０~８９５nm)、近红外波段２(９００~１３５０nm)为最优波段组合;６波段优化方案中,绿、红、红边、近红外１(７７０~８９５nm)、近红外２(９００~１３５０nm)、短波红外１(１５６０~１６６０nm)为最优波段组合;８波段优化方案中,蓝、绿、红、红边、近红外１(７７０~８９５nm)、近红外２(９００~１３５０nm)、短波红外１(１５６０~１６６０nm)和短波红外２(２１００~２３００nm)为最优波段组合.研究结果表明Land satTM OLI、SPOT等陆地资源遥感器波段设置还存在一定优化调整空间,特别是红边波段在目前传感器波段设置中没有得到足够重视.     

Hyperspectral band selection with objective image quality assessment

Hyperspectral sensors often collect hundreds of bands at a time, so hyperspectral images can accurately characterize different land-cover types with abundant spectral information. However, these spectral bands also contain redundant information that needs to be removed. Band selection is one of the most widely used methods to remove noised or redundant bands. Because labelled samples are difficult to collect, most band selection methods adopt unsupervised ways to select diverse and representative bands. Still, noised bands are often selected because they usually have low correlation with other bands. In this article, objective image quality assessment is introduced to indicate the quality of every band, and combined with the redundancy measure, a new unsupervised band selection method is proposed. Three real hyperspectral images are used to demonstrate the effectiveness of the proposed algorithm.