Unsupervised change detection in season-varying images remains challenging due to the pseudo-changes induced by the seasonal variation of the images acquired at different times. Although various transformations [e.g., multivariate alteration detection (MAD), slow feature analysis (SFAs), etc.] have been developed to alleviate this issue, the complicated seasonal variation in a heterogenous landscape cannot be well addressed. This study developed a novel “difference-in-differences”-based change detection (DIDCD) method, which calculates the change magnitude by performing two difference operations between the changes observed in the target pixel and those in the control group. The control group is established by a combined method of k-means clustering and minimum covariance determinant (MCD), which identifies similar pixels that undergo parallel seasonal changes compared to the target pixels. Through DID operation, DIDCD eliminates seasonal variation and accurately identifies land cover change. DIDCD is evaluated using four pairs of images under both season-consistent and season-varying conditions across two scenarios (urban expansion and forest fire disturbance). The results demonstrate that DIDCD effectively suppresses the pseudo-changes induced by the seasonal variation and outperforms the existing unsupervised change detection algorithms.

As the second largest oilseed crop globally, the consumption of rapeseed experiences a gradual increase due to population growth and the high energy demand. Timely and accurate rapeseed mapping is urgently needed to ensure food and security safety. Due to the lack of ground truth in the target year, transferring historical classifiers or generating samples in the current year based on historical crop rotations are commonly employed for in-season mapping. However, transfer-based methods underperform in regions with significant interannual feature variation, while sample-generation methods fail in areas lacking evident rotation patterns. To address these challenges, this paper proposes a novel method that integrates rotational prediction and historical classifiers based on classification error estimation and sample weighting. The method first generates prediction samples from crop rotations and classification samples from the classification results of historical classifiers for the target year. Subsequently, the classification error of these samples is estimated based on the assumption that crop rotation is independent of the spectra of the target year. Finally, the prediction samples and classification samples are weighted according to the estimation error to form an optimal sample set and rapeseed is mapped using the classifier trained with the sample set. Validation in two typical rapeseed planting areas demonstrates the stability of the proposed method across regions and its superiority over both transfer-based and rotation-based methods, thereby proving the effectiveness of integration.

Nighttime light (NTL) products have become emerging instruments for studying human activity patterns in various applications, including demarcating urban areas, estimating the residential population, and monitoring the economic livelihood of cities. However, the deployment of such products, particularly the Defense Meteorological Satellite Program (DMSP)/Operational Line Scan (OLS) stable NTL data, is also subject to issues of data quality and methodological biases, including interannual inconsistency, saturation and blooming effect, and, in particular, the saturation problem, posing a challenge for further applications. This study proposed a novel saturation correction method based on regression model and radiance-calibrated NTL data (SARMRC) to correct the saturation effect in annual stable NTL data. The proposed method was applied to mainland China. The results show that SARMRC can effectively recover the light intensity distribution within the saturated areas of the stable NTL data. The corrected images were proven to be more spatially consistent with the reference National Polar-orbiting Partnership (NPP)/Visible Infrared Imaging Radiometer Suite (VIIRS) data and can better reflect the temporal development of cities compared to existing saturation correction methods (e.g., the index-based vegetation adjusted NTL urban index (VANUI) method and the interpolation method). The superior performance of SARMRC can be attributed to: 1) better utilization of both radiance-calibrated and stable NTL products and 2) the employment of training pixel selection and logarithmic model as well as double-year adjustment. The new method is expected to improve the data quality of DMSP/OLS stable NTL data for analyzing both local and regional socioeconomic activities.

In recent years, ground penetrating radar (GPR) has become increasingly important as a nondestructive way to explore plant roots. Automatic recognition and localization of root objects from GPR images presents a significant challenge. GPR images for the root system contain complicated hyperbolic signals that appear deformation depending on root size, orientation, aggregation degree and soil background. This paper presents a new deep learning approach, YOLOv4-hyperbola, that provides fully automatic recognition and localization of root objects from GPR images. YOLOv4-hyperbola improves the YOLOv4 (You Only Look Once v4) architecture by introducing keypoints detection branch in order to accurately locate roots while identifying them. The YOLOv4-hyperbola model was trained by combining field datasets and simulated datasets to simultaneously identify and locate hyperbolic features representing potential root objects across GPR images, and evaluated on datasets of root detection from two experiments in the field. Compared with Randomized Hough transform (RHT) method, the proposed approach demonstrated higher accuracy and efficiency in root object detection on GPR image. YOLOv4-hyperbola was able to accurately recognize and locate abnormal hyperbolic signals caused by the complexity of root system in nature. The validation on the two independent datasets showed that the proposed approach had good generalization and great application potential for real-time detection and location of roots over large areas in the field.

Deep semantic segmentation networks perform better in cloud detection of satellite imagery than traditional methods due to their ability to extract high-level features over a large receptive field. However, a large receptive field often leads to loss of spatial details and blurring of boundaries. Therefore, it is crucial to understand the role of the receptive field on the segmentation results, which has rarely been investigated for cloud detection tasks. This study, for the first time, explored the relationship between the receptive field size and the performance of a cloud detection network. Six typical networks commonly used for cloud detection and nine modified UNet variants with different depths, dilated convolutions, and skip connections were evaluated based on the Landsat 8 Biome (L8 Biome) dataset. The theoretical receptive field (TRF) and the effective receptive field (ERF) were introduced to measure the receptive field sizes of different networks. The results revealed a negative correlation between the ERF size and cloud segmentation accuracies for different cloud distributions and a relatively weak negative correlation between the TRF size and segmentation accuracies. Furthermore, ERFs were considerably smaller than the corresponding TRFs for most networks, implying that large-scale contextual information was not learned after training. This result indicates the importance of using networks with a small receptive field for cloud detection of Landsat 8 OLI imagery. Moreover, as the boundary accuracies are significantly lower than the region accuracies, future efforts should be devoted to addressing inaccurate boundary localization rather than exploring the contextual information over a large receptive field.

Hyperspectral image classification is a challenging task when a limited number of training samples are available. It is also known that the classification performance highly depends on the quality of the labeled samples. In this work, a cluster-based conditional generative adversarial net (CCGAN) is proposed as an effective solution to increase the size and quality of the training data set. The proposed method is able to automatically select the most representative initial samples with a subtractive clustering-based strategy, which keeps the diversity for sample generation. Moreover, compared to the traditional semisupervised classification frameworks, the CCGAN is able to generate realistic spectral profiles by considering the class-specific labels. Experiments on well-known Pavia University data set demonstrate that the proposed CCGAN can significantly boost the classification accuracy, even using a small number of initial labeled samples.

Spatiotemporal data fusion is a methodology to generate images with both high spatial and temporal resolution. Most spatiotemporal data fusion methods generate the fused image at a prediction date based on pairs of input images from other dates. The performance of spatiotemporal data fusion is greatly affected by the selection of the input image pair. There are two criteria for selecting the input image pair: the “similarity” criterion, in which the image at the base date should be as similar as possible to that at the prediction date, and the “consistency” criterion, in which the coarse and fine images at the base date should be consistent in terms of their radiometric characteristics and imaging geometry. Unfortunately, the “consistency” criterion has not been quantitatively considered by previous selection strategies. We thus develop a novel method (called “cross-fusion”) to address the issue of the determination of the base image pair. The new method first chooses several candidate input image pairs according to the “similarity” criterion and then takes the “consistency” criterion into account by employing all of the candidate input image pairs to implement spatiotemporal data fusion between them. We applied the new method to MODIS-Landsat Normalized Difference Vegetation Index (NDVI) data fusion. The results show that the cross-fusion method performs better than four other selection strategies, with lower average absolute difference (AAD) values and higher correlation coefficients in various vegetated regions including a deciduous forest in Northeast China, an evergreen forest in South China, cropland in North China Plain, and grassland in the Tibetan Plateau. We simulated scenarios for the inconsistency between MODIS and Landsat data and found that the simulated inconsistency is successfully quantified by the new method. In addition, the cross-fusion method is less affected by cloud omission errors. The fused NDVI time-series data generated by the new method tracked various vegetation growth trajectories better than previous selection strategies. We expect that the cross-fusion method can advance practical applications of spatiotemporal data fusion technology.

In the scenario of mobile fog computing (MFC), communication between vehicles and fog layer, which is called vehicle-to-fog (V2F) communication, needs to use bandwidth resources as much as possible with low delay and high tolerance for errors. In order to adapt to these harsh scenarios, there are important technical challenges concerning the combination of network coding (NC) and multipath transmission to construct high-quality V2F communication for cloud-aware MFC. Most NC schemes exhibit poor reliability in burst errors that often occur in high-speed movement scenarios. These can be improved by using interleaving technology. However, most traditional interleaving schemes for multipath transmission are designed based on round robin (RR) or weighted round robin (WRR), in practice, which can waste a lot of bandwidth resources. In order to solve those problems, this paper proposes a novel multipath transmission scheme for cloud-aware MFC, which is called Bidirectional Selection Scheduling (BSS) scheme. Under the premise of realizing interleaving, since BSS can be used in conjunction with a lot of path scheduling algorithms based on Earliest Delivery Path First (EDPF), it can make better use of bandwidth resources. As a result, BSS has high reliability and bandwidth utilization in harsh scenarios. It can meet the high-quality requirements of cloud-aware MFC for transmission.

Vegetation phenology has an important impact on ecosystem processes and biosphere-atmosphere feedbacks, which is also highly sensitive to the climate change. The observation and extraction of Phenology require the highest possible temporal resolution and spatial resolution. Digital camera photographs can supplement long-term observations, which meet the above conditions. The usual treatment for digital repeat RGB-cameras is to compute compared greenness chromatic coordinate (GCC) series representing the state of vegetation growth. In order to eliminate the effect of solar angle and atmospheric scattering on GCC series, we used the histogram features of the R, G, and B band of each photo to obtain a stable GCC time series through multiple linear regression. The DART model was used to simulate winter wheat scenes under different light and atmospheric conditions. Simulation experiments show that this method can eliminate the instability caused by atmospheric scattering, but the correction of the solar elevation angle is not obvious.

Arid and semiarid grasslands around the world are heavily suffering from shrub encroachment. However, the shrub encroachment process is still poorly understood, especially when historical data field samples are insufficient. High-spatial-resolution (HSR) remotely sensed data provides the only large-scale historical record of grasslands. This study investigated the population structure of Caragana genus plants, a typical shrub in the temperate arid and semiarid grasslands of Inner Mongolia, China. Shrub individuals were identified from HSR images and the age of shrub individuals were estimated, then a caragana population age structure which can calculate mortality and regeneration was established with a negative exponential function.

Due to its in situ and nondestructive nature, ground-penetrating radar (GPR) has recently been applied to the field investigation of plant roots. The discrepancy between the roots and surrounding soils creates a dielectric constant contrast, forming clear hyperbolic reflections on the GPR radargram. The intensity and shape of the reflecting signals from roots are substantially affected by the root orientation as well as the relative geometry between the root in the subsurface and the GPR survey direction on the ground surface. However, no previous study has utilized the information on the intensity and shape of a root's GPR reflection to map its orientation, which is crucial in interpreting radargrams and rebuilding 3-D root system architecture. In this paper, a mathematical formulation of hyperbolic reflection formed by a single root was first deduced based on the principles of electromagnetic wave propagation. Then, using this formulation, curve fitting was conducted on both simulated and field collected data sets by GPR. Information on the horizontal orientation and vertical inclination of a single root was acquired according to the formulation coefficient retrievals. Conditions for this method of application and factors impacting the extraction of root orientation information were analyzed. The results indicated fairly precise root orientation estimations. The proposed method has extended the application of GPR in root investigation, thus advancing the frontier of noninvasive root system architecture mapping.

Remote estimation of fractional vegetation cover (FVC) in arid and semiarid areas is crucial for understanding their roles in global climate changes and maintaining their ecological sustainability. Among the existing algorithms for remote estimation of FVC, the linear spectral mixture analysis (LSMA) has been widely adopted owing to its simplicity and flexibility. However, the spectral variability of endmembers is still a big challenge that would largely decrease the estimation accuracy of LSMA. In this letter, we proposed a novel unmixing algorithm by integrating an orthogonal Fisher transformation into the LSMA (fLSMA). Two evaluation experiments were conducted: one was based on simulations; the other was based on a field survey in Xilingol grassland, China. The proposed fLSMA yielded remarkably higher accuracies and precisions than the conventional LSMA (cLSMA), weighted SMA (wSMA) in the first experiment. In the second experiment, a root-mean-square error (RMSE) of 0.11 was derived for the fLSMA, compared with the RMSE values larger than 0.36 for the cLSMA and wSMA. Although the performance of fLSMA was somehow similar to the multiple endmember SMA (MESMA) in the two evaluation experiments, the fLSMA was much less time-consuming than the MESMA in massive computations. The results indicate the potential of the proposed fLSMA in long-term monitoring of FVC in semiarid areas based on satellite observations.

Cloud removal is significantly needed for enhancing the further utilization of Landsat imagery, since such optical remote sensing satellite images are inevitably contaminated by clouds. Clouds dynamically affect the signal transmission due to their different shapes, heights, and distribution. Generally, pixel replacement is the only and common method used to remove thick opaque clouds, and radiometric correction techniques has been widely adopted to remove the thin clouds. However, no methods can remove both thick and thin clouds at the same time. In this paper, a new method is proposed based on fitting “trajectory” of cloudy pixels with the help of IHOT spatially charactering clouds for pixel correction, which considers signal transmission including not only the additive reflectance from the clouds but also the energy attenuation when solar radiation passes through them. The experimental results show that the proposed approach performs effective removal for thick and thin clouds, and possesses the highest accuracy with the reference image, which can restore land cover information accurately.

Endmember variability associated with impervious layer has been a serious problem in spectral mixture analysis (SMA). A reliable spectral library which ideally models the endmember variability is required for precise SMA. Even though many endmember bundles extraction algorithms have been proposed, there are still some problems in these methods which blur the threshold and endmember numbers. In this paper, an iterative classification extraction endmember bundles algorithm (ICEEA) is proposed. Impervious and pervious training sample are provided with GlobeLand30 product, and Maximum Likelihood Classifier (MLC) is used to conduct iteratively classification. After each classification, the artificial layer pixels which are misclassified as pervious class are excluded from artificial cover, and the impervious sample is selected again in new artificial cover. It stops when there are none misclassified pixels existing in the artificial layer. According to the results of simulated 30m data and real TM data, ICEEA has two advantages over PPI: (1) producing more reliable impervious endmember bundles which can model the endmember variability well; (2) having none threshold setting problem; (2) running much faster than PPI.

Most previous haze/cloud detection methods for Landsat imagery, e.g., haze optimized transformation (HOT), cannot adequately suppress land surface information and, in particular, often overestimate haze thickness over bright surfaces. This paper proposes an iterative HOT (IHOT) for improving haze detection with the help of a corresponding clear image. With an iterative procedure of regressions among HOT, the reflectance difference at the top of atmosphere (TOA) between hazy and clear images, and TOA reflectances of hazy and clear images, the land surface information can be removed, and the iterative HOT (IHOT) result is derived to spatially characterize the haze contamination in the Landsat images. A group of Landsat images that were acquired in different landscapes and seasons were used to test IHOT. Visual comparisons indicate that IHOT performed better than previous haze detection methods for images that were acquired in diverse landscapes and also performed robustly for hazy images that were acquired at different seasons when using the same reference clear image. Additionally, two indirect quantitative validations were used to illustrate that IHOT can provide the best transformation for accurately determining haze information. Therefore, it is expected that the proposed IHOT method will be used for automatic cloud/haze detection for large numbers of Landsat images if data sets of clear Landsat imagery are available.

Spectral mixture analysis (SMA) is widely used to quantify the fraction of each component (endmember) of mixed pixels that contain spectral signals from more than one land surface type. Generally, nonlinear SMA (NSMA) outperforms linear SMA (LSMA) in the vegetation (tree, shrub, crop, and grass) and soil mixture case because NSMA considers the significant multiple scattering that exists for these mixtures. However, compared to LSMA, the bilinear NSMA method, which is a typical physical-based NSMA method, is undermined by its susceptibility to the collinearity effect. In this paper, a two-step constrained NSMA method (referred to as TsC-NSMA) is proposed to mitigate the collinearity effect in the bilinear NSMA method. The theoretical maximum likelihood range is mathematically derived for each endmember fraction, and the ranges are used as additional constraints for the bilinear NSMA method to optimize the unmixing results. Three different data sets, including simulated spectral data, an in situ ground plot spectral measurement, and a Landsat8 Operational Land Imager image, were used to assess the performance of the TsC-NSMA method. The results indicated that TsC-NSMA achieved the highest estimation accuracy for all mixed scenarios which either contain severe endmember collinearity or high noise levels, thereby suggesting its ability to mitigate the collinearity effect in the bilinear NSMA method with the potential to improve the estimation of endmember fractions in practical applications.

Positive and unlabeled learning (PUL) algorithm, an one-class classifier which is trained by positive samples and unlabeled samples, has been used in remote sensing classification. However, the effect of training strategy of PUL has not been investigated. This study tested the performances of PUL-SVM on cropland mapping by Landsat TM data using the training samples with different sizes and different purity levels. It is found that the highest accuracy is achieved when the sizes of positive sample and unlabeled sample are comparable if using the random strategy. In contrast, if using the purer positive samples, it is more difficult to find the optimal unlabeled sample size. Therefore, it is recommended the random strategy for the positive samples, and the balanced sizes for positive and unlabeled samples when using PUL-SVM.

Plant self-thinning power law is regarded as an essential regulation which plays an important role in determining population dynamics and community structure. However, little experimental and theoretical studies on the power law have been developed for shrub, and also little has been done from the perspective of below-ground plant parts to reveal the mechanisms underlying the self-thinning process of natural plant especially in field. Taking the shrub, Caragana microphylla, as example, this study revealed the below-ground root biomass-density relationship of Caragana microphylla and explained how and to what extent do below-ground root competition affect the above-ground self-thinning process. The below-ground root systems were surveyed by Ground- penetrating radar (GPR). The root biomass was estimated based on GPR data. Some of the statistical relationships discussed in this study may be useful for predicting root biomass of shrub community in water-limited ecosystems and will also serve as a frame of reference for future studies.

Fractional vegetation cover (FVC) is one of the fundamental parameters for characterizing terrestrial ecosystems, with wide uses in various environmental and climate-related modeling applications. The remote sensing technique provides a unique opportunity for estimating FVC over large geographical areas by employing spectral mixture analysis (SMA). The effectiveness of SMA depends largely on the accurate extraction of representative and pure endmembers. However, in arid and semiarid environments that have sparse vegetation distributions, most current SMA models may produce large biases due to difficulties in obtaining pure vegetation spectra from the satellite images. This letter developed a new approach to estimate FVC from satellite observations by integrating an endmember spectrum purification procedure into a nonlinear SMA model. The proposed method is capable of extracting pure endmember spectra even though pure vegetation endmember is not present in target images in arid and semiarid environments, which improves the accuracy of FVC retrievals. Validation experiments conducted in the Xilingol grassland, Inner Mongolia, China, demonstrate that the proposed method produces more accurate FVC estimates (RMSE <; 0.13, AD <; 0.06) than do current algorithms. The better performance of the proposed method can be attributed to the purified vegetation spectra that more closely resemble the real pure vegetation spectra.

Spectral mixture analysis is an important technique to extract desired information from the mixed remotely sensed data. However, current spectral mixture analysis techniques suffered from the endmember variability. Quantitative assessment of SMA techniques with simulated data is critical to understand the influence of endmember variability. For that reason, this study has compared five typical spectral mixture analysis addressing endmember variability issue with simulated data. The comparison result shows that MESMA seems to be the best in unmixing accuracy. However, sensitive to noise and large computation loads also made MESMA less satisfactory, while other methods could supersede MESMA at specific situations.

The Fraunhofer line discriminator (FLD) principle applied on the atmospheric oxygen absorption feature around 761 nm ( O2-A band) has been widely used to retrieve solar-induced chlorophyll fluorescence (Fs) from remotely sensed data. In this letter, however, we address a violation of the basic assumption caused by O2absorption feature changes, and we evaluate the impact of diminishing O2absorption with the increase in ground altitude on Fsretrieval accuracy. The Fs retrieval accuracy substantially decreases in higher ground altitude areas for the standard FLD and three-band FLD methods, in which relative estimation errors increase approximately 70%-80% and 10%-16%, respectively, with an increase in ground altitude from 0.01 to 4.5 km. However, this increasing trend in Fs retrieval error does not occur with the use of the improved FLD (iFLD) method, which exhibits smaller than 5% of changes in relative estimation errors. Analytical analyses reveal the causes of the changes in Fs retrieval accuracy by the three methods. Based on these findings, the iFLD method is recommended to be used in cross-altitude studies or for Fs estimation in higher ground altitude areas under conditions of low radiometric noise, although its high sensitivity to noise should be taken into account. The investigations in this letter further indicate that the impact of ground altitude should be included in the uncertainty budgets of Fs retrievals and should be considered in the interpretation of Fs signals at the O2-A absorption feature.

Roots have important functions in the ecosystem. Therefore, establishing root-related parameters such as root size, biomass, and 3-D architecture is necessary. Traditional methods for measuring tree roots are labor intensive and destructive to nature, limiting quantitative and repeated assessments in long-term research. Ground-penetrating radar (GPR) provides a nondestructive method for measuring tree roots. This study investigates the feasibility of a GPR system with 500-MHz, 900-MHz, and 2-GHz measurement frequencies for detecting tree roots and estimating root biomass under controlled experimental conditions in a sandy area. After energy attenuation correction and velocity analysis, not only the individual root in subsurface is able to be located but also the parameters that correlate well with root biomass can be extracted from the processed GPR data. The major findings were as follows. First, both the amplitude and amplitude-area indices were confirmed to be more effective for estimating root biomass after attenuation-effect compensation. This result suggests that the calibration of GPR wave-attenuation effects and velocity changes with depth are helpful in estimating root biomass from GPR parameters. Second, the selection of GPR system frequency was mainly dependent on field conditions, particularly soil water content. Lower frequency was recommended for developing root biomass estimation model under varied soil conditions. Third, the new method based on the metal reflector experiment was effective and easy to perform in situ for attenuation-effect correction.

Vegetation-index-based spatial sharpening technologies were developed for improving the spatial resolution of thermal-infrared (TIR) images. Previous studies showed that the relationship between vegetation index and surface temperature is independent with spatial resolution. However, spatial extent is another scale factor which may affect the relationship of vegetation index and surface temperature but was neglected in the previous studies. In this letter, we investigated both of these two aspects of scale effect (spatial resolution and spatial extent) of the relationship of vegetation index and surface temperature based on two scenes of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) data respectively acquired in grassland and crop-land. The result shows that the relationship of vegetation index and temperature on different spatial extents varies greatly. Therefore, previous methods which apply the relationship established on the whole image with coarse resolution to the small area (pixel size of the TIR image) with fine resolution may produce a large error. We modified the original sharpening method by establishing the vegetation-index-temperature relationship on the minimum available spatial extent (2 × 2 pixels of the TIR image) instead of the whole image. Then, the new method was tested and compared with the original method by simulated data. The result shows that the new method performs more robustly than the original method.

Vegetation index based thermal sharpening technologies have been studied recently. Previous studies showed that the relationship between vegetation index and surface temperature is independent with spatial resolution; consequently, the relationship established on coarse resolution can be applied on fine resolution. However, in this study, we found that the resolution independence can not grant the effectiveness of the sharpening method. Instead, relationship of vegetation index temperature established on the correct spatial size (size of thermal resolution) should be used for sharpening temperature image. Unfortunately, the relationship of vegetation index temperature on different spatial sizes varies greatly. Therefore, previous methods which apply the relationship established on the whole image to the small spatial size (size of thermal resolution) may produce large error. In this paper, an improved vegetation index based sharpening method considering the effect of spatial size was proposed and tested by a simulation study. The result shows that it can acquire higher sharpening accuracy than original method.

Postclassification comparison (PCC) and change vector analysis (CVA) have been widely used for land use/cover change detection using remotely sensed data. However, PCC suffers from error cumulation stemmed from an individual image classification error, while a strict requirement of radiometric consistency in remotely sensed data is a bottleneck of CVA. This letter proposes a new method named CVA in posterior probability space (CVAPS), which analyzes the posterior probability by using CVA. The CVAPS approach was applied and validated by a case study of land cover change detection in Shunyi District, Beijing, China, based on multitemporal Landsat Thematic Mapper data. Accuracies of “change/no-change” detection and “from-to” types of change were assessed. The results show that error cumulation in PCC was reduced in CVAPS. Furthermore, the main drawbacks in CVA were also alleviated effectively by using CVAPS. Therefore, CVAPS is potentially useful in land use/cover change detection.

In the past decades, spectral unmixing has been studied for deriving the fractions of spectrally pure materials in a mixed pixel. However, limited attention has been given to the collinearity problem in spectral mixture analysis. In this paper, quantitative analysis and detailed simulations are provided, which show that the high correlation between the endmembers, including the virtual endmembers introduced in a nonlinear model, has a strong impact on unmixing errors through inflating the Gaussian noise. While distinctive spectra with low correlations are often selected as true endmembers, the virtual endmembers formed by their product terms can be highly correlated. It is found that a virtual-endmember-based nonlinear model generally suffers more from collinearity problems compared to linear models and may not perform as expected when the Gaussian noise is high, despite its higher modeling power. Experiments were conducted on a set of in situ measured data, and the results show that the linear mixture model performs better in 61.5% of the cases.