Sub-pixel Land Cover Research Articles

Assessing a Temporal Change Strategy for Sub-Pixel Land Cover Change Mapping from Multi-Scale Remote Sensing Imagery

Remotely sensed imagery is an attractive source of information for mapping and monitoring land cover. Fine spatial resolution imagery is typically acquired infrequently, but fine temporal resolution systems commonly provide coarse spatial resolution imagery. Sub-pixel land cover change mapping is a method that aims to use the advantages of these multiple spatial and temporal resolution sensing systems. This method produces fine spatial and temporal resolution land cover maps, by updating fine spatial resolution land cover maps using coarse spatial resolution remote sensing imagery. A critical issue for sub-pixel land cover change mapping is downscaling coarse spatial resolution fraction maps estimated by soft classification to a fine spatial resolution land cover map. The relationship between a historic fine spatial resolution map and a contemporary fine spatial resolution map to be estimated at a more recent date plays an important role in the downscaling procedure. A change strategy based on the assumption that the change for each land cover class in a coarse spatial resolution pixel is unidirectional was shown to be a promising means to describe this relationship. This paper aims to assess this change strategy by analyzing the factors that affect the accuracy of the change strategy, using six subsets of the National Land Cover Database (NLCD) of USA. The results show that the spatial resolution of coarse pixels, the time interval of the previous fine resolution land cover map and the current coarse spatial resolution images, and the thematic resolution of the used land cover class scheme have considerable influence on the accuracy of the change strategy. The accuracy of the change strategy decreases with the coarsening of spatial resolution, an increase of time interval, and an increase of thematic resolution. The results also indicate that, when the historic land cover map has a 30 m resolution, like the NLCD, the average accuracy of the change strategy is still as high as 92% when the coarse spatial resolution data used had a resolution of ~1000 m, confirming the effectiveness of the change strategy used in sub-pixel land cover change mapping for use with popular remote sensing systems.

Partially and Fully Constrained Least Squares Linear Spectral Mixture Models for Subpixel Land Cover Classification Using Landsat Data

Partially and Fully Constrained Least Squares Linear Spectral Mixture Models for Subpixel Land Cover Classification Using Landsat Data

The effect of imposing ‘fractional abundance constraints’ onto the multilayer perceptron for sub-pixel land cover classification

To be physically interpretable, sub-pixel land cover fractions or abundances should fulfill two constraints, the Abundance Non-negativity Constraint (ANC) and the Abundance Sum-to-one Constraint (ASC). This paper focuses on the effect of imposing these constraints onto the MultiLayer Perceptron (MLP) for a multi-class sub-pixel land cover classification of a time series of low resolution MODIS-images covering the northern part of Belgium. Two constraining modes were compared, (i) an in-training approach that uses ‘softmax’ as the transfer function in the MLP’s output layer and (ii) a post-training approach that linearly rescales the outputs of the unconstrained MLP. Our results demonstrate that the pixel-level prediction accuracy is markedly increased by the explicit enforcement, both in-training and post-training, of the ANC and the ASC. For aggregations of pixels (municipalities), the constrained perceptrons perform at least as well as their unconstrained counterparts. Although the difference in performance between the in-training and post-training approach is small, we recommend the former for integrating the fractional abundance constraints into MLPs meant for sub-pixel land cover estimation, regardless of the targeted level of spatial aggregation.

An Improved Endmember Selection Method Based on Vector Length for MODIS Reflectance Channels

Endmember selection is the basis for sub-pixel land cover classifications using multiple endmember spectral mixture analysis (MESMA) that adopts variant endmember matrices for each pixel to mitigate errors caused by endmember variability in SMA. A spectral library covering a large number of endmembers can account for endmember variability, but it also lowers the computational efficiency. Therefore, an efficient endmember selection scheme to optimize the library is crucial to implement MESMA. In this study, we present an endmember selection method based on vector length. The spectra of a land cover class were divided into subsets using vector length intervals of the spectra, and the representative endmembers were derived from these subsets. Compared with the available endmember average RMSE (EAR) method, our approach improved the computational efficiency in endmember selection. The method accuracy was further evaluated using spectral libraries derived from the ground reference polygon and Moderate Resolution Imaging Spectroradiometer (MODIS) imagery respectively. Results using the different spectral libraries indicated that MESMA combined with the new approach performed slightly better than EAR method, with Kappa coefficient improved from 0.75 to 0.78. A MODIS image was used to test the mapping fraction, and the representative spectra based on vector length successfully modeled more than 90% spectra of the MODIS pixels by 2-endmember models.

Subpixel land cover change mapping with multitemporal remote-sensed images at different resolution

Due to the lack of support for a high-resolution image in a short time, land cover change detection is always applied on the multitemporal remote-sensed images with different resolutions. The coarse-resolution image contains a large number of mixed pixels, which can seriously limit the utility of the change detection. Soft classification (SC) can be applied to improve this situation through deriving the abundances and generating the fractional change map, but it cannot provide the spatial distribution of the subpixels. Subpixel mapping (SPM) is a potential solution to resolve this problem, and is designed to use the proportions to obtain a sharpened thematic map at a subpixel scale. Based on this thought, the subpixel scale land cover change mapping result can be realized by integrating these two key techniques. However, in practice, there is a serious limitation to the detail and accuracy of the result, because when the scale factor between the different resolution images is large, the subpixel configuration is complex and the data volumes will be amplified. Moreover, with the high proportion of the changed area in the whole image, the change detection process at the subpixel level gets more difficult. The SPM technique is generally performed based only on the abundances of each and the spatial dependence assumption, so it cannot satisfy the demand. In order to overcome this shortcoming, several new reasonable subpixel scale change detection rules are defined in this paper, which dictate the land cover change map must be constructed according to the existing high-resolution image. The output from SC and prior information of the subpixel feature arrangement is applied into a modified cellular automata (CA) model, which can be regarded as a more reasonable tool to monitor the subpixel changes to resolve the big-data problem. Two experiments are performed and the results prove that the proposed method can effectively improve the accuracy of the change detection maps of the spatial distribution in a subpixel scale.

On quantifying post-classification subpixel landcover changes

The post-classifications change matrix is well defined for hard classifications. However, for soft classifications, where partial membership of pixels to landcover classes is allowed, there is no single definition for such a matrix. In this paper, we argue that a natural definition of the post-classification change matrix for subpixel classifications can be done in terms of a constrained optimization problem, according to which the change matrix should allow an optimal prediction of the subpixel landcover fractions at the latest date from those of the earliest date. We first show that the traditional change matrix for crisp classification corresponds to the optimal solution of the unconstrained problem. Then, the formulation is generalized for subpixel classifications by incorporating certain constraints pertaining to desirable properties of a change matrix, thus resulting in a constrained least square (CLS) change matrix. In addition, based on intuitive criteria, a generalized product (GPROD) was parameterized in terms of an exponent parameter and used to derive another change matrix. It was shown that when the exponent parameter of the GPROD operator tends to infinity, one of the most commonly used methods for map comparison from subpixel fractions, namely the MINPROD composite operator, results. The three matrices (CLS, GPROD and MINPROD) were tested on both simulated and real subpixel changes derived from QuickBird and Landsat TM images. Results indicated that, for small exponent values (0–0.5), the GPROD matrix yielded the lowest errors of estimated landcover changes, whereas the MINPROD generally yielded the highest errors for the same estimations.

A New Super Resolution Mapping Algorithm by Combining Pixel and Subpixel-Level Spatial Dependences With Colorimetry

Super resolution mapping is a continuously growing area of remote sensing. Satellite images coupled with a very high spectral resolution, and are suitable for detection and classification of surfaces and different elements in the observed image. The main problem with high resolution data for these applications is the (relatively) low spatial resolution, which can vary from a few to tens of meters. In the case of classification purposes, the major problem caused by low spatial resolution is related to subpixels, i.e., pixels in the image where more than one land cover class is within the same pixel. In such a case, the pixel cannot be considered as belonging to just one class, and the assignment of the pixel to a single class will inevitably lead to a loss of information, no matter what class is chosen. A new super resolution mapping (SRM) algorithm by combining pixel and subpixel-level spatial dependences with colorimetry is proposed in this paper. The pixel-level dependence is measured by the spatial attraction model with either surrounding or quadrant neighborhood, while the subpixel-level dependence is characterized by either the mean filter or the exponential weighting function. Both pixel-level and subpixel-level dependences are then fused as the weighted dependence for quickly obtaining the optimal spatial distribution of subpixels by employing the colorimetric algorithm. Synthetic imagery and a QuickBird image are tested for validation of the proposed method. The results demonstrate that the proposed method can achieve results with greater accuracy than two traditional subpixel mapping (SPM) methods and the mixed spatial attraction model method. Meanwhile, the proposed method needs considerably less computation time than the conventional mixed spatial attraction model method, and hence it provides a new solution to subpixel land cover mapping.

A subpixel mapping algorithm combining pixel-level and subpixel-level spatial dependences with binary integer programming

A new subpixel mapping (SPM) algorithm combining pixel-level and subpixel-level spatial dependences is proposed in this letter. The pixel-level dependence is measured by the spatial attraction model (SAM) with either surrounding or quadrant neighbourhood, while the subpixel-level dependence is characterized by either the mean filter or the exponential weighting function. Both pixel-level and subpixel-level dependences are then fused as the weighted dependence in the constructed objective function. The branch-and-bound algorithm is employed to solve the optimization problem, and thus, obtain the optimal spatial distribution of subpixel classes. An artificial image and a set of real remote sensing images were tested for validation of the proposed method. The results demonstrated that the proposed method can achieve results with greater accuracy than two traditional SPM methods and the mixed SAM method. Meanwhile, the proposed method needs less computation time than the mixed SAM, and hence it provides a new solution to subpixel land cover mapping.

A Comparison of Advanced Regression Algorithms for Quantifying Urban Land Cover

Quantitative methods for mapping sub-pixel land cover fractions are gaining increasing attention, particularly with regard to upcoming hyperspectral satellite missions. We evaluated five advanced regression algorithms combined with synthetically mixed training data for quantifying urban land cover from HyMap data at 3.6 and 9 m spatial resolution. Methods included support vector regression (SVR), kernel ridge regression (KRR), artificial neural networks (NN), random forest regression (RFR) and partial least squares regression (PLSR). Our experiments demonstrate that both kernel methods SVR and KRR yield high accuracies for mapping complex urban surface types, i.e., rooftops, pavements, grass- and tree-covered areas. SVR and KRR models proved to be stable with regard to the spatial and spectral differences between both images and effectively utilized the higher complexity of the synthetic training mixtures for improving estimates for coarser resolution data. Observed deficiencies mainly relate to known problems arising from spectral similarities or shadowing. The remaining regressors either revealed erratic (NN) or limited (RFR and PLSR) performances when comprehensively mapping urban land cover. Our findings suggest that the combination of kernel-based regression methods, such as SVR and KRR, with synthetically mixed training data is well suited for quantifying urban land cover from imaging spectrometer data at multiple scales.

A spatial–temporal Hopfield neural network approach for super-resolution land cover mapping with multi-temporal different resolution remotely sensed images

The mixed pixel problem affects the extraction of land cover information from remotely sensed images. Super-resolution mapping (SRM) can produce land cover maps with a finer spatial resolution than the remotely sensed images, and reduce the mixed pixel problem to some extent. Traditional SRMs solely adopt a single coarse-resolution image as input. Uncertainty always exists in resultant fine-resolution land cover maps, due to the lack of information about detailed land cover spatial patterns. The development of remote sensing technology has enabled the storage of a great amount of fine spatial resolution remotely sensed images. These data can provide fine-resolution land cover spatial information and are promising in reducing the SRM uncertainty. This paper presents a spatial–temporal Hopfield neural network (STHNN) based SRM, by employing both a current coarse-resolution image and a previous fine-resolution land cover map as input. STHNN considers the spatial information, as well as the temporal information of sub-pixel pairs by distinguishing the unchanged, decreased and increased land cover fractions in each coarse-resolution pixel, and uses different rules in labeling these sub-pixels. The proposed STHNN method was tested using synthetic images with different class fraction errors and real Landsat images, by comparing with pixel-based classification method and several popular SRM methods including pixel-swapping algorithm, Hopfield neural network based method and sub-pixel land cover change mapping method. Results show that STHNN outperforms pixel-based classification method, pixel-swapping algorithm and Hopfield neural network based model in most cases. The weight parameters of different STHNN spatial constraints, temporal constraints and fraction constraint have important functions in the STHNN performance. The heterogeneity degree of the previous map and the fraction images errors affect the STHNN accuracy, and can be served as guidances of selecting the optimal STHNN weight parameters.

Unsupervised Subpixel Mapping of Remotely Sensed Imagery Based on Fuzzy C-Means Clustering Approach

Subpixel mapping (SPM) is a technique to obtain a land cover map with finer spatial resolution than the original remotely sensed imagery. An image-based SPM model that directly uses the original image data as input by integrating both the spectral and spatial information has been demonstrated as a promising SPM model. However, all existing image-based SPM models are based on a supervised approach, since the spectral term in these SPM models is composed of a supervised unmixing method. The endmembers and training samples for different land cover classes must be determined before implementing these supervised SPM algorithms. In this letter, a novel unsupervised image-based SPM model based on the fuzzy c-means (FCM) clustering approach (usFCM_SPM) was proposed. By incorporating the unsupervised unmixing criterion of the FCM clustering algorithm and the maximal land cover spatial-dependence principle, the proposed usFCM_SPM can generate a subpixel land cover map without any prior endmember information. Both synthetic multispectral image and real IKONOS image experiments demonstrate that the usFCM_SPM can generate higher accuracy subpixel land cover maps than the traditional unsupervised pixel-scale classification approaches and the unsupervised pixel-swapping model.

A Spatio–Temporal Pixel-Swapping Algorithm for Subpixel Land Cover Mapping

The aim of this letter is to present a spatio-temporal pixel-swapping algorithm (STPSA), based on conventional pixel-swapping algorithms (PSAs), in which both spatial and temporal contextual information from previous land cover maps or observed samples are well integrated and utilized to improve subpixel mapping accuracy. Unlike conventional pixel-swapping algorithms, STPSA is capable of utilizing prior information, which was previously ignored, to predict the attractiveness based on pairs of subpixels. This algorithm involves three main steps and operates in an iterative manner: 1) it predicts the maximum and minimum attractiveness of each pair of pixels; 2) ranks the swapping scores based on the attractiveness of all the pairs; and 3) swaps the locations of the pair of pixels with a maximum score to increase the objective function. Experiments with actual satellite images have demonstrated that the proposed algorithm performs better than other algorithms. In comparison, the proposed STPSA's better performance is due to the fact that prior information used in other algorithms is restricted to a percentage level rather than the real subpixel level.

Indicator Cokriging-Based Subpixel Land Cover Mapping With Shifted Images

Subpixel mapping (SPM) is a technique for predicting the spatial distribution of land cover classes in remote sensing images at a finer spatial resolution level than those of the input images. Indicator cokriging (ICK) has been found to be an effective and efficient SPM method. The accuracy of this model, however, is limited by insufficient constraints. In this paper, the accuracy of the ICK-based SPM model is enhanced by using additional information gained from multiple shifted images (MSIs). First, each shifted image is utilized to compute the conditional probability of class occurrence at any fine spatial resolution pixel (i.e., subpixel) using ICK, and a set of conditional probability maps for all classes are generated for each image. The multiple ICK-derived conditional probability maps are then integrated, according to the estimated subpixel shifts of MSI. Lastly, class allocation at the subpixel scale is implemented to produce SPM results. The proposed algorithm was tested on two synthetic coarse spatial resolution remote sensing images and a set of real Moderate Resolution Imaging Spectroradiometer (MODIS) data. It was evaluated both visually and quantitatively. The experimental results showed that more accurate SPM results can be generated with MSI than with a single observed coarse image in conventional ICK-based SPM. In addition, the accuracy of the proposed method is higher than that of the existing Hopfield neural network (HNN)-based SPM and the HNN with MSI.

Using Modified Normalized Spectral Mixture Analysis to Acquire Sub-Pixel Land Cover in a Coastal City

Spectral mixture analysis has been widely used in quantifying and monitoring urban land surface composition. Most previous research chose inland cities on flat plains as study areas, masking water before unmixing to acquire acceptable results. As an important type with distinct characteristics, coastal cities should be considered separately. This study took Xiamen as an example of mountainous coastal cities and compared four methods of unmixing using Landsat TM images, before assessing their accuracy through the visual interpretation of a SPOT 5 image. The results show that, for mountainous coastal cities, (1) NSMA (Modified normalized spectral mixture) will lead to spectral confusion between water and impervious surface; (2) Modified NSMA without masking water will lead to spectral confusion between dark mountainous vegetation and impervious surface; (3) Modified NSMA with masking both water and mountainous vegetation achieves better results; and (4) pure pixels of impervious surface and soil chosen from high-resolution images and V through PPI achieve the most accurate unmixing results.

Linear unmixing of MODIS albedo composites to infer subpixel land cover type albedos

We use a linear unmixing approach to test how land use and forestry maps, in combination with the MODIS BRDF/albedo product, can be used to estimate land cover type albedos in boreal regions. Operational land use maps from three test areas in Finland and Canada were used to test the method. The resulting endmember albedo estimates had low standard errors of the mean and were realistic for the main land cover types. The estimated albedos were fairly consistent with albedo measurements conducted with a telescope mast and pure pixel albedos. Problems with the method are the possible errors in the land cover maps, lack of good quality winter MODIS albedo composites and the mismatch between the MODIS pixels and the true observation area. The results emphasize the role of tree species as determinant of forest albedo. Comprehensive spatial and temporal measurements of land cover albedo are usually not possible with in situ mast measurements, and the spatial resolution of MODIS albedo product is often too low to allow direct comparison of pixel albedos and land cover types in areas with heterogeneous vegetation. Hence, and since local forestry maps exist for most temperate and boreal regions, we believe that the proposed method will be useful in estimating average regional land cover type albedos as well as in tracking changes in them.

Development of Variable Threshold Models for detection of irrigated paddy rice fields and irrigation timing in heterogeneous land cover

Accurate monitoring of irrigated paddy field area and irrigation timing are of a great concern in agricultural water management due to the substantial consumption of fresh water when irrigating paddy fields. Spectral threshold methods (Xiao et al., 2005) have been widely applied to detect irrigated paddy fields and irrigation timing using Moderate Resolution Imaging Spectroradiometer (MODIS) Enhanced Vegetation Index (EVI) and Land Surface Water Index (LSWI). These methods applied constant additive threshold values (T) to LSWI and compared it to EVI to detect the irrigated paddy fields. In this study, we developed Variable Threshold Models that utilized different pixel-based threshold values depending on sub-pixel land cover heterogeneity and hence, improve detection performance on distributed small-scale paddy fields. Non-irrigated sub-pixels were quantified with irrigation maps produced by Synthetic Aperture Radar (SAR) microwave images. Significant positive correlation between EVI and the sub-pixel numbers of non-irrigated area were found (r=0.87), which resulted in higher T for MODIS pixels with more non-irrigated sub-pixels. Accordingly, a Variable Threshold Model, i.e. a regression model between T and EVI, was developed. With the Variable Threshold Model, agreement rates between MODIS and SAR-based irrigated small-scale paddy field classification doubled compared with that from a fixed threshold value. In comparison with field observations, the Variable Threshold Models showed a mean error of +0.9 days, an improvement over the mean error of +2.8 days from a fixed threshold model. Combined utilization of SAR and MODIS images provides a useful tool for developing a Variable Threshold Model that can enhance accurate monitoring of irrigation dates across heterogeneous paddy field regions.

Subpixel Land Cover Mapping by Integrating Spectral and Spatial Information of Remotely Sensed Imagery

Subpixel mapping (SPM) is a technique to predict spatial locations of land cover classes within mixed pixels in remotely sensed imagery. The two-step approach first estimates fraction images by spectral unmixing and then inputs fraction images into an SPM algorithm to generate the final subpixel land cover map. A shortcoming of this approach is that the information about the credibility of fraction images is not considered. In this letter, we proposed a general framework of SPM which is directly applied to original coarse resolution remotely sensed imagery by integrating spectral and spatial information. Based on the proposed framework, the linear unmixing model and the maximal spatial dependence model were combined to construct a novel SPM model aiming to minimize the least squares error of spectral signature and make the subpixel land cover map spatially smooth, simultaneously. By applying to an Airborne Visible/Infrared Imaging Spectrometer hyperspectral image, the proposed model was evaluated both visually and quantitatively by comparing it with hard classification and the two-step SPM approach. The results showed that the regularization parameter, which balances the influence of spectral and spatial terms, plays an important role on the solution. The L-curve approach was a reasonable method to select the regularization parameter, with which an increased accuracy of the proposed model was obtained.

Integrating Spatial Structure in Super-Resolution Mapping of Hyper-Spectral Image

This paper proposes and validates a super-resolution mapping method based on minimization of energy function defined by spatial dependence and spatial texture structure. The method is a three-step classification routine that involves the integration of 1) Obtaining fractions of land cover classes of mixture pixels accurately; 2) Using energy function derived from the spatial dependence and spatial distribution of class proportions within and between pixels to formulate a super-resolution mapping model; and 3) Solving the nonlinear model with simulated annealing algorithm to obtain a sharpened image. A case study with real Hyperion dataset associated with ALOS image was implemented to validate the method. The experimental result confirms that this model can potentially capture the spatially anisotropic structure of sub-pixel land cover classes by incorporating spatial structure to spatial dependence, and thus of having better performance.

Efficient collection of training data for sub-pixel land cover classification using neural networks

Artificial neural networks (ANNs) are a popular class of techniques for performing soft classifications of satellite images. They have successfully been applied for estimating crop areas through sub-pixel classification of medium to low resolution images. Before a network can be used for classification and estimation, however, it has to be trained. The collection of the reference area fractions needed to train an ANN is often both time-consuming and expensive. This study focuses on strategies for decreasing the efforts needed to collect the necessary reference data, without compromising the accuracy of the resulting area estimates. Two aspects were studied: the spatial sampling scheme (i) and the possibility for reusing trained networks in multiple consecutive seasons (ii). Belgium was chosen as the study area because of the vast amount of reference data available. Time series of monthly NDVI composites for both SPOT-VGT and MODIS were used as the network inputs. The results showed that accurate regional crop area estimation (R2>80%) is possible using only 1% of the entire area for network training, provided that the training samples used are representative for the land use variability present in the study area. Limiting the training samples to a specific subset of the population, either geographically or thematically, significantly decreased the accuracy of the estimates. The results also indicate that the use of ANNs trained with data from one season to estimate area fractions in another season is not to be recommended. The interannual variability observed in the endmembers’ spectral signatures underlines the importance of using up-to-date training samples. It can thus be concluded that the representativeness of the training samples, both regarding the spatial and the temporal aspects, is an important issue in crop area estimation using ANNs that should not easily be ignored.

Development of a circa 2000 land cover map of northern Canada at 30 m resolution from Landsat

Previous land cover maps covering northern Canada have been of insufficient spatial or thematic detail to address emerging northern issues such as wildlife habitat, land use planning, and fine-scale land cover dynamics. Mapping northern land cover requires medium-resolution (~30 m) remote sensing data to effectively characterize cover types that are spatially heterogeneous and cannot be consistently represented at coarser (250 m – 1 km) scales. In this paper, we present a land cover map of northern Canada at 30 m spatial resolution suitable for application in northern land use planning, wildlife habitat assessment, and climate change impact assessment and adaption. Orthorectified circa 2000 Landsat data were acquired from the Centre for Topographic Information, with coverage from the treeline to the northern tip of Ellesmere Island, and combined into 16 radiometrically balanced large-area mosaics. A stratified unsupervised cluster labelling approach was used for map generation. Literature on northern land cover and vegetation mapping and numerous northern vegetation surveys were examined to define a land cover legend containing 15 classes. Field data gathered during several campaigns were used in conjunction with other available medium-resolution land cover maps to develop a dataset for training and validation. Standardized accuracy assessment is limited due to the cost of field data acquisition and the small archive of reference data in northern regions. Comparison with field data used only to aid cluster labelling suggests 81.5% accuracy for 76 plots, and examination of subpixel land cover distribution within each 1 km Circumpolar Arctic Vegetation Map (CAVM) class shows good agreement. The map is publicly available through the Natural Resources Canada Geogratis portal in 1 : 250 000 scale National Topographic System (NTS) map sheets.