Subpixel Proportions Research Articles

Leveraging high resolution classifications and random forests for hindcasting decades of mesic ecosystem dynamics in the Landsat time series

Mesic ecosystems are fundamental to conservation efforts in semi-arid systems, but are threatened by climate change and development. Newer earth observation datasets, including Sentinel-1 and −2, provide opportunities to monitor mesic ecosystems at meaningful spatial scales, but are insufficient for measuring decadal-scale changes. Conversely, the Landsat time series has decades of data, but images are spatially coarse relative to many of the mesic ecosystem areas that sustain dryland systems, resulting in classifications with mixed pixels inadequate for effective monitoring. We developed a workflow that uses 10-m classifications produced from fusion of the Sentinel-1 and −2 time series (2017–2020) to estimate sub-pixel proportions of Landsat time series observations (2004–2020). Using random forest regression models, we quantified water resource proportions (WRP) of surface water, mesic vegetation, and upland land covers within each 30-m Landsat pixel. We incorporated ancillary covariates to account for varying topographic conditions, land cover, and climate. Results indicate that our approach consistently estimates sub-pixel proportions of Landsat pixels more accurately compared to spectral mixture analysis (SMA). The WRP product for surface water had up to 8% less error than SMA as measured by Mean Absolute Error (MAE) and up to 17% less error as measured by Root Mean Squared Error (RMSE). For mesic vegetation, the WRP product outperformed SMA by up to 4% (MAE) and 7% (RMSE). Finally, we demonstrated the ability of our time series to characterize historical water resource availability at a case study site with a well documented restoration history by qualitatively examining the mesic vegetation dynamics time series to identify system responses to restoration efforts. Our approach allows us to hindcast observations of Sentinel products and measure water resource dynamics with greater precision over larger temporal scales. We envision these WRP data to be useful for measuring the impacts of conservation interventions, disturbance recovery, or land use changes that pre-date the Sentinel time series.

Assessing the Accuracy of Landsat Vegetation Fractional Cover for Monitoring Australian Drylands

Satellite-derived vegetation fractional cover (VFC) has shown to be a promising tool for dryland ecosystem monitoring. This model, calibrated through biophysical field measurements, depicts the sub-pixel proportion of photosynthetic vegetation (PV), non-photosynthetic vegetation (NPV) and bare soil (BS). The distinction between NPV and BS makes it particularly important for drylands, as these fractions often dominate. Two Landsat VFC products are available for the Australian continent: the original Joint Remote Sensing Research Program (JRSRP) product, and a newer Digital Earth Australia (DEA) product. Although similar validation statistics have been presented for each, an evaluation of their differences has not been undertaken. Moreover, spatial variability of VFC accuracy within drylands has not been comprehensively assessed. Here, a large field dataset (4207 sites) was employed to compare Landsat VFC accuracy across the Australian continent, with detailed spatial and temporal analysis conducted on four regions of interest. Furthermore, spatiotemporal features of VFC unmixing error (UE) were explored to characterize model uncertainty in large areas yet to be field sampled. Our results showed that the JRSRP and DEA VFC were very similar (RMSE = 4.00–6.59) and can be employed interchangeably. Drylands did not show a substantial difference in accuracy compared to the continental assessment; however contrasting variations were observed in dryland subtypes (e.g., semi-arid and arid zones). Moreover, VFC effectively tracked total ground cover change over time. UE increased with tree cover and height, indicating that model uncertainty was low in typical dryland landscapes. Together, these results provide guiding points to understanding the Australian ecosystems where VFC can be used with confidence.

Multi-decade, multi-sensor time-series modelling—based on geostatistical concepts—to predict broad groups of crops

We have mapped the broad groups of crops grown each summer and winter, from 1987 to 2017, for a 300,000-km2 region of Queensland, Australia. These maps are part of a legislated decision-making process for the protection of prime agricultural land. For summer, the two groups of crops are ‘Coarse-grain & Pulse’ and ‘Cotton’. For winter, the two groups of crops are ‘Cereal’ and ‘Pulse’. Non-crop groups, present in both summer and winter, are ‘Bare soil’ and ‘Other’ (comprising pastures, woody vegetation, and crop residues). The foundation of the maps is time-series modelling—specifically, applying the concepts of geostatistics in the temporal domain—to model the variation in land-surface phenology within a growing season. The time-series model is flexible, robust, parsimonious, parallelisable, and able to deal with irregular observations. We combined satellite imagery from the Landsat sensors, as well as, when available, Sentinel-2A and MODIS (with the last two reprojected to the 30-m grid of Landsat). We applied the time-series model pixel-wise across the study region, to three variables derived from satellite imagery gathered for an individual growing season: enhanced vegetation index, and the sub-pixel proportions of bare-ground and non-photosynthetic vegetation. Weekly-averaged predicted phenological metrics then served as explanatory variables in a tiered, tree-based classification model, for the prediction of the groups. The classification model comprised two expert rules and two random forests. Prior to fitting the classification model, geospatial object-based image analysis was used to change the scale of analysis from individual pixels to (approximately) field-based segments. From the perspective of a map-user, in any given growing season we predicted ‘Coarse-grain & Pulse’ correctly in 79% of cases; the values for ‘Cotton’, ‘Cereal’, and ‘Pulse’ were 90%, 84%, and 73%, respectively; ‘Bare soil’ was 72% in summer, and 88% in winter. ‘Other’ was the most accurately mapped group (98% correct in summer, and 99% correct in winter).

Enhancing spectral unmixing by considering the point spread function effect

The point spread function (PSF) effect exists ubiquitously in real remotely sensed data and such that the observed pixel signal is not only determined by the land cover within its own spatial coverage but also by that within neighboring pixels. The PSF, thus, imposes a fundamental limit on the amount of information captured in remotely sensed images and it introduces great uncertainty in the widely applied, inverse goal of spectral unmixing. Until now, spectral unmixing has erroneously been performed by assuming that the pixel signal is affected only by the land cover within the pixel, that is, ignoring the PSF. In this paper, a new method is proposed to account for the PSF effect within spectral unmixingto produce more accurate predictions of land cover proportions. Based on the mechanism of the PSF effect, the mathematical relation between the coarse proportion and sub-pixel proportions in a local window was deduced. Area-to-point kriging (ATPK) was then proposed to find a solution for the inverse prediction problem of estimating the sub-pixel proportions from the original coarse proportions. The sub-pixel proportions were finally upscaled using an ideal square wave response to produce the enhanced proportions. The effectiveness of the proposed method was demonstrated using two datasets. The proposed method has great potential for wide application since spectral unmixing is an extremely common approach in remote sensing.

A lake detection algorithm (LDA) using Landsat 8 data: A comparative approach in glacial environment

Glacial lakes show a wide range of turbidity. Owing to this, the normalized difference water indices (NDWIs) as proposed by many researchers, do not give appropriate results in case of glacial lakes. In addition, the sub-pixel proportion of water and use of different optical band combinations are also reported to produce varying results. In the wake of the changing climate and increasing GLOFs (glacial lake outburst floods), there is a need to utilize wide optical and thermal capabilities of Landsat 8 data for the automated detection of glacial lakes. In the present study, the optical and thermal bandwidths of Landsat 8 data were explored along with the terrain slope parameter derived from Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version2 (ASTER GDEM V2), for detecting and mapping glacial lakes. The validation of the algorithm was performed using manually digitized and subsequently field corrected lake boundaries. The pre-existing NDWIs were also evaluated to determine the supremacy and the stability of the proposed algorithm for glacial lake detection. Two new parameters, LDI (lake detection index) and LF (lake fraction) were proposed to comment on the performances of the indices. The lake detection algorithm (LDA) performed best in case of both, mixed lake pixels and pure lake pixels with no false detections (LDI=0.98) and very less areal underestimation (LF=0.73). The coefficient of determination (R2) between areal extents of lake pixels, extracted using the LDA and the actual lake area, was very high (0.99). With understanding of the terrain conditions and slight threshold adjustments, this work can be replicated for any mountainous region of the world.

Sensitivity analysis for predicting continuous fields of tree-cover and fractional land-cover distributions in cloud-prone areas

The use of multi-temporal datasets, such as vegetation index time series or phenological metrics, for improved classification and regression performance is well established in the remote-sensing science community. However, the usefulness of such information is less apparent for areas with distinct wet season periods and heavily concentrated cloud cover. In view of this, this study examines the potential of multi-temporal datasets for the estimation of sub-pixel land-cover fractions and percentage tree cover in an area having distinct wet and dry seasons. Prediction is based on a regression tree algorithm in combination with linear least-squares regression planes, which relate multi-spectral and multi-temporal satellite data from the Moderate-Resolution Imaging Spectroradiometer (MODIS) sensor to sub-pixel land-cover proportions and percentage tree cover, derived from high-resolution land-cover maps. Furthermore, several versions of the latter were produced using different classification approaches to evaluate the sensitivity of the response variable on overall prediction accuracy. The results were evaluated according to absolute accuracy levels and according to their long-term inter-annual robustness by applying the regression models to MODIS data over a period of 11 years. The best regression model based on dry season information only estimated continuous fields of percentage tree cover with a prediction error of less than 7% and an inter-annual variability of less than 4% over a time period of 11 years. The inclusion of intra-annual information did not contribute to any improvements in model accuracy compared to information from the dry season alone, and furthermore, deteriorated inter-annual robustness of model predictions. In addition, it has been shown that the quality of the response variable in the training data had significant effects on overall accuracy.

Validation of operational AVHRR subpixel snow retrievals over the European Alps based on ASTER data

Snow is of great economic and social importance for the European Alps. Accurate monitoring of the alpine snow cover is a key component in studying regional climate change as well as in daily weather forecasting and snowmelt run‐off modelling. These applications require snow cover information on a high temporal resolution in near‐real time. For the European Alps, operational snow cover fraction maps are generated on a daily basis using data from the Advanced Very High Resolution Radiometer (AVHRR) on board the National Oceanic and Atmospheric Administration (NOAA) platforms. Snow cover distribution is inherently discontinuous and heterogeneous in this mountainous region. We have therefore implemented a straightforward multiple endmember unmixing approach to estimate fractional snow cover. Subpixel proportions are difficult to validate because similar products are not available and appropriate ground‐based observations do not exist. In this study, we validate AVHRR subpixel snow retrievals using binary classified data sets from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) to establish absolute errors of our operational approach at three test sites. Our analysis indicates that the AVHRR subpixel maps compare well with the aggregated ASTER data, showing an overall correlation of 0.78 and providing subpixel estimates with a mean absolute error of 10.4% fractional snow cover. Discrepancies between AVHRR and ASTER snow fraction maps can be attributed to varying snow conditions, terrain effects and density in forest cover.

Nonlinear estimation of subpixel proportion via kernel least square regression

Spectral mixture analysis is an efficient approach to spectral decomposition of hyperspectral remotely sensed imagery, using land cover proportions which can be estimated from pixel values through model inversion. In this paper, a kernel least square regression algorithm has been developed for nonlinear approximation of subpixel proportions. This procedure includes two steps. The first step is to select the feature vectors by defining a global criterion to characterize the image data structure in the feature space and the second step is the projection of pixels onto the feature vectors and the application of classical linear regressive algorithm. Experiments using simulated data, synthetic data and Enhanced Thematic Mapper (ETM)+ data have been carried out, and the results demonstrate that the proposed method can improve proportion estimation. By using the simulated and synthetic data, over 85% of the total pixels in the image are found to lie between the 10% difference lines, and the root mean square error (RMSE) is less than 0.09. Using the real data, the proposed method can also perform satisfactorily with an average RMSE of about 0.12. This algorithm was also compared with other widely used kernel based algorithms, i.e. support vector regression and radial basis function neutral network and the results show that the proposed algorithm outperforms other algorithms about 5% in subpixel proportion estimation.

Resolution dependent errors in remote sensing of cultivated areas

Remote sensing has become a common and effective method for estimating the areal coverage of land cover classes. One class of particular interest is agriculture as area estimates of cultivated lands are important for purposes such as estimating yields or irrigation needs. The synoptic coverage of satellite imagery and the relative ease of automated analysis have led to widespread mapping of agriculture using remote sensing. The accuracy of area estimates derived from these maps is known to be related to the accuracy of the maps. However, even in the situation where the map is very accurate, errors in area estimates may occur. These errors result from the behavior of the distribution of subpixel proportions of cultivated areas, and how that behavior changes as a result of sensor spatial resolution and class definitions. The sensitivity of estimates of cultivated areas to sensor spatial resolution and to the choice of threshold used to define cultivated land is explored in six agriculturally distinct locations around the world. Using a beta model for the distribution of subpixel proportions that is parameterized using variograms, it is possible to model the distribution of subpixel proportions for any spatial resolution. When the spatial resolution is small with respect to the spatial structure of the landscape (as measured by the variogram range) use of any class definition threshold produces an estimate very close to the true area coverage. On the other hand, as the resolution becomes coarse in relation to the variogram range, the subpixel proportions are no longer concentrated at the extremes of the distribution and the difference between the estimated and the true area has greater sensitivity to the selected threshold used to define classes. Thus, for the cases examined here, both the resolution and the class definition threshold have a strong influence on area estimates. The spatial resolutions where errors can be large depend on landscape spatial structure, which can be quantified using variograms. The net effect is that for the same spatial resolution, some places will exhibit much larger errors in area estimates than others. For the site in the Anhui province of China, where agricultural fields are very small (0.07 ha on the average), area estimates are highly sensitive to class definition thresholds even at the relatively fine resolution of 45 m. Conversely, in California (USA) spatial resolutions as coarse as 500 m can be used to reliably estimate cultivated areas. Results also suggest that the proportion of the total area that is cultivated significantly influences the accuracy of area estimates. When the area proportion is low, the class definition threshold must also be low to achieve an accurate area estimate. Conversely, in areas dominated by agriculture, a very stringent class definition of cultivated lands is required for accurate area estimates. While explored in the context of estimation of cultivated areas, the findings presented here are generic to the problem of area estimation using remote sensing.

Cross-Sensor Calibration Between Ikonos and Landsat ETM+ for Spectral Mixture Analysis

Spectral mixture analysis is an algorithm that is developed to overcome the weakness in traditional land-use/land-cover (LULC) classification where each picture element (pixel) from remote sensing is assigned to one and only one LULC type. In reality, a remotely sensed signal from a pixel is often a spectral mixture from several LULC types. Spectral mixture analysis can derive subpixel proportions for the endmembers from remotely sensed data. However, one frequently faces the problem in determining the spectral signatures for the endmembers. This study provides a cross-sensor calibration algorithm that enables us to obtain the endmember signatures from an Ikonos multispectral image for spectral mixture analysis using Landsat ETM+ images. The calibration algorithm first converts the raw digital numbers from both sensors into at-satellite reflectance. Then, the Ikonos at-satellite reflectance image is degraded to match the spatial resolution of the Landsat ETM+ image. The histograms at the same spatial resolution from the two images are matched, and the signatures from the pure pixels in the Ikonos image are used as the endmember signatures. Validation of the spectral mixture analysis indicates that the simple algorithm works effectively. The algorithm is not limited to Ikonos and Landsat sensors. It is, in general, applicable to spectral mixture analysis where a high spatial resolution sensor and a low spatial resolution sensor with similar spectral resolutions are available as long as images collected by the two sensors are close in time over the same place.

A SPLIT Model for Extraction of Subpixel Impervious Surface Information

This paper introduces a Subpixel Proportional Land cover Information Transformation (SPLIT) model to extract proportions of impervious surfaces in urban and suburban areas. High spatial resolution airborne Digital Multispectral Videography (DMSV) data provided subpixel information for Landsat TM data. The SPLIT model employed a Modularized Artificial Neural Network (MANN) to integrate multi-sensor remote sensing data and to extract proportions of impervious surfaces and other types of land cover within TM pixels. Through a control unit, the MANN was able to decompose a complex task into multiple subtasks by using a group of sub-networks. The SPLIT model identified spectral relations between TM pixel values and the corresponding DMSV subpixel patterns. The established relationship allows extrapolation of the SPLIT model to the areas beyond DMSV data coverage. We applied five intervals, i.e., 81 percent, to map the subpixel proportions of land cover types. We extrapolated the SPLIT model from training sites that have both TM and DMSV coverage into the entire DuPage County with TM data as the input. The extrapolation received 82.9 percent overall accuracy for the extracted proportions of urban impervious surface.

Approaches to fractional land cover and continuous field mapping: A comparative assessment over the BOREAS study region

Subpixel land cover mapping involves the estimation of surface properties using sensors whose spatial sampling is coarse enough to produce mixtures of the properties within each pixel. This study evaluates five algorithms for mapping subpixel land cover fractions and continuous fields of vegetation properties within the BOREAS study area. The algorithms include a conventional “hard”, per-pixel classifier, a neural network, a clustering/look-up-table approach, multivariate regression, and linear least squares inversion. A land cover map prepared using a Landsat TM mosaic was adopted as the source of fine scale calibration and validation data. Coarse scale mixtures of five basic land cover classes and continuous vegetation fields, both corresponding to the field of view of SPOT-VEGETATION imagery (1.15-km pixel size), were synthesised from the TM mosaic using a modelled point spread function. Two measures of land cover distribution were used, fractions of fine scale land cover categories and continuous fields of vegetation structural characteristics. The subpixel algorithms were applied using both proximate (<100 km) and distant (>400 km) separation between training and validation regions. “Hard” classification performed poorly in estimating proportions or continuous fields. The neural network, look-up-table and multivariate regression algorithms produced good matches of spatial patterns and regional land cover composition for the proximate treatment. However, all three methods exhibited substantial biases with the distant treatment due to the characteristics of the training data. Linear least squares inversion offers a relatively unbiased but less precise alternative for subpixel proportion and fraction mapping as it avoids calibration to the a priori distribution of land cover in the training data. In general, a combination of multivariate regression for proximate training data and linear least squares inversion for distant training data resulted in woody fraction estimates within 20% of the Landsat TM classification-based estimates.

On the use of prior and posterior information in the subpixel proportion problem

Although the problems of classification and subpixel proportion estimation in remote sensing land cover characterization generally are held to be distinct (though related), often elements of the former are adopted in addressing the latter that blur this distinction - particularly, regarding the use of prior and posterior information. The author examines this issue in more detail, using simple, canonical versions of the two problems and, in the course of this examination, provides analytical expressions upon which to build discussion of improvements to subpixel proportion estimation from a statistical viewpoint.

Spectral unmixing ofcompact airborne spectrographic imager (CASI) data for quantifying sub-pixel proportions of biophysical parameters in a coastal dune system

In environmentally sensitive and large coastal dune systems, identification and mapping of favourable (sandy), and unfavourable (scrub rich) habitats form the key to coastal conservation and management. In highly mixed floral environments, however, such an identification is difficult with low resolution multispectral imagery. In such cases, spectral unmixing is useful to resolve the “mixed pixel” effects. The Kenfig NNR (National Nature Reserve), South Wales. UK, bestowed with high biodiversity, is facing loss of successionally young slack habitats due to dune stabilisation and vegetation succession. To map such habitats, linear spectral unmixing of airborne MSS (CASI-Compact Airborne Spectrographic Imager) data was performed using the Constrained Least Square (CLS) method, and the sub-pixel proportions of the spectral end members viz. sand, vegetation and shade/moisture were defined. Comparison of the estimated fractions with the growth forms of the dominant vegetation species Salix repens (Creeping willow) with classified digital aerial photographs shows a positive correlation, thus giving us confidence in the mixture modelling technique. Apart from aiding in conservation management, such a fuzzy classification of multi-date imagery helps to delineate sandy and vegetated areas for change detection and landscape/ habitat succession studies.

Subpixel Land Use Classification and Retrieval of Forest Stem Volume in the Boreal Forest Zone by Employing SSM/I Data

T he use of microwave radiometry for the retrieval of land use distribution (subpixel proportions) and forest stem volume for 25 km and 50 km pixels is investigated, with new algorithms introduced for the spaceborne multichannel SSM/I instrument. The test sites are located in the boreal forest zone in Finland. SSM/I data from snow-free periods (from May to September) are used in the study, together with a digital high-resolution land use map as reference data. A stepwise algorithm is developed, in which the surface emissivity is estimated for each of the pixels in the first stage of the algorithm. By using a training data set (50% of all data), empirical coefficients for the subpixel land use and forest stem volume are determined. The results for subpixel land use classification (retrieval of fractions of individual land use categories contained within study pixels) are promising. When there are two or three categories to retrieve, the rms errors range from 3% to 9%, and the respective relative errors from 3% to 32%. When there are six categories to retrieve, the performance is less satisfactory, with relative errors exceeding 100% for some categories. For the retrieval of forest stem volume, a similar approach is used. The validation data set shows rms errors (standard errors) of approximately 20 m 3/ha when the mean estimated forest stem volume is around 70 m 3/ha. When compared to reported Landsat TM-based biomass estimates for a study area in Finland, the r 2 (coefficient of determination) values are comparable. The methods are also validated for a wider use by limiting the training data set to either northern or southern Finland, since the surface and forest types in these areas are systematically different (subgroups of boreal forests). The results show that the methods are applicable to remote areas with less training data. This is especially true for the stem volume estimation and limited (two or three classes) subpixel land use classification. These can be considered as promising results for a coarse resolution, high-frequency microwave instrument such as the SSM/I.

Mapping sub-pixel proportional land cover with AVHRR imagery

A problem with NOAA AVHRR imagery is that the intrinsic scale of spatial variation in land cover in the U.K. is usually finer than the scale of sampling imposed by the image pixels. The result is that most NOAA AVHRR pixels contain a mixture of land cover types (sub-pixel mixing). Three techniques for mapping the sub-pixel proportions of land cover classes in the New Forest, U.K. were compared: (i) artificial neural networks (ANN); (ii) mixture modelling; and (iii) fuzzy c -means classification. NOAA AVHRR imagery and SPOT HRV imagery, both for 28 June 1994, were obtained. The SPOT HRV images were classified using the maximum likelihood method, and used to derive the 'known' sub-pixel proportions of each land cover class for each NOAA AVHRR pixel. These data were then used to evaluate the predictions made (using the three techniques and the NOAA AVHRR imagery) in terms of the amount of information provided, the accuracy with which that information is provided, and the ease of implementation. The ANN was the most accurate technique, but its successful implementation depended on accurate co-registration and the availability of a training data set. Supervised fuzzy c -means classification was slightly more accurate than mixture modelling.

Subpixel measurement of tropical forest cover using AVHRR data

Abstract Tropical forest assessment using data from the Advanced Very High Resolution Radiometer (AVHRR) may lead to inaccurate estimates of forest cover in regions of small subpixel forest or non-forest patches and in regions where the pattern of clearance is particularly convoluted. Test sites typifying these two patterns were chosen in Ghana and Rondonia, respectively. To capture the subpixel proportions of forest cover, a linear mixture model was applied to two AVHRR test images over the test sites. The model produced image outputs in which pixel intensities indicated the proporton of forest cover per km2. For comparison, supervised maximum likelihood classifications were also performed. The outputs were assessed against classified Landsat TM scenes, converted to proportions maps and coregistered to the AVHRR images. An empirical method was applied for determining the critical forest cover per km2 needed for an AVHRR pixel to be classified as forest. The critical values exceeded 50 per cent, indicating a tendency for AVHRR classification to underestimate forest cover. This was confirmed by comparing estimates of total forest cover obtained from the AVHRR and TM classifications. In the case of Ghana, a more accurate estimate of forest cover was obtained from the AVHRR mixture model than from the classification. Both mixture model outputs were found to be well correlated with those from Landsat TM. Further work should test the robustness of the approach adopted here when applied to much larger areas.