V.A. Tolpekin

Quantification of the Effects of Land-Cover-Class Spectral Separability on the Accuracy of Markov-Random-Field-Based Superresolution Mapping

This paper explores the effects of class separability in Markov-random-field-based superresolution mapping (SRM). We propose to account for class separability by means of controlling the balance tuned by a smoothness parameter between the prior and the likelihood terms in the posterior energy function. A generally applicable procedure estimates the optimal smoothness parameter, based on local energy balance analysis. The study shows how the optimal value of the smoothness parameter depends quantitatively and monotonically upon the class separability and the scale factor. Effects are studied on an image synthesized from an agricultural scene with field boundary subpixels. We varied systematically the class separability, the scale factor, and the smoothness parameter values. The accuracy of the resulting land-cover-map image is assessed by means of the kappa statistic at the fine-resolution scale and the class area proportion at the coarse-resolution scale. Performance is compared with a hard and a soft classification of the coarse-resolution image. We demonstrate that an optimal value of the smoothness parameter exists for each combination of scale factor and class separability. This allows us to reach a high classification accuracy (kappa = 0.85) even for poorly separable classes, i.e., with a transformed divergence equal to 0.5 and a scale factor equal to 10. The developed procedure agrees with the empirical data for the optimal smoothness parameter. The study shows that SRM is now applicable to a larger set of images with class separability ranging from poor to excellent.

Context-sensitive extraction of tree crown objects in urban areas using VHR satellite images

Municipalities need accurate and updated inventories of urban vegetation in order to manage green resources and estimate their return on investment in urban forestry activities. Earlier studies have shown that semi-automatic tree detection using remote sensing is a challenging task. This study aims to develop a reproducible geographic object-based image analysis (GEOBIA) methodology to locate and delineate tree crowns in urban areas using high resolution imagery. We propose a GEOBIA approach that considers the spectral, spatial and contextual characteristics of tree objects in the urban space. The study presents classification rules that exploit object features at multiple segmentation scales modifying the labeling and shape of image-objects. The GEOBIA methodology was implemented on QuickBird images acquired over the cities of Enschede and Delft (The Netherlands), resulting in an identification rate of 70% and 82% respectively. False negative errors concentrated on small trees and false positive errors in private gardens. The quality of crown boundaries was acceptable, with an overall delineation error <0.24 outside of gardens and backyards.

Application of the Expectation Maximization Algorithm to Estimate Missing Values in Gaussian Bayesian Network Modeling for Forest Growth

The leaf area index (LAI) is a biophysical variable related to atmosphere-biosphere exchange of CO2. One way to obtain LAI value is by the Moderate Resolution Imaging Spectroradiometer (MODIS) biophysical products. In this paper, we use this product to improve the physiological principles predicting growth model within a Gaussian Bayesian network (GBN) setup. The MODIS time series, however, contains gaps caused by persistent clouds, cloud contamination, and other technique problems. We used the Expectation Maximization (EM) algorithm to estimate these missing values. During a period of 26 successive months, the EM algorithm is applied to four different cases: successively and not successively missing values during two different winter seasons, successively and not successively missing values during one spring season, and not successively missing values during the full study. Results show that the maximum value of the averaged absolute error between the original values and those estimated equals 0.16. This low value indicates that the estimated values well represent the original values. Moreover, the root mean square error of the GBN output reduces from 1.57 to 1.49 when performing the EM algorithm to estimate the not successively missing values. We conclude that the EM algorithm within a GBN can adequately handle missing MODIS LAI values and improves the estimation of the LAI.

Angular Backscatter Variation in L-Band ALOS ScanSAR Images of Tropical Forest Areas

Scanning synthetic aperture radar (ScanSAR) systems provide continuous information over large areas, but for effective use of such products in tropical forest, the decrease of radar backscatter with large variation of incidence angles requires attention. This letter analyzes the dependence of radar backscatter on incidence angle for L-band ScanSAR images of tropical forest. We investigated and modeled the angular backscatter effect per land-cover class in three ScanSAR images of the Colombian Orinoco. We found that there is an evident effect of incidence angle on radar backscatter, depending on land-cover class, moisture content, and physical structure of the reflecting targets. To normalize the angular backscatter variation, we proposed two methods. The first one applies a cosine correction estimated through linear regression. The second one models the radar backscatter of flooded forest considering second-order signal interactions. The model explains the observed backscatter of flooded forest areas in the rainy season (R2 that is larger than 0.77).

Fuzzy Super Resolution Mapping Based on Markov Random Fields

Recent research has used Markov Random Fields (MRF) as a method for super-resolution mapping (SRM). This paper investigated the per-pixel uncertainty associated with MRF based SRM. This provided insight into the spatial distribution of uncertainty associated with SRM. Furthermore, the map of per-pixel uncertainty clearly shows the boundary between land-cover classes and this may provide an input for image segmentation. The insight provided by the per-pixel uncertainty together with the class boundaries will be valuable for development of the MRF approach to super-resolution mapping.

Detection of built-up area in optical and synthetic aperture radar images using conditional random fields

Abstract Classifying built-up areas from satellite images is a challenging task due to spatial and spectral heterogeneity of the classes. In this study, a contextual classification method based on conditional random fields (CRFs) has been used. Spatial and spectral information from blocks of pixels were employed to identify built-up areas. The CRF association potential was based on support vector machines (SVMs), whereas the CRF interaction potential included a data-dependent term using the inverse of the transformed Euclidean distance. In this way, accuracy was stable for a varying smoothness parameter, while preserving class boundaries and aggregating similar labels, and a discontinuity adaptive model was obtained and conditioned on data evidence. The classification was applied on satellite towns around the city of Nairobi, Kenya. The accuracy exceeded that of Markov random fields, SVM, and maximum likelihood classification by 1.13%, 2.22%, and 8.23%, respectively. The CRF method had the lowest fraction of false positives. The study concluded that CRFs can be used to better detect built-up areas. In this way, it provides accurate timely spatial information to urban planners and other professionals.

Local interpolation of coseismic displacements measured by InSAR

Abstract Coseismic displacements play a significant role in characterizing earthquake causative faults and understanding earthquake dynamics. They are typically measured from InSAR using pre- and post-earthquake images. The displacement map produced by InSAR may contain missing coseismic values due to the decorrelation of ASAR images. This study focused on interpolating missing values in the coseismic displacement map of the 2003 Bam earthquake using geostatistics with the aim of running a slip distribution model. The gaps were grouped into 23 patches. Variograms of the patches showed that the displacement data were spatially correlated. The variogram prepared for ordinary kriging (OK) indicated the presence of a trend and thus justified the use of universal kriging (UK). Accuracy assessment was performed in 3 ways. First, 11 patches of equal size and with an equal number of missing values generated artificially, were kriged and validated. Second, the four selected patches results were validated after shifting them to new locations without missing values and comparing them with the observed values. Finally, cross validation was performed for both types of patch at the original and shifted locations. UK results were better than OK in terms of kriging variance, mean error (ME) and root mean square error (RMSE). For both OK and UK, only 4 out of 23 patches (1, 5, 11 and 21) showed ME and RMSE values that were substantially larger than for the other patches. The accuracy assessment results were found to be satisfactory with ME and RMSE values close to zero. InSAR data inversion demonstrated the usefulness of interpolation of the missing coseismic values by improving a slip distribution model. It is therefore concluded that kriging serves as an effective tool for interpolating the missing values on a coseismic displacement map.

Improving Forest Growth Estimates Using a Bayesian Network Approach

Estimating the contribution of forests to carbon sequestration is commonly done by applying forest growth models. Such models inherently use field observations, such as leaf area index (LAI), whereas relevant information is also available from remotely sensed images. The purpose of this study is to improve the LAI estimated from the physiological principles predicting growth (3-PG) model by combining its output with LAI derived from Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) satellite imagery. A Bayesian network (BN) approach is proposed to take care of the different structure of the inaccuracies in the two data sources. It addresses the bias in the 3-PG model and the noise of the ASTER images. Moreover, the EM algorithm is introduced into BN to estimate missing the LAI ASTER data, since they are not available for long time series due to the atmospheric conditions. This paper shows that the outputs obtained with the BN were more accurate than the 3-PG estimate, as the root mean square error reduces to 0.46, and the relative error to 5.86%. We conclude that the EM-algorithm within a BN can adequately handle missing LAI ASTER values, and BNs can improve the estimation of LAI values. Ultimately, this method may be used as a predicting model of LAI values, and handling the missing data of ASTER images time series.

Sensitivity of Reflectance to Water Vapor and Aerosol Optical Thickness

The atmospheric condition parameters used in the radiative transfer-based atmospheric correction (AC) are often uncertain. This uncertainty propagates to the estimated reflectance. The reflectance, is, however, not equally sensitive to all the parameters. A sensitivity analysis (SA) helps in prioritizing the parameters. The objective of this study was to perform an SA of reflectance to water vapor concentration (wv) and aerosol optical thickness (AOT). SA was performed using the Fourier amplitude sensitivity test (FAST) method, which computes sensitivity indices (SI) of these parameters. Besides variation in the two parameters, we also studied the effect of surface albedo on the SI by quantifying SI for three target surfaces (in the spectral range 0.44-0.96 μm): 1) a dark target (water); 2) a bright target (bare soil); and 3) a target having low albedo in the visible and high albedo in near-infrared range (forest). For AOT, high (≈0.9) SI values were observed at the nonwater absorption wavelengths. For wv, high SI values were observed at wavelengths, where strong absorption features are located and when the surface albedo was high. For the dark target, the effect of AOT was prominent throughout the spectral range. We found that the sensitivity of reflectance to wv and AOT is a function of wavelength, strength of the absorption features, and surface albedo. We conclude that AOT is a more important parameter for dark targets than wv even at the principal absorption feature. For bright targets, the importance of wv and AOT depends on the strength of the absorption feature.

Markov random field-based method for super-resolution mapping of forest encroachment from remotely sensed ASTER image

Forest encroachment (FE) is a problem in Andaman and Nicobar Islands (ANI) in India for environment and planning. Small gaps created in the forest slowly expand its periphery disturbing the biodiversity. Therefore, intrusion of poachers, slash and burn and other factors causing FE must be carefully detected and monitored. Remote sensing offers a great opportunity to accomplish this task because of its synoptic view. Conventional classification methods with remotely sensed images are problematic because of small size of FE and mixed landcover composition. This study presents an application of super-resolution mapping (SRM) based on Markov random field for detection of FE using ASTER (15 m) images. The SRM results were validated using multispectral IRS LISS-IV (5.8 m) image. Non-contiguous FE patches of various sizes and shapes are characterized using the spatial contextual information. The novelty of this approach lies in the identification and separability of small FE pockets which could not be achieved with pixel-based maximum likelihood classifier (MLC). The SRM parameters were optimized and found comparable to previous studies. Classification accuracy obtained with SRM at scale factor 3 is κ = 0.62 that is superior to accuracy of MLC (κ = 0.51). SRM is a promising tool for detection and monitoring of FE at Rutland Island in ANI, India.