Tuomas Häme

Multiscale analysis and validation of the MODIS LAI product. I. Uncertainty assessment

The development of appropriate ground-based validation techniques is critical to assessing uncertainties associated with satellite data-based products. Here we present a method for validation of the Moderate Resolution Imaging Spectroradiometer (MODIS) Leaf Area Index (LAI) product with emphasis on the sampling strategy for field data collection. This paper, the first of two-part series, details the procedures used to assess uncertainty of the MODIS LAI product. LAI retrievals from 30 m ETM+ data were first compared to field measurements from the SAFARI 2000 wet season campaign. The ETM+ based LAI map was thus as a reference to specify uncertainties in the LAI fields produced from MODIS data (250-, 500-, and 1000-m resolutions) simulated from ETM+. Because of high variance of LAI measurements over short distances and difficulties of matching measurements and image data, a patch-by-patch comparison method, which is more realistically implemented on a routine basis for validation, is proposed. Consistency between LAI retrievals from 30 m ETM+ data and field measurements indicates satisfactory performance of the algorithm. Values of LAI estimated from a spatially heterogeneous scene depend strongly on the spatial resolution of the image scene. The results indicate that the MODIS algorithm will underestimate LAI values by about 5% over the Maun site if the scale of the algorithm is not matched to the resolution of the data.

Multiscale analysis and validation of the MODIS LAI product II. Sampling strategy

The development of appropriate ground-based validation techniques is critical to assessing uncertainties associated with satellite data-based products. In this paper, the second of a two-part series, we present a method for validation of the Moderate Resolution Imaging Spectroradiometer Leaf Area Index (MODIS LAI) product with emphasis on the sampling strategy for field data collection. Using a hierarchical scene model, we divided 30-m resolution LAI and NDVI images from Maun (Botswana), Harvard Forest (USA) and Ruokulahti Forest (Finland) into individual scale images of classes, region and pixel. Isolating the effects associated with different landscape scales through decomposition of semivariograms not only shows the relative contribution of different characteristic scales to the overall variation, but also displays the spatial structure of the different scales within a scene. We find that (1) patterns of variance at the class, region and pixel scale at these sites are different with respect to the dominance in order of the three levels of landscape organization within a scene; (2) the spatial structure of LAI shows similarity across the three sites, that is, ranges of semivariograms from scale of pixel, region and class are less than 1000 m. Knowledge gained from these analyses aids in formulation of sampling strategies for validation of biophysical products derived from moderate resolution sensors such as MODIS. For a homogeneous (within class) site, where the scales of class and region account for most of the spatial variation, a sampling strategy should focus more on using accurate land cover maps and selection of regions. However, for a heterogeneous (within class) site, accurate point measurements and GPS readings are needed.

AVHRR-Based Forest Proportion Map of the Pan-European Area.

Abstract A methodology was developed and applied to estimating forest area and producing forest maps. The method utilizes satellite data and ground reference data. It takes into consideration the fact that a pixel rarely represents any single ground cover class. This is particularly true for low-spatial-resolution data. It also takes into consideration that the spectral classes overlap. The image was first classified using an unsupervised clustering method. A (multinormal) spectral density function was estimated for each class based on the spectral vectors (reflectance values) of the cluster members. Values of the target variable — the proportion of forested area — were determined for the spectral classes using sampling from CORINE (Coordination of Information on the Environment) Land Cover database. Each pixel was assigned class membership probabilities, which were proportional to the value of the density function of the respective class evaluated at the spectral value of the pixel. The estimate of forest area for the pixel was finally computed by multiplying the class membership probabilities by the class forest area and summing over all the classes. The method was applied over a mosaic of 49 Advanced Very High Resolution Radiometer (AVHRR) images acquired from the National Oceanic and Atmospheric Administration (NOAA)-14 satellite. The estimated forest areas were compared with those extracted from the full-coverage CORINE data and with official forest statistics reported to the European Commissions Statistical Office (EUROSTAT). The forest percentage (proportion of forest area of the total land area) of 12 countries of the European Union was underestimated by 1.8% compared to the CORINE data. It was underestimated by 4.2% when compared with EUROSTATs statistics and 6.0% when compared to United Nations Economic Commission for Europe/Food and Agricultural Organization (UN-ECE/FAO) statistics. The largest underestimation of forest percentage within a country (compared to CORINE) was in France (5.9%). The largest overestimation was found in Ireland, 15.6%.

Volume Scattering Modeling in PolSAR Decompositions: Study of ALOS PALSAR Data Over Boreal Forest

Model-based approaches for decomposing polarimetric backscatter data from boreal forest are discussed in this paper. Several model-based decompositions are analyzed with respect for the most accurate estimation of the volume scattering component. A novel generalized model for description of the volume contribution is proposed when observed backscatter from forest indicates that media does not follow azimuthal symmetry case. The model can be adjusted to the polarimetric synthetic aperture radar (PolSAR) data itself, taking into consideration higher sensitivity of HH against VV backscattering term to the presence of canopy at L-band. The model is general enough to allow a broad range of canopies to be modeled and is shown to comply with several earlier proposed volume scattering mechanism models. It is afterward incorporated in the Freeman-Durden three-component decomposition, yielding an improved modification. The performance of the proposed modification is evaluated using multitemporal ALOS PALSAR data acquired over Kuortane area in central Finland, representing typical mixed boreal forestland. Several decompositions are also benchmarked in order to see how they satisfy physical requirements when decomposing covariance matrix into a weighted sum of individual scattering mechanism contributions. When using experimental data, the proposed decomposition is shown to better satisfy non-negativity constraints for the covariance matrix eigenvalues at each decomposition step with less additional PolSAR data averaging needed. Discussed decompositions are also evaluated for the accuracy of initial stratification based on dominating scattering mechanism using ground reference data.

A new parameterization of canopy spectral response to incident solar radiation: case study with hyperspectral data from pine dominant forest

A small set of independent variables generally seems to suffice when attempting to describe the spectral response of a vegetation canopy to incident solar radiation. This set includes the soil reflectance, the single-scattering albedo, canopy transmittance, reflectance and interception, the portion of uncollided radiation in the total incident radiation, and portions of collided canopy transmittance and interception. All of these are measurable; they satisfy a simple system of equations and constitute a set that fully describes the law of energy conservation in vegetation canopies at any wavelength in the visible and near-infrared part of the solar spectrum. Further, the system of equations specifies the relationship between the optical properties at the leaf and the canopy scales. Thus, the information content of hyperspectral data can be fully exploited if these variables can be retrieved, for they can be more directly related to some of the physical properties of the canopy (e.g. leaf area index). This paper demonstrates this concept through retrievals of single-scattering albedo, canopy absorptance, portions of uncollided and collided canopy transmittance, and interception from hyperspectral data collected during a field campaign in Ruokolahti, Finland, June 14–21, 2000. The retrieved variables are then used to estimate canopy leaf area index, vegetation ground cover, and also the ratio of direct to total incident solar radiation at blue, green, red, and near-infrared spectral intervals. D 2003 Elsevier Science Inc. All rights reserved.

Detecting Man-Made Structures and Changes in Satellite Imagery With a Content-Based Information Retrieval System Built on Self-Organizing Maps

The increasing amount and resolution of satellite sensors demand new techniques for browsing remote sensing image archives. Content-based querying allows an efficient retrieval of images based on the information they contain, rather than their acquisition date or geographical extent. Self-organizing maps (SOMs) have been successfully applied in the PicSOM system to content-based image retrieval in databases of conventional images. In this paper, we investigate and extend the potential of PicSOM for the analysis of remote sensing data. We propose methods for detecting man-made structures, as well as supervised and unsupervised change detection, based on the same framework. In this paper, a database was artificially created by splitting each satellite image to be analyzed into small images. After training the PicSOM on this imagelet database, both interactive and off-line queries were made to detect man-made structures, as well as changes between two very high resolution images from different years. Experimental results were both evaluated quantitatively and discussed qualitatively, and suggest that this new approach is suitable for analyzing very high resolution optical satellite imagery. Possible applications of this work include interactive detection of man-made structures or supervised monitoring of sensitive sites

Improved Mapping of Tropical Forests With Optical and SAR Imagery, Part II: Above Ground Biomass Estimation

Performance of the above ground (dry) biomass estimation with the medium resolution optical (ALOS AVNIR) data and radar (ALOS PALSAR) data was evaluated on a tropical forest site in Lao PDR (Laos). The average biomass of ground reference plots was relatively low, 78 t/ha, due to strong anthropogenic influence in most of the study area. The biomass estimates were computed using linear regression analysis and the Probability method that combines unsupervised clustering and fuzzy estimation. The predictions were validated with independent field plot data. With all the methods and data types, the root mean square error (RMSE) ranged from 33.6 t/ha to 40.1 t/ha (44.2% and 52.8% of mean biomass, respectively). The Probability method produced a larger dynamic range to the predictions than the regression models, which saturated at approximately 100 t/ha. Large errors for higher biomass plots increased the RMSE of Probability over the RMSE of the regression models. The bias ranged from -0.8 to 3.9% except with the Probability model for PALSAR data where the bias was 12.5%. Our study showed that PALSAR data were nearly as good for the biomass estimation as the AVNIR data. A combination of mono-temporal ALOS PALSAR and ALOS AVNIR data did not improve biomass estimation over the performance obtained with AVNIR data alone. For the Probability method, ground reference data should be more representative than that available in this study.

Estimating net primary production of boreal forests in Finland and Sweden from field data and remote sensing

Abstract We calculated annual mean stem volume increment (AMSVI) and total litter fall to produce forest net primary production (NPP) maps at 1-km2 and half-degree resolutions in Finland and Sweden. We used a multi-scale methodology to link field inventory data reported at plot and forestry district levels through a remotely sensed total plant biomass map derived from 1-km2 AVHRR image. Total litter fall was estimated as function of elevation and latitude. Leaf litter fall, a surrogate for fine root production, was estimated from total litter fall by forest type. The gridded NPP estimates agreed well with previously reported NPP values, based on point measurements. Regional NPP increases from northeast to southwest. It is positively related to annual mean temperature and annual mean total precipitation (strongly correlated with temperature) and is negatively related to elevation at broad scale. Total NPP (TNPP) values for representative cells selected based on three criteria were highly correlated with simulated values from a process-based model (CEVSA) at 0.5° × 0.5° resolution. At 1-km2 resolution, mean above-ground NPP in the region was 408 g/m2/yr ranging from 172 to 1091 (standard deviation (SD) = 134). Mean TNPP was 563 (252 to 1426, SD = 176). Ranges and SD were reduced while the mean values of the estimated NPP stayed almost constant as cell size increased from 1-km2 to 0.5° × 0.5°, as expected. Nordic boreal forests seem to have lower productivity among the world boreal forests. Abbreviations: ABIO = Above-ground biomass; AMSVI = Annual mean stem volume increment; ANPP = Above-ground net primary production; AVHRR = Advanced Very High Resolution Radiometer; NPP = Net primary production; RS = Remote sensing; SBIO = Stem biomass; TNPP = Total net primary production; TPB = Total plant biomass.

Land Cover and Soil Type Mapping From Spaceborne PolSAR Data at L-Band With Probabilistic Neural Network

This paper evaluates performance of fully polarimetric SAR (PolSAR) data in several land cover mapping studies in the boreal forest environment, taking advantage of the high canopy penetration capability at L-band. The studies included multiclass land cover mapping, forest-nonforest delineation, and classification of soil type under vegetation. PolSAR data used in the study were collected by the ALOS PALSAR sensor in 2006-2007 over a managed boreal forest site in Finland. A supervised classification approach using selected polarimetric features in the framework of probabilistic neural network (PNN) was adopted in the study. It has no assumptions about statistics of the polarimetric features, using nonparametric estimation of probability distribution functions instead. The PNN-based method improved classification accuracy compared with standard maximum-likelihood approach. The improvement was considerably strong for soil type mapping under vegetation, indicating notable non-Gaussian effects in the PolSAR data even at L-band. The classification performance was strongly dependent on seasonal conditions. The PolSAR feature data set was further modified to include a number of recently proposed polarimetric parameters (surface scattering fraction and scattering diversity), reducing the computational complexity at practically no loss in the classification accuracy. The best obtained accuracies of up to 82.6% in five-class land cover mapping and more than 90% in forest-nonforest mapping in wall-to-wall validation indicate suitability of PolSAR data for wide-area land cover and forest mapping.

Stand-Level Stem Volume of Boreal Forests From Spaceborne SAR Imagery at L-Band

This paper presents a modified robust stem volume retrieval approach suitable for use with L-band SAR imagery. Multitemporal dual-polarization SAR imagery acquired by ALOS PALSAR during the summer-autumn 2007 is used in the study, along with stand-wise forest inventory data from two boreal forest sites situated in central Finland. The average sizes of forest stands at the study sites were 3 ha and 4.8 ha. The method used employs model fitting with an inverted semi-empirical boreal forest model, and takes advantage of the multitemporal aspect in order to improve the stability and accuracy of stem volume estimation. Multitemporal combination of model output in a multivariate regression framework allows volume estimates to be obtained with an RMSE about 43% of the mean of 110 m3 /ha, and a coefficient of determination R 2 of 0.71 in the best case. The methodology used can be employed to produce large-area stem volume maps from dual-polarization ALOS PALSAR imagery mosaics.