Wyniki wyszukiwania - BazTech

1

Zróżnicowanie kosodrzewiny w Tatrach, w świetle badań teledetekcyjnych

Zwijacz-Kozica M.

Teledetekcja Środowiska

|

2010

|

T. 44

1--61

EN

Dwarf mountain pine (Pinus mugo Turra) is the main component in the subalpine belt in the Tatra National Park, where the study was conducted. From the ecological point of view dwarf pine plays an important role in the sensitive mountainous area. Until now there were no studies focused on structure of dwarf pine community and there were also no attempts to work out methodology for detailed qualitative and quantitative description of dwarf pine. In this study for the first time it was aimed to prepare methodology of dwarf pine characterization and monitoring using hyperspectral data. Analysis involved processing of airborne and satellite images data and field measurements. Presented study evaluated linear predictive models between vegetation indices derived from radiometrically corrected air- borne imaging spectrometer ROSIS, spectral field and laboratory measurements and field measurements of dwarf pine biophysical variables (LAI, fAPAR). Narrow band vegetation indices were computed on the basis of all possible two-band combinations of set of vegetation indices (VI, NDVI, PVI, SAVI2, TSAVI). VI based on ROSIS wavebands 510 nm and 630 nm was linearly related to leaf area index (R2=0,48). VI and NDVI based on FieldSpec HH wavebands 886 nm and 518 nm performed better and were linearly related to LAI (R2=0,72). TSAVI based on ROSIS wavebands 658 nm and 570 nm was linearly related to the fraction of absorbed photosynthetically active radiation (R2=0,72). SAVI2 based on FieldSpec HH wavebands 747 nm and 703 nm was linearly related to fAPAR (R2=0,81). Analysed indices of vegetation condition were correlated (R2>0,90) with spectral vegetation indices based on FieldSpec Pro laboratory data. The study shows that for hyperspectral image data covering spectral region of visible light and near infrared, linear regression models can be applied to quantify LAI and fAPAR with satisfying accuracy. Models involving spectral information from sensors that have wider spectral range have better potential to linearly quantify biophysical vegetation parameters involving spectral vegetation indices. Vegetation indices that have the best relation to LAI and fAPAR were based on wavebands related to spectral features. It can be assumed that hyperspectral data contain information relevant to the estimation of vegetation biophysical parameters. In this study it was investigated if dwarf pine community differs spectrally within study site. To assess presence and extent of the spectral differentiation the set of field and laboratory spectral measurements were used. Reflectance curves were compared visually and using the statistical test. It was demonstrated that the majority of the studied dwarf pine plots have a characteristic signature. Parts of the electromagnetic spectrum which offer greatest information content for discriminating between and identifying dwarf pine spectral types were indicated. It was also examined if any of abiotic components of environment (altitude above sea level, aspect, slope, soil type, geology, global radiation and temperature) has an influence on the spatial distribution of LAI and fAPAR values. WMP (index of tie strength) and MP (tie strength) were used to assess an extent of the influence. It was found that neither of investigated abiotic factors affects LAI and fAPAR values.

2

Crop classification with neural networks using airborne hyperspectral imagery

Olesiuk D., Bachmann M., Habermeyer M., Heldens W., Zagajewski B.

Roczniki Geomatyki

|

2009

|

T. 7, z. 2

107-112

EN

Mainly due to size of input data, the artificial neural networks (ANNs) methods for remote sensing image classification can be expensive to use, in terms of computer resources and expert analyst time (Mahesh, Mather, 2006). In the case of hyperspectral data, neural networks training process may take weeks of time, in order to determine the number of input nodes in network structure needed by hundreds of image bands. In addition, not every neural networks package, such as the Stuttgart Neural Network Simulator (SNNS) used in this study, works with binary data, which makes dimensionality data reduction methods necessary to develop an effective classification scheme based on an ASCII text file. Despite these reservations, ANNs offer a wide field of research and investigation in crop and land cover classification, because they are a non-parametric method in the sense that they make no assumptions about the statistical distribution of the classes to be identified. As additional benefit, they can accept non-numeric inputs as well as ratio and interval-scale data. Moreover, the SNNS software provides the user a unique opportunity to design the input layers in a network structure, such as sub pattern window, which makes it possible to include texture information as additional data in the classification process (Zell et al., 1995). This method is especially useful in discrimination of non-homogeneous classes (Zagajewski, Olesiuk, 2008), and has been applied in this study. The objective of this work was to compare the results of crop classifications based on two data sets derived from hyperspectral HyMap imagery: (1) after MNF transformation, (2) vegetation and soil indices. The minimum noise fraction (MNF) transformation is used to segregate noise in the data, to determine the inherent dimensionality of the image data, and to reduce the computational requirements for subsequent processing (Boardman, Kruse, 1994). Essentially, it is two cascaded transformations. The first transformation, based on an estimated noise covariance matrix, de-correlates and rescales the noise in the data. This first step results in transformed data in which the noise has unit variance and no band-to-band correlations. The second step is a standard Principal Components transformation of the noise-whitened data. MNF bands are in a descending order of eigen values with almost no noise in the bands where the eigen values are near unity and below unity indicating signal-to-noise ratio (S/N) decreasing with decreasing order of MNF bands. The second data set contains hyperspectral indices which were selected to estimate pigment, nitrogen, cellulose and water content in vegetation, and clay and iron content in soil. The study area is located in the Demmin region in north Germany (Figure 1). This is a previously mapped agricultural area, where the main land cover/ land use types are represented by agriculture and grassland farming, with intermixed forestry and urban areas. This area is used as an agricultural and multi-disciplinary test site, and is included in the Committeee on Earth Observation Satellites (CEOS) catalogue for calibration and validation sites.

PL

Celem opracowania jest porównanie wyników klasyfikacji upraw uzyskanych ze zdjęć hiperspektralnych HyMap. Teren badań znajduje się w rolniczym regionie Demmin w północnych Niemczech. Do klasyfikacji wykorzystano dwa zestawy danych: 1) obrazy po transformacji Minimum Noise Fraction (MNF) oraz 2) mapy wskaźników roślinnych i glebowych. Transformacja MNF polega na redukcji wymiarów przestrzeni spektralnej (kompresji danych) i składa się z dwóch kaskadowych transformacji. Pierwszy etap polega na dekorelacji szumu, a drugi to standardowa transformacja PCA przeprowadzona na danych po oddzieleniu szumu. W rezultacie powstają nowe kanały, które uszeregowane są od największej do najmniejszej wariancji, przez co do dalszych prac mogą być wykorzystane najbardziej przydatne informacje. Drugi zestaw danych zawiera utworzone na podstawie obrazu hiperspektralnego wskaźniki roślinne i glebowe. Definiują one zawartość pigmentów, azotu, celulozy oraz wody w roślinność, a także iłu i żelaza w glebie. Klasyfikacja przeprowadzona została z wykorzystaniem sztucznych sieci neuronowych. Wykorzystano do tego celu oprogramowanie Stuttgart Neural Network Simulator (SNNS). Zastosowano sieć wielowarstwową, jednokierunkową, uczoną z użyciem metody wstecznej propagacji błędów (back- propagation errors). Klasyfikacje obu zestawów danych wykonano z zastosowaniem dwóch typów struktury neuronów w warstwie wejściowej. Pierwszy typ to struktura standardowa, gdzie liczba neuronów wejściowych odpowiada liczbie wykorzystywanych kanałów obrazowych. Druga struktura zaprojektowana została poprzez zdefiniowanie okna maski w postaci macierzy 3x3 piksele, dzięki czemu do procesu klasyfikacji włączona została informacja o teksturze badanego obiektu. Najlepszą dokładność całkowitą klasyfikacji wynoszącą 92,5% oszacowano dla zestawu zawierającego kanały wynikowe transformacji MNF i przeprowadzonej z wykorzystaniem struktury sieci odpowiadającej masce 3x3 piksele. Dla zestawu danych składającego się ze wskaźników roślinnych i glebowych dokładność klasyfikacji wyniosła około 80% w obydwu zastosowanych strukturach sieci.

3