The problem of feature selection is a significant one in classification problems, where the addition of too many features to the classification fails to lead to significant increases in classification accuracy. This problem is especially significant within the context of multitemporal remote sensing classifications, where the costs and efforts associated with the acquisition of additional imagery can be extensive. It would thus be beneficial to identify the most important seasons for acquiring imagery for specific land cover types. This study uses a phenologically-adjusted 21 date RapidEye time-series in order to evaluate two methods of feature selection. The two methods compared in this study are a genetic algorithm (GA) and a semi-exhaustive method (EXH), both of which compare permutations of sequential date and band combinations. These methods are employed using a seven class support vector machine classification on a Normalized Difference Vegetation Index (NDVI)-transformed dataset. Overall accuracy (OAA) is used as the performance metric, and OAA significance is assessed using the McNemar test. The results from the feature selection methods are compared on the basis of phenological seasons selected across all iterations and the ideal number of combinations, based on the ratio of better performing classifications to all other classifications. The results suggest that the GA has a moderate but insignificant correlation when compared with the EXH for identifying ideal phenological seasons (overall Spearman’s ρ= 0.60, p = 0.13), but is comparable when considering the number of seasons and image combinations.
The recognition and monitoring of vegetation and habitats for nature conservation is a vital point of research within the remote sensing community. It has been agreed on that there is no general solution on deriving information on habitats due to different data availability and spectral as well as textural behaviour of habitat main types (e.g. woodlands, grasslands, etc.). Therefore the monitoring should be rather multi-scale, versatile, user-friendly, and cost-efficient for predefined indicators.
In the presented study, five Central European test sites of natural vegetation communities in an Alpine area (1), a temperate forest (2), a Pannonian grassland (3), a shallow lake (4) and a Carpathian grassland (5) have been investigated by multi-temporal remote sensing. For these studies, different time-series of RapidEye images from the years 2009 to 2011 were acquired. The amount of the images was depending on required acquisitions dates as well as weather conditions. The definition of the indicators was relying on the available ground truth data as well as the demands and judgement of the managing authorities in the nature conservation areas. The selected methods for deriving of the indicators depend on the time-series as well as the available calibration and validation data. The techniques vary between unsupervised classification, object-based approaches and supervised classification methods with algorithms such as support vector machines (e.g. SVM) or classification trees (e.g. See5). Often the named methods are utilized in combined approaches.
The resulting indicators for the monitoring are shrub encroachment (for 1), share of naturally occurring tree type (for 2), differentiation of grassland types (for 3 and 5) and the changing extent of a reed belt (for 4). All indicators seem to be valid and useful. However, a transferability of the methods or a general statement on good-practice remote sensing applications can hardly be derived from these specific case studies.
Gathering vibrational data from the human middle ear is quite difficult. To this date the well-known acoustic probe is used to estimate audiometric parameters, e.g. otoacoustic emissions, wideband reflectance and the measurement of the stapedius reflex. An acoustic probe contains at least one microphone and one loudspeaker. The acoustic parameter determination of the ear canal is essential for the comparability of test-retest measurement situations. Compared to acoustic tubes, the ear canal wall cannot be described as a sound hard boundary. Sound energy is partly absorbed by the ear canal wall. In addition the ear canal features a complex geometric shape (Stinson and Lawton1). Those conditions are one reason for the inter individual variability in input impedance measurement data of the tympanic membrane. The method of Laser-Doppler-Vibrometry is well described in literature. Using this method, the surface velocity of vibrating bodies can be determined contact-free. Conventional Laser-Doppler-Systems (LDS) for auditory research are mounted on a surgical microscope. Assuming a free line of view to the ear drum, the handling of those laser-systems is complicated. We introduce the concept of a miniaturized vibrometer which is supposed to be applied directly in the ear canal for contact-free measurement of the tympanic membrane surface vibration. The proposed interferometer is based on a Fabry-Perot etalon with a DFB laser diode as light source. The fiber-based Fabry-Perot-interferometer is characterized by a reduced size, compared to e.g. Michelson-, or Mach-Zehnder-Systems. For the determination of the phase difference in the interferometer, a phase generated carrier was used. To fit the sensor head in the ear canal, the required shape of the probe was generated by means of the geometrical data of 70 ear molds. The suggested prototype is built up by a singlemode optical fiber with a GRIN-lens, acting as a fiber collimator. The probe has a diameter of 1.8 mm and a length of 5 mm.