Catálogo de publicaciones - libros

Image analysis of sedimentary particles using backscatter electron (BSE) microscopy shows great promise in paleoclimatic and paleoenvironmental studies. Prior to the last few years BSE microscopy has been used primarily for compositional (provenance) studies. Our preliminary work on Paleozoic loessite, as well as previous work on recent sediments (Francus 1998; Francus and Karabanov 2000), suggests that BSE microscopy image analysis is an effective tool for deriving textural data for use as a paleoclimate proxy. Our data on the Paleozoic loessite shows that we are able to document changes in grain size of quartz through several loessite-paleosol couplets. In each case, the quartz was coarser in the loessite facies relative to the overlying paleosol, which is similar to grain size trends observed in the Quaternary Chinese Loess Plateau. Image acquisition is a critical step in this methodology, however, special precautions are needed to make sure that 1) the samples are suitably prepared, 2) the acquisition instrument’s settings are controlled and maintained, and 3) the acquisition system provides an output of suitable resolution. Processing is similar to other types of imagery subjected to image analysis, and includes calibration, filtering, and image segmentation and thresholding. An important component of processing is testing how different filters affect grain boundaries, particularly if grain size or grain shape is to be measured. In terms of image measurements, the magnification is an important consideration, and should be consistent; with standard BSE detectors, grains smaller than approximately 2 μm can not be resolved because of the size of the interaction volume of the backscatter electrons. As the case study illustrates, measurements of grain size or grain perimeter in this methodology do not translate into actual grain size information because of stereological considerations, however, relative changes in grain parameters yield useful information. The two biggest drawbacks of the present methodology are that it is difficult to keep acquisition conditions constant, and that data collection is time consuming. As instruments with BSE capabilities improve with more digital controls, acquisition will become much more stable, and as protocols are developed, it will be possible to semi-automate the procedure, allowing for a much faster rate of data collection.

This chapter looked at the use of automatic techniques for the analysis of microscopic objects. It showed all stages of the analysis, from the initial hardware set-up, through the focusing procedure and object detection, to the final classification of the objects. It presented a case study of a pollen classification system that can be trained on a sample of images of pollen grains. The system was based on a flexible neural network architecture: such networks trained on individual species can be combined to produce a discriminator for the set of species. The chapter presented results based on a trial with a large number of images. Experience in the face recognition community has shown that, for small data sets, it is easy to get good results; problems arise when attempts are made to scale up techniques to realistic amounts of data. Comparing the results of different techniques is also problematic. Large public datasets are required along with protocols for testing the recognition systems.

Conventional spectral analysis is mainly based on Fourier transform. This kind of transform provides excellent information in terms of frequencies (with their associated amplitudes) constituting the original signal, but does not keep the spatial information: it is possible to determine the elementary bricks that compose the signal, but not the way they are ordered along the signal. This limitation of the method has been noticed by Gabor (1946) who proposed a sliding window along the signal, in which the Fourier transform could be performed. In this way, part of the local (spatial) information is not lost. Nevertheless, this time-frequency tiling is still rigid and not really appropriate for natural complex signals. In the eighties, mathematicians introduced the concept of wavelet transform. The wavelet is a localized function, sort of a probe, capable of dilation (spreading out of the wavelet along the Oy axis) and translation (along the Ox axis). The transformation of the original signal by the wavelet results in coefficients, which are another expression of the signal. In addition, the wavelet transform acts as a mathematical microscope. In the discrete wavelet transform, two wavelets are used: the mother wavelet (the probe) and the scaling function. Therefore, it is possible to observe the signal at various scales, which is equivalent to the extent of the smoothing effect on the signal. This results in approximation coefficients computed by the scaling function. However, the mother wavelet will provide the detail coefficients. In conclusion, the signal is decomposed in two series of coefficients for each scale of observation. This extremely powerful tool has been used to detect cycles in the growth of lacustrine shells. By removing detail and/or approximation coefficients at different scales, and using image reconstruction (the wavelet transform has an inverse wavelet transform), annual, seasonal, tidal (monthly), and fortnight cycles in shell growth increments can easily be detected. Because of the very low amplitude of some of these cycles, it would not be possible to detect them without using the scaling effect and the detail coefficients associated with the lower scales. This method is much more powerful than the conventional Fourier transform when the aim of the study is to look for specific local periods and scale-sensitive information.