4. MARCO TEORICO
4.8 Recursos de la Biblioteca Escolar
4.1.4. Espacio físico e infraestructura de la biblioteca escolar
7.2.1 Introduction
In the main, in conventional geophysical survey over a defined area, the data will usually be handled and presented as a raster. A raster is an image made of cells, where the display properties of the cell are related to the value of the cell. By this means, it is possible to display the 400 readings taken in a 1 x 1 metre survey over a 20m Grid as a greyscale image where the darkness/lightness of each ‘cell’ in the image is governed by the reading that corresponds to that location in the survey Grid. Almost all digital images (including photographs) are rasters; they just vary in complexity, both in terms of the numbers of cells in the raster (commonly called pixels in digital photography) and in terms of the palette of colours used to display them. Figure 7.1 is an example of a simple binary raster, displayed next to a 20 x 20 raster of resistivity survey data.
In this research, the techniques that have been processed using this basis are the gradiometry, the area resistivity surveys, and the EM surveys. This is because all of these techniques, regardless of the actual depth they are sensitive to, or the depth information that can be inferred from them, essentially deal in two-dimensional information; each reading corresponds to a location on a single plane.
While there are a number of software solutions for image processing, specialised to various purposes, including several options specially made for archaeological geophysical data, GEOPLOT3 (Geoscan Research 2006) was selected as the primary tool. The main reasons for this decision were that for more than half of the surveys (especially taking into account the multiple resistivity data sets where the multiplexer was used) it is the ‘native’ software, developed by Geoscan Research specifically for their equipment, so the instruments could be directly downloaded into the programme.
It also required no investment in terms of learning a new programme inside out, and there was access to significant expertise using it in the department. As a programme specifically designed for archaeogeophysics, it has a number of data correction and processing options that would be quite hard to implement in other raster processing software. However, the publishing options are somewhat limited, so the data was exported to ArcGIS 9.1 (ESRI 2005) for the creation of figures, and to digitise the interpretation drawings.
7.2.2 Process overview
Generally speaking there are two aims to geophysical data processing, and two slightly different philosophies between them. First, it is often necessary to make corrections to the data: for example removing spikes in resistivity data (falsely high readings) caused by poor probe contacts, or staggering introduced in an incorrectly walked zig-zag gradiometry survey, where the paired readings are slightly out of step, due to different pace-lengths on the forward and back runs. It can also be necessary to compensate for unavoidable ‘errors’ in the data, for example, ferrous spikes and drift in gradiometer data, or mis-matched grids in resistivity survey resulting from having to move the remote probes and not getting a close background match.
Just making these minor corrections can render a dataset much more easily interpretable, and sometimes they are all that is needed. Sometimes, however, a dataset can be considerably enhanced by further processing. There is a very fine line between enhancing a dataset to the benefit of the archaeological interpretation, and introducing unnecessary processing that results in a ‘pretty’ image. In the latter there is a danger that misleading anomalies and features are created as a result of the processes, rather than reflecting any buried features. It is also equally possible to remove archaeologically relevant information from an image, either with correction processes or enhancement processes. Careful consideration and comparisons are needed at each stage to check for this. In the end, the resolution of archaeological information, rather than a consideration of aesthetics must be the governing principle in any data processing. Sometimes, therefore, it is best just to leave the data alone.
As a principle, and to work towards epistemic transparency, any geophysical data plot should be accompanied by a detailed account of the processes applied and the display properties used to produce it. In this work, these descriptions are located in Appendix A. This allows the process to be deconstructed, clearly demonstrating how
interpretations have been arrived at.
There are two main types of process that can be applied to a raster image. Point operators transform a single cell, based on its original value, to a new value in the resulting raster. Neighbourhood operators (also called convolution) examine the values of a given region around the cell being transformed (sometimes called a
window, or kernel) and use those values to determine the value of the central cell in the new raster. Neighbourhood operators are subject to edge effects, where the kernel is cut off by the proximity of the target cell to the edge of the image; unpredictable responses can happen, or the filter may simply be programmed not to run to the edge of the image.
When one or other of these functions is calculated for each cell in the raster, a new raster image is generated with the new values (see Figure 7.2). This simple principle can be used in an almost infinite number of ways, to do very useful operations on images, such as showing regions of rapid change, or performing adaptive contrast balancing on aerial photographs, allowing greater levels of detail to be recovered from the image. The following sections will discuss common correction and enhancement filters, and will state whether they are point or neighbourhood operators, or work on some other principle.
7.2.3 Data corrections
Most geophysical surveys require some corrections to the raw data collected in the field. This can variously be due to operator errors, minor instrument problems, or inconsistencies in the survey environment. Many of the filters used for correction and enhancement use the overall statistics of the dataset as the mathematical basis for their actions, so the first port of call (other than any changes that need to be made to
arrange the data correctly in the grid) is usually to remove outlying values from the image, to allow the parts of the image with the most variation to be displayed using a greater range of values. This is essentially improving the contrast of the image. In other image processing software, these might be transformations of the histograms such as a contrast stretch, or histogram equalisation process. In GEOPLOT there are two means of achieving this, either a neighbourhood operation called ‘despike’ or a simple clipping process which is a point operator. The despike tool allows the operator to define the shape of the kernel, in terms of the number of readings that make up the window in the x and y direction (as some surveys might have more in-line readings than traverses), and set a threshold above which the value will be discarded and replaced with the mean of the values surrounding it. The threshold is set as a number of standard deviations of the mean of the whole dataset. By changing the kernel size and the thresholds, increasingly harsh effects can be produced. The
image. It also does not work very well if the image has a large standard deviation; the threshold may be too low to capture all of the spikes in the image.
These problems can generally be overcome by using a point operator, (the clip
function) in GEOPLOT to set a minimum and maximum value (usually determined by the image statistics, but selected by the operator). The filter examines each cell in the image, replacing those that fall outside this range with the mean value in the image.
This can drastically reduce the spikes, but should be used with caution as the replacement value is the mean for the image, not the surrounding cells, and so may artificially reduce or raise some higher/lower areas of the image that reflect
archaeologically interesting variations.
These despiking processes can be used in tandem, with a clip function being used to
‘fix’ the image statistics and remove outliers, then despike being used to reduce noise in the image. Despiking tools should always be used first, after positional corrections, as if left in place the spikes can be concentrated, smeared or otherwise enhanced by the other filters and processes, potentially resulting in misinterpretations of the data.
They may be required in any type of two-dimensional survey, as spikes could result from noise introduced by the instrument, by modern material in the topsoil, or by poor probe contacts or changes in the ground surface.
Other corrections are more closely linked to the type of survey undertaken. In resistivity survey, the background resistivity values for adjacent grids may vary slightly due to the repositioning of the remote probes. This can easily be corrected with a point operation to add or subtract the required offset from all of the cells in the offending grid(s).
EM and gradiometry surveys might be subject to drift; a gradual change in the values over time resulting from systematic changes in the background (diurnal shift in gradiometer surveys) or the instrument warming or cooling (in EM surveys), that is unrelated to the actual values being detected and occurs incrementally over the grid.
One solution is to use a point operator that take the change across the grid (the far edge bias) and applies an incremental increase or decrease to each value in the cell to offset the imbalance at the same rate it occurred at. For data that in theory has a
central point of 0, i.e. gradiometer data, GEOPLOT also has a specific filter that would be very hard to replicate, the 0 mean traverse. There should be very little drift over one traverse of data, and GEOPLOT understands how the data was collected, and can correctly identify traverses in the data. The zero mean traverse looks at each such run of values, and adjusts each point in the sub-set to make the mean of the traverse zero. This removes any drift in the grid, and if applied to a whole dataset, any grid mismatches as well.
In the hands of inexperienced operators, gradiometer surveys can be prone to heading errors, particularly at the start and end of lines as the surveyor makes small changes in the orientation of the gradiometer when switching it on or off, or stepping over a grid tape. These can be corrected with a simple point process where the affected values are selected (perhaps those from the first or last meter) along each grid edge, and a simple addition or subtraction employed to bring the values back in line with their close neighbours. The exact value must be chosen by the operator with careful inspection of adjacent values to determine how much the heading error has biased the reading by.
They can also be subject to periodic errors caused by the gait of the operator, or perhaps a regular pattern of small height changes over a ploughed field. These can be removed by careful analysis of the frequency spectrum of the image, and filtering for specific components of the spectrum. These two corrections fall somewhere between point operators and neighbourhood operators, or use a combination of the techniques.
7.2.4 Image enhancements
Further processes may be applied to enhance the data, rather than just correct mistakes and survey problems. Typically, these might involve sharpening (high pass filters) or smoothing (low pass filters), both neighbourhood operators, to emphasise different aspects of the data. For example, resistivity data is quite often high pass filtered, which essentially preserves areas of rapid change and high contrast, and removes gradual changes. This serves to sharpen up potentially archaeological anomalies, whilst removing gradual background changes that are assumed to reflect geology-scale variations, but which are possibly swamping smaller, more localised changes.
Low pass filters are often used to smooth gradiometer data, as this can be quite visually noisy, due to small scale but large changes in the readings. These can obscure
operation reduces the noise and makes the image easier to interpret. In both cases, in GEOPLOT the user controls the intensity of the function by dictating the shape of the kernel applied to the data, and how the kernel elements are weighted. It is in using these processes that problems can arise, either with overly smoothed data, or with data with processing artefacts that look like archaeology. It is tempting to filter data to produce a smooth output that easy to look at, but there are times when the simply corrected dataset is equally informative, or perhaps even more so. There are a number of other filtering and enhancement options, but they were not needed for this research.
7.2.5 Other processes
GEOPLOT also has a number of tools not strictly for data processing. The most commonly used is interpolation; this can either expand or reduce the dataset, and is often used to increase the readings in one direction to match the other (for example a 1 x 0.25m gradiometer interpolated twice in the y direction to a 0.25 x 0.25m survey).
This works by the software inserting a new data point in between two values, taking its value from a combination of its neighbours. It significantly increases the file size and processing time, and is therefore often done in the final stages of processing as part of smoothing the image. Care must be taken to avoid introducing processing artefacts by this method, and smearing noise or spikes into apparent features. The process can also be used in reverse, to de-sample a survey, to allow direct comparison with another technique, or to combine two surveys collected at different reading intervals.
It is also possible to use selective filters to separate out areas of high and low resistivity, or positive and negative gradiometer responses according to user set thresholds, or generate contour plots of the data.
7.3 Ground penetrating radar
7.3.1 Introduction
The nature of GPR survey means that large volumes of data are collected and then analysed, especially when conducting area surveys with the intention of producing horizontal time-slices, as in this instance. The individual radargrams were not studied in great detail, or processed prior to the timeslicing as for this research, simple
timeslicing, as outlined below has proven effective. This three dimensional approach to the data was necessary due to the nature of the peat landscapes and multi-layered archaeology expected on the sites, particularly in the lowlands.
7.3.2 Timeslicing
Producing timeslices is a complex task with many stages. Timeslicing produces a number of plan views of the radar amplitudes at regular pseudo-depths through the collected profiles. The software used, GPR-SLICE (Goodman 2008) uses a subjective gain curve that is determined by the user visually and as such does not have a ‘value’
that can be reproduced in the Appendix (A) on data manipulation. The gain-curves used in the processing are retained with the dataset however to allow re-processing under the same parameters if needed.
The data is downloaded from the tablet as raw radargrams and imported into GPR-SLICE. The data is then converted which involves re-sampling the data to 32 in-line samples/m. The radargrams are then arranged in a Grid and where appropriate (as in the case of zig-zag survey) the readings in selected lines are reversed. Horizontal grids of data are then built from the data in the profiles. This is then interpolated into timeslices. The thickness, in terms of the time window averaged in the image, of the slices is decided by the operator. Unless otherwise stated the timeslices presented are the squared amplitude of the values for that particular stack of samples within the grid.
Once the slices have been created, they can be processed in ways similar to two-dimensional surveys such as low and high pass filtering, and histogram adjustments to correct the contrast of the images. One important function allows all of the images in any created dataset to all be displayed to the same histogram, meaning a particular shade represents the same squared amplitude in all of the slices, and the intensity of anomalies is preserved relative to each other in all of the images, rather than each image using its own greyscale.
7.3.3. Mosaic corrections
There are mosaic errors, that is, zones with different background signal responses and anomaly strengths (Ernenwein & Kvamme 2008; Goodman 2008, Sections XV.A & E) in some of the radar datasets caused by the survey being done on a number of
different dates, sometimes weeks apart and therefore under different conditions. A
number of the suggested processes for dealing with these have been attempted, and the most satisfying result has come from applying a filter at the stage immediately after slicing the data, but before producing any gridded datasets, that creates a zero mean for each line of data. This function has a threshold based on a certain percent of the values in a line. The most satisfactory results have been with it set to ignore the top 50% of all of the values when calculating the average. This means anomalies (and especially linear anomalies in the survey direction) are more likely to be preserved, and a better background match achieved. It is similar to the zero mean traverse function used in GEOPLOT to correct for drift. This has not totally removed the mosaic problems within the dataset, but is comparable, in terms of the visible
anomalies as processing each block of readings collected on the same day separately, which was done for one of the datasets to make a comparison between the two techniques for dealing with this common problem.