Remote sensing studies have always shown various discrepancies as researchers have always identified errors and uncertainties in image interpretations. It also highlights that these properties can be human induced due to mis-interpretation of the data, miss handling of equipment (calibration problems, poor decision making in when to take data, if cloud problems.
However, these issues can be technological problems in the sub-orbital and orbital craft; wrong flight path, poor sensor range.
The objective of this essay is to highlight the certain types of errors, the causes of uncertainties and scale problems in remote sensing. The extent of this area of terms is substantial in terms.
Error, according to Heuvelink(1991) is defined as “the difference between reality and our representation of reality”. Error produces ramifications, as “error is a bad thing”. However Heuvelinks definition does not account for subtle and random errors “in a statistical way”. The concept of error is illustrated by Jensen’s belief that error is based on two factors internal and external geometric error. Jensen (2005) backs this up by stating it is important to recognize the problem sources of internal and external error, “whether it is systematic or nonsystematic”. Apparently, geometric error of the systematic approach is generally easier to define and correct.
The problems of internal geometric errors are caused by the Earth’s curvature characteristics and remote sensing system being used.
Jensen breaks down where internal geometric errors happen
2. Scanning system
3. Relief Displacement
Skew effects happen because of the remote sensing data being affected by the earth’s rotation on the axis and the remote sensors orbit of the planet whether it is phased, sun synchronous or geosynchronous. This combination of factors causes the image of the IFOV being examined to be skewed, this was based on using the LandSat enhanced thematic plus using its linear array of 16 lines, which scanned 3 times. This skew effect could be the result of a faulty scanner. Jensen (2005, states if the image is not deskewed, the data will be displayed incoherently. The belief is that the image is skew in an eastwardly approach. The deskewed images being scanned will have an abrupt change on the pixels being read by the scanner. Every orbital sensor that collects image from the Earth will incur skewed images due the spectral curve of the land and also because of the overlaps in images.
Since a large amount of data is retrieved from various scanning sensors aboard orbital and sub-orbital craft. Jensen makes the comparison that multi-spectral orbital craft have minimal distortion to multi-spectral sub-orbital craft. This distortion is reduced in orbital craft because of their nadir equipment, altitude and IFOV in the terms of swath. Jensen(2005, p 227-235) places the problems of geometric distortion on aircraft because of there AGL and their operating height. The only way that this distortion is reduced is only by using the central 70% of the swath width, because scientists noticed “ground resolution elements have larger cell sizes the farther way from the nadir”
The use of aerial photography illustrate that photographs are exposed to perspective geometry , where all the objects are displaced from their plan metric positions outwardly from the principal point. The greater planimetric distance the greater, the reliefs distance(Jensen, 2005, p 227-235). The displacement occurs because of the direction that is perpendicular to the line of flight of each scan. The problems of one-dimensional relief displacement cause objects from their terrain to be displaced from their true position. This causes maps created from imagery to contain plainimetric errors.
Externals errors happen because of unexpected ramifications “in nature through space and time”. The widely known errors of this are; altitude alterations and attitude changes, which can be yaw pitch and roll.
The majority of remote systems operate well above AGL so that images can produce a uniform pattern. If the orbital or suborbital craft changes any of its orbits along the designated flight path, this will result in the scale of the image to change. Jensen(2005, p 227-235)), points out that these changes happen due to the elevation of the aircraft and the terrain. The only ways these issues are corrected by are the use geometric rectification algorithms.
The problem of sub-orbital craft are that they affected by issues of turbulence and wind. This happens when sub-orbital craft are collecting data and have to contend with up and down drafts, cross, tail and head winds. This results in the aircraft changing its flight path by rotating its various axes (roll, pitch and yaw. This intern causes geometric distortions to image by introducing compression and expansion of the image. Most satellite and aircraft used in remote sensing have gyro-stabilization equipment to offset these errors in their flight paths. The problem can also be located in mishandling of equipment, as sensors can be miss-calibrated for recording information.
The remotely sensed data can attribute to uncertainties in the processing outcomes, which in turn can affect the sound decision-making. Uncertainty pertains to areas of inference and prediction (May,2001). Researchers often confuse these terms in remote sensing. Uncertainty plays a large part in remote sensing as it causes problems especially the classification of land types, Measurement of Sea Surface Temperature, Image interpretation, Image mapping and many other areas of remote sensing. Uncertainty comes from many sources such as ignorance, “through measurement of prediction”. Uncertainty relates to being not 100% sure of something. The problems in uncertainty are that sometimes it is the most exciting in remote sensing. This is where remote sensing can explored “to find things so that a base is made for better understanding of how the world operates. The problems of uncertainty can be located in AVHRR data used for measurements of vegetation can have levels of inaccuracy. The best example where uncertainty can cause calamities was the Kyoto protocol in 2000, which collapsed due to high levels of uncertainties in the measurement and understanding of carbon emissions.
Woodcock(2002) relates uncertainty to three areas
Accuracy is often described as the closeness of results, observations that correspond to values being accepted as being true
Bias is often seen as an over-estimation of a true value. Bias is often model based.
Precision is often the exact value expressed whether the value is right or wrong.
These three terms are often seen in projects of measurement of changes in ice sheets, Tropical forests and land classification. Grant & Leavenworth (1988) emphasis “during the life cycle of remotely sensed data, uncertainties are introduced and propagated in an often unknown way”. By constructing a listing where an uncertainty occurs provides the factors of how uncertainty is found in remote sensing.
The sensor system.
Using the landsat as an example, the velocity of the scanning optical mirror, the number of spectral bands, orbit and altitude height are, parameters that determine the signal noise and “goodness” of the measurement. In addition to this, all these factors are translated in the terms of resolution; spatial, spectral, temporal. The data gathered from the sensor are affected by these characteristics. This happens in multi temporal systems, especially tropical regions, where there is high cloud cover, this results in incomplete data sets.
The problem relates to a sensor trying to correctly capture areas with high complexity. The detection of objects in mixed classifications can be difficult with areas of land and urban together. These objects can affect the area being examined as spectral influence could offset sensors reading in IFOV, although this can be improved by high spatial resolution. In all, this causes the image to become fuzzy and results in uncertainty if the pixels values in the image do not correspond to the classification structure, as two pixels of different values represent the same type of land cover.
Geometric and atmospheric distortions.
Jensen (2005) illustrates that geometric cause many errors in raw data interpretation especially when affected by the height and altitude of a sensor or by the ground control points(GCP). The concern with atmospheric distortions is that electromagnetic radiation interaction with the atmosphere can diminish information signal of a sensor. This happens, as radiation is subjected to back scattering and absorption. This causes some uncertainty in capturing a truly perfect image of a flat representation.
The processing of data causes a lot of uncertainty as the amount uncertainty allowed in data is based on
• Success of the radiometric and geometric corrections
• The loss of data during conversion
Scale variations have long been a thorn in remote sensing, as scale can be constraints for detail in which information needs to be observed and analyzed. Altering the scale in image causes the representation of patterns to differ from the actual size. Scale (Maher, 1997) is used as a basis for measurement as a scale of 1:1,000 means for every one unit on the map, you would need to measure one-thousand of those units on the earth. For example, 1mm on the map represents 1,000 mm on the ground and 1 metre represents 1,000 metres. Although these relations of unit measurement can cause problems as some users may miss-interpret a measurement unit, as 1:10000 is a large scale in comparison to a 1:100000(l.
Lillesand(2004, p 617-622) belief is that scale can be very problematic in remote sensing as the definition and understanding varies from researchers. Lillesand breaks down scale in to two areas temporal and spatial. In remote sensing scale is an ambiguous term and is often defined as the relationship between the size of feature on a map or image to the corresponding dimensions on the ground. However ecologists understanding of spatial scale is based on two factors, grain (finest resolution of data) and extent (area under observation). Other problems of spatial scale are the understanding of large and small scales in an image. This is evident as small scale represents coarser spatial objects while large scale represents clearer objects, however individualist (ecologists) reverse the meanings of small and large scale for image analyzing.
The only way ensuring the scale of the image is correct for research is using the three principles that Lillesand(2004,p 622) created.
1. Spatial resolution of the sensor
2. The spatial area under observation
3. The nature of information sought in any given image processing operation
These key factors should always be relevant in deciding which type of sensor to be used for image analyzing of an area in spatial scale.
However, scale varies have changed since the deployment of imaging systems on satellites. Sabins (1997) show this.
Small scale 1:500,000 as 1cm= 5km or moiré
These descriptions are different from the aerial photography. However, Harvey and Hill (2003) see that detailed large-scale date extracted from aerial photography was superior as similar scale data could not be extract from the Landsat TM, SPOT XS satellite. This was due to errors of spectral data found in the classification for different vegetation covers.
The best available photographic data are panchromatic and at scales of 1:25 000 or 1:50 000.
It is generally believed that all environmental processes are scale dependent. The different scale measurements manifest a homogenous side to one scale and a heterogeneous to another side(Atkinson & Tate,2000). This can be evident as scale dependence on spatial variation can be problematic in the processing of data, as the techniques of averaging, smoothing and extrapolation can be dangerous for replacing missing datasets. In spatial scale it is often desirable to focus on the particular scale of spatial variation (mean and sample of data), as not all scales need to utilized for specific process(Atkinson & Tate,2000). However that can problem if using a specific scale such as the drainage basin scale to represent contours on a topography map, however it would be useless for sheet flow on a hill slope. The general belief is, that it is widely accepted that scales of measurement are determined by the sampling strategy, the sampling generally refers to spatial pattern of the sample observation. Unfortunately, all samples of spatial data are not stationary and can affect the spatial scale of the area under study(Atkinson & Tate,2000). That is the fundamental reason for the continuing interest in scale in remote sensing is that spatial resolution is the primary scale of measurement (Atkinson and Aplin).
Zhou & Liu(2004) says that classification of land produces errors in multi- temporal data acquisition. The use of classification causes many problems in mapping of land and urban areas which results in uncertainty of accuracy. Using spatial resolution, a decrease in pixel size can cause major ramifications in an image being inferred for a prediction model (H. Liu & Q. Zhou). This is evident as pixels being inferred in a spatial swath may represent more than one classification. This can found by comparing different satellites spatial resolution. As landsat TM uses a 30 * 30 metre resolution that is acceptable for some classifications, however if the resolution was needed for small buildings (5 x5 metres), it would fail. This also highlights that uncertainties can enter in to classification of land types as distinguish uncertainties can happen (Atkinson & Foody), Page 14).
• The occurrence of ambiguous definition of classes
• The problems of land transitions over period due to rapid changes of building into new land types.
There also more sources of uncertainty that can occur during the post classification stage as different users can create different conclusions from the IFOV under observation. This is a result of their map reading skills, understanding of raster information and the ability to distinguish significant objects.
Neels (2005) defines that uncertainty is found in most of the big ice sheets. This happens due to poor observation of spatial coverage of measuring their variable changes in changes in surface elevation. This was due to the implementation of the budget method. This method is insufficient because of the fluctuation changes in ice sheet volume are determined by the snakk residual alterations in the large terms. This causes temporal and spatial fluctuations, which then cause uncertainty in measuring the ice sheet volume. Not all these measurements in high accuracy and dense data can provide a reliable estimate of ice sheet volume and this highlight errors in analysis of the surface elevation measurements. It also manifests the scale problem of obtaining this volume of data.
However Harvey and Hill (2003) notice, that errors are common in image acquisition in remote sensing studies of wetland areas. They believe that spectral overlap may cause a reduction in the utility of imagery collection of a certain season, although spectral classification can be enhanced by image optimizing. This is manifested by their work in tropical wetland environments, as tropical wetlands have large variations in the nature of rainfall. These factors can affect the overall classifications and cause various errors to the image been inferred.
Errors can be found in most areas as Czaplewski (2003) questions the methods used in remote sensing especially in monitoring the of global deforestation trends. He manifests this by agreeing with the arguments of Townsend and Tucker over the 10% stratified random samples used by FAO to estimate tropical deforestation. These samples are quite questionable as tropical deforestation is spatially concentrated. The belief here is that these samples can have significant problems for estimations for the FAO. This was seen in study models of dense tropical regions and showed that some regions are less spatially concentrated. Therefore this provided that sampling errors from Landsat sensor scenes in these studies are higher that other regions.
In Rajeev and Saxena (2004) argument, they conclude that large scale soil mapping of different scales are dependent upon the requirements of the user. However the uses of satellite have being problem for large scaling mapping because of their coarse resolution. The only way this scale of operation was feasible was to employ conventional methods, which were costly and time consuming. This has changed with the introduction of high resolution PAM and LISS III data from IRS-1C/CD satellite. Although scale is seen mostly as a measurement, scale can be used as a term in remote sensing as the area of data to be analyzed .This was seen with the central America dataset as the scale of coverage was 619 048km2 and includes the countries of Belize, Guatemala, El Salvador, Nicaragua, Costa Rica and Panama. However the assessment of this scale of mapping proved to be difficult, as it show the limits of AVHRR data for classification of NDVI in comparison to Landsat TM data, which has higher classification accuracies (Friedl et al, 2000).
It also evident that the temporal scales of these studies have been relatively short (days, weeks, months), and few studies have exceeded years in duration. Despite this, researchers are now being called upon to lengthen their studies to longer temporal scales.
The use of these terms illustrate the problems that are found in remote sensing and highlights the problems they cause to images being taken from orbital and sub-orbital craft. It demonstrates the issues of pixel classification of data in images for the representation of land, woodland, urban and water classes.
It also raises the concern about the understanding of these terminologies in remote sensing as some individuals lack the true meaning of what these terms are. That is why Woodcock (2002) states that “where relevant, adopt the terminology used within statistics and otherwise should adopt terms that convey clearly the authors meaning.” Scale apparently has being problem for researchers as they are concerned in deciding which scale will provide the most accurate measurement in spatial. The problems of uncertainty and error is also seen as new techniques such as using different types of sensors, prediction models are still not capable of yielding the trends of ice sheet mass balance, unless decades of observation are made. Error can also be human induced to poor understanding and misinterpretation of data, technology and the area of observation. It appears that it is by far unfeasible to assess all errors , uncertainties and scale problems as they will always appear in studies, observations and new types of equipment in remote sensing, as nothing can be 100% perfect in any environment.
Finally the biggest concern is the constant changes of electromagtic radiation as this causes major problems to remote sensing when attaining information on a large scale region of interest. these changes in the spectrum will cause uncertainties for the researcher and the scientist.