We have demonstrated so far in this Tutorial that remote sensing is an efficient way to gather large amounts of information from vast areas without being on the observed surface. But, interpreters will seldom apply this knowledge effectively unless they have first-hand familiarity with the surface of interest, or at least, with models of the surface. They gather this intelligence by several means: from circumspect field observations, judicious investigations at training sites, sophisticated measurements in the laboratory, on the ground, in the air, and from space, and, ultimately, from a rigorous mathematical analysis of the data to test for validity and correlation.
In remote sensing, ground truth is just a jargon term for at-surface or near-surface observations. The term in its simplest meaning refers to "what is actually on the ground that needs to be correlated with the corresponding features in the scene (usually as depicted in a photo or image)". Some purists dislike the term - "reference data" is an alternative, but this fails to convey the sense of "onsite observations and measurements". As applied generally to any planetary body, the term connotes gathering data on-site and deriving information therefrom that properly characterizes states, conditions, and parameters associated with the surface. With appropriate sensors, we can derive aspects of the subsurface and any gaseous envelope (atmosphere) above it, as well. The purpose in acquiring ground truth is ultimately to aid in calibrating and interpreting remotely recorded surveys by checking realities within the scene. Since human interpreters normally experience Earth as ground dwellers, their view of the world from a horizontal or low-angle panorama is the customary frame of reference. In fact, the remote sensing specialist and the novice should retain a surface-based perspective during all phases of data collection, analysis, and applications, because they will implement most interpretations and decisions dealing with natural resources and land use at the ground level.
A few examples of the process of acquiring ground truth should clarify the paragraph above. Let us assume that we have a Landsat image. We see various features of interest within the image. Some we can identify by their shapes, spectral character, or other descriptors. Other patterns are of uncertain nature. If the image is of some region near us, we can actually get into a vehicle, drive into the region, locate ourselves relative to the pattern, and look about, actually spotting image features as their counterparts on the ground, and in most instances, recognizing and identifying them. With this information, we can return to our base, get the scene data in digital format, choose all or part of the image to classify, and specify the classes of interest by having found areas in the scene that they identify with. These become "training sites". The resulting classification map shows the distribution of the features/classes everywhere within the scene or subscene we choose to examine; included are the training sites and other areas assigned to each class. We can then go back into the "field" and visit a select number of other areas to find what is there. These areas either are what we expect if the classification is a good one or they aren't what the classification process says they are - they are misclassified, that is, their identity is erroneous. By tabulating these "hits" and "misses" we determine the accuracy of classification as a percentage of correct identities. Most such ground truthing is "after the fact", done in a time frame after the image was acquired.
But there is also the option of realtime ground truthing. On the day of a satellite pass - assuming that the user has a vested interest in studying the scene on that date - a field crew can be somewhere in the scene. Various tasks can be performed: taking spectral measurements of several classes, determining atmospheric properties, getting actual temperatures, specifying irradiances, determining bidirectional reflectances for key features, quantitatively estimating silt loads in passing rivers, measuring soil moisture. These are some of the varied data obtainable "on the spot" that could aid in image interpretation.
The writer (NMS) remembers distinctly two field experiences that are classic examples of obtaining ground truth. He and a colleague took a boat onto the Potomac River a day after the October 10, 1972 Landsat-1 overflight above Washington, D.C. The purpose was to measure the amount of mud and silt carried by the river right after runoff from a big storm the previous day. This served to calibrate the streamwater status in the image acquired on the 10th, one of the prime objectives of the investigation.
The other instance involved the writer (NMS) in the late summer of 1973, while going into the field to check on an unsupervised classification made by the LARSYS (Purdue University) processing system and to designate new training sites for a subsequent supervised one. The classified area centered on glacial Willow Lake on the southwestern flank of the Wind River Mountains of west-central Wyoming. Prior to arriving onsite, I prepared a series of computer-generated printouts (long since misplaced), in which an alphanumeric symbol represented each spectral class (separable statistically but not identified). Different clusters of the same symbols suggested that discrete land use/cover classes were present. In the processing, I allowed the total number of classes to vary. Printouts with seven to ten such classes looked most realistic. But there was no a priori way to decide which was most accurate. In touring the area, I had to revise or modify my preconceived notions about classes. I had not considered the importance of grasses and of sagebrush, nor anticipated clumps of trees below the 79 m (259 ft) resolution of the Landsat Multispectral Scanner. After a tour through the site, I gazed over the scene from a slope top and tried to fit the patterns in the different maps into the distribution of features on the ground. The result was convincing: the eight-class map was mildly superior to the others. Without this bit of ground truth, I would not have felt any confidence in interpreting the map and deriving any measure of its accuracy. Instead, what happened was a "proofing" of reliability for a mapped area of more than 25 square kilometers (about 10 square miles) through a field check of a fraction of that area. In this way my confidence in the capability of Landsat remote sensing data for identifying the contents of the real world was established. This improved even more when, back at Goddard, I reclassified the test site - this time by a supervised method - using the field observations to provide better training sites. The resulting map appeared to increase class accuracy by about 20%.
To re-iterate: The Table below summarizes the types of tasks and operations associated with obtaining and using ground truth data:
Correlate surface features and localities as known from familiar ground perspectives with their expression in satellite imagery
Provide input and control during the first stages of planning for analysis, interpreting, and applying remote sensing data (e.g., identifying landmarks, logistics of access. etc.)
Reduce data and sampling requirements (e.g., areas of needed coverage) for exploring, monitoring, and inventory activities
Select test areas for aircraft and other multistage support missions (e.g., underflights simultaneous with spacecraft passes)
Accurately survey ground control points to be used in image rectification and other geometric corrections
Develop standard sets of spectral signatures by using ground-based instruments
Identify classes established by unsupervised classification
Select and categorize training sites for supervised classification
Verify accuracy of classification (error types and rates) by using quantitative statistical techniques
Obtain quantitative estimates relevant to class distributions (e.g. field size; forest acreage)
Collect physical samples for laboratory analysis of phenomena detected from remote sensing data (e.g., water quality, rock types, and insect-induced disease)
Acquire supplementary (ancillary) non-remote sensing data for interpretive model analysis or for integration into Geographic Information Systems
Measure spectral and other physical properties needed to stipulate characteristics and parameters pertinent to designing new sensor systems
Examples of typical observations and measurements conducted in the field, commonly as the remote sensing platform passes over, or shortly thereafter, include these:
Ground truth activities are an integral part of the "multi" approach, discussed later in this Section. This simply means gathering the data under different conditions, such as the use of several sensors simultaneously and repeat coverage over time. We will explore the "payoffs" from this idea later in this Section.
Probably the most common reasons for conducting field activities are to select training sites prior to supervised classification or to identify key classes after unsupervised classification. The best way to collect field data, if feasible, is simply to spend a few days in the field, examining the terrain for the classification. Obviously, the scale of this effort depends on the size of the area we want to classify. One or more full Landsat scenes may require considerable travel and field time, whereas we can often examine a typical subscene (such as 512 x 512 pixels) in a day or two. If logistics or circumstances (e.g., an inaccessible foreign area or during an off-season such as winter) limit field operations, then we may use instead aerial photography, maps, literature research, interviews with residents (perhaps over the Internet), etc. In practice, to specify training sites generally means integrating the following sources of information: direct observations, photo documentation, a variety of maps, personal familiarity.Source: http://rst.gsfc.nasa.gov