• No results found

Change detection

In document List of Figures (sider 29-32)

Figure 2.7: SAR images of the same scene recorded in single polarisation(left)and quad-polarisation(right). Images from [28].

The added information content of polarimetric SAR can be appreciated in Figure 2.7.

2.2 Change detection

The introduction of the concept of CD for time series of remote sensing images dates back to the 1960s [30]. Even from its early definitions, it has always been referring to the detection or the assessment of both natural and human-caused phenomena affecting the Earth surface [4]. Singh [30] calls CD the process of identifying differences in the state of an object or phenomenon by observing it at different times.

2.2.1 What do we consider as a change?

On the contrary, defining which events should be highlighted as changes is still debatable, and the question whether a CD algorithm should also detect differences due to, e.g., weather conditions, seasonal trends or phenological processes is still open and highly application-dependent. Nonetheless, this ambiguity must be solved before proposing a CD framework, in order to evaluate its performance objectively. Arguably, a good definition should be flexible and adaptive, that is, when a change stands out over minor ones, the main event should be of major interest and the others should be ignored. For example, the growth stage of plants is an important aspect when monitoring

Chapter 2. Heterogeneous change detection in remote sensing 14 agricultural productions, but it should be irrelevant when assessing a forest fire aftermath. On the other hand, one may think of a more complex frame-work able to detect and distinguish all the diverse changes without discarding any [17, 21, 31, 32, 33, 34, 35].

2.2.2 Change detection methods pipeline

Traditionally, a CD framework usually consists of the three main phases listed below, as described in [6, 23, 24, 36]. Postclassification methods constitute a notable exception [16, 17, 31]. Although these methods may be fit for specific-purpose applications, they are generally considered as inferior due to the accumulation of error from the underlying classifications, approximated as the product of the overall accuracies of the individual classifications [23, 36].

Image preprocessing

The image rectification and restoration aims to correct distorted or degraded image data to create a more faithful representation of the original scene.

This typically involves the initial processing of raw image data to correct for geometric distortions, to calibrate the data radiometrically, and to eliminate noise present in the data. Thus, the nature of any particular restoration process is highly dependent upon the characteristics of the instrument itself.

These procedures are termed preprocessing operations since they precede further image manipulation and data analysis.

Geometric distortionsare both systematic and random: some are well un-derstood and mathematically modelled effects due to for example the previ-ously mentioned panoramic distortion, the Earth’s curvature, and the Earth’s rotation; others are caused by a wrong positioning and inclination of the sen-sor (most frequently happening to airborne and drone systems). To geocode and georeference an image means to take care of these problems and make sure that each pixel represents a well-defined position on the Earth. Coreg-istration is another fundamental preprocessing step: in order to perform meaningful analyses, one must bring all the images to a common spatial grid where a pixel represents the exact same area of the Earth in all of them.

Depending on the spatial resolution, this operation might require more than simple geometric transformations such as translations and rotations.

For optical data, also the radiometry degradation sources can be

distin-15 2.2. Change detection guished between systematic and random. The corrections of these account for Earth-sun distance and sun elevation to normalise the reflectance with respect to the seasonal position of the sun, but also for unpredictable atmo-spheric distortions. Finally,noiseremoval includes the restoration of missing lines (destriping), median filtering,multilooking and other techniques to im-prove the quality of the data before it is actually processed.

Change extraction

Once the images are ready for inspection, the next step is the extraction of change features: after a meaningful comparison of the images, the changes stand out from the background. Traditional CD methods are based on the comparison of homogeneous images, i.e. two or more images acquired by the same kind of sensor. Hence, the most logical and straightforward feature to consider when dealing with optical data affected by additive noise is the image difference, and the image ratio when dealing with SAR data and their multiplicative signal model. Clearly, the idea is to highlight the changes across images while removing the noise at the same time. For the bitemporal case, the result generally reduces to a difference image with a single value per pixel that represents to which degree (or probability) the pixel is likely to belong to changed areas. For a time series of N images, each pixel can be associated to N −1 values corresponding to the difference images between consecutive acquisitions.

Before proceeding with the next phase, a very common postprocessing step is filtering. Local, nonlocal, or global information can be used to smooth the difference image and further eliminate outliers caused by input noise or other issues. Without this procedure these pixels could turn into false positives or false negatives at the end of the CD pipeline. Examples range from simple local median filtering [37] to rather complex algorithms such as the Gaussian filtering that exploits fully connected conditional random field models [38].

Change image thresholding

Finally, the last operation required to distinguish changed parts from un-changed parts is thresholding the difference images or alternative test statis-tics. By splitting their histogram into two, thresholding allows to classify their pixels into changes (foreground) and no changes (background). The optimal thresholds can be set either manually after visual inspection or

au-Chapter 2. Heterogeneous change detection in remote sensing 16 tomatically by exploiting an algorithm such as [39, 40, 41, 42], or by using them in an ensemble fashion by a majority vote [43].

In document List of Figures (sider 29-32)