PhD topic thesis: Spatio-temporal patterns mining in image time series.
Remote sensing imagery is a type of raster data that is being collected on a regular basis by a large number of sensors on board of different satellite platforms. Remote sensing imagery is being represented by multi-dimensional gridded data, with two spatial dimensions related to the location of the value measured by the sensor (spatial resolution), and other dimensions related to the spectral range for the value (spectral resolution) and the temporal dimension related to the time when the data was acquired. These dimensions are considered as parameters of imagery data and vary depending on the sensor.
Traditionally, the analysis of remote sensing image databases is done in a fixed-time or scene-by-scene way, but one scene is a snapshop of a constantly changing landscape. The temporal aspect of this dataset is not fully exploited preventing us from detecting patterns and processes related to change that are fundamental in the context of the two application areas being addressed in this joint research project. As a motivational example, we consider the Program for Deforestation Assessment in the Brazilian Legal Amazonia (PRODES) using remote sensing images and digital image processing techniques. In the past each image (a snapshot) was fully analysed not considering past images, in order to detect the deforested areas. Now the deforested areas detected in the past are used to mask out areas that don’t have to be analysed again, thus reducing the volume of data to be dealt with. But yet we cannot detect spatio-temporal patterns of change over time.
The following research questions arise from this representative example: how can we characterize patterns of change in image time series? Which general methods for time series analysis should be developed/adapted in order to characterize patterns of change in image time series? Another challenge is how to integrate them in database systems developed to handle massive volumes of image time series. These questions will be addressed in a PhD thesis carried out at the Applied Computer Graduate program at INPE. The experiments will be conducted on the large collection of remote sensing images available at INPE, which contains large time image series with different spatial and spectral resolutions, using and expanding on the image database tools available in the TerraLib library.
References:
Possible supervisers: Gilberto Câmara (INPE), Lúbia Vinhas (INPE), and Edzer Pebesma (IFGI).
Mobility measures: as this work will be based on the complementarities from IFGI (statistical analysis of time series) and INPE (image databases), we foresee that the student will spent at least 3 to 6 months at IFGI in his/her second year.
