BEGIN:VCALENDAR VERSION:2.0 PRODID:-//pretalx//talks.osgeo.org//foss4g-2022-academic-track//speaker//9W ZKHZ BEGIN:VTIMEZONE TZID:CET BEGIN:STANDARD DTSTART:20001029T040000 RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10 TZNAME:CET TZOFFSETFROM:+0200 TZOFFSETTO:+0100 END:STANDARD BEGIN:DAYLIGHT DTSTART:20000326T030000 RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3 TZNAME:CEST TZOFFSETFROM:+0100 TZOFFSETTO:+0200 END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT UID:pretalx-foss4g-2022-academic-track-VUQSVM@talks.osgeo.org DTSTART;TZID=CET:20220826T123000 DTEND;TZID=CET:20220826T130000 DESCRIPTION:Generalization is one of the fundamentals of scientific researc h. In the context of spatial information\, generalization needs to allow f or finding common properties but also for spatial contiguity. Therefore\, such generalization is often made through regionalization - partitioning o f space into spatial clusters or regions. This process is vital for enviro nmental studies\, where many patterns and processes are autocorrelated spa tially. Examples of regionalizations include delineation of ecoregions\, d etection of homogeneous zones for precision agriculture\, definition of cl imate regions\, and so on.\n\nTraditionally spatial generalization was per formed manually\, often based on a compilation of pre-existing\, independe ntly conducted studies. This approach lack of quantitative framework\, and thus no systematic checks\, modifications or objective updates are possib le. Currently\, the abundance of remote sensing spatial data\, such as sat ellite imagery\, gridded climate data\, or land cover maps\, allows fast e xtraction of relevant spatial information on regional and global scales\, making possible studies rooted in a clear quantitative framework.\n\nSuch data\, however\, still requires spatially-aware generalization to formulat e general concepts or claims. Remote sensing data stores information as a set of raster cells\, where a single cell is unaware of its spatial contex t. This is often not enough to understand underlying objects or processes. \n\n(Geographic) object-based image analysis (OBIA) (Blaschke 2010) is fre quently applied to resolve this issue. It is an approach to partition spac e consisting of raster cells into homogeneous objects and thus make spatia l regionalization possible. Several generalization techniques were develop ed for OBIA\, including a superpixels approach that proved to perform best for image processing and remote sensing data analysis (Csillik 2017).\n\n The main idea of superpixels is to create connected groupings of cells wit h similar values (Ren and Malik 2003\; Achanta et al. 2012). Each superpix el represents a desired level of homogeneity while at the same time mainta ining spatial structures. Superpixels also carry more information than eac h cell alone\, and thus they can speed up the subsequent processing effort s (Ren and Malik 2003\; Achanta et al. 2012).\n\nThe original superpixels algorithm has\, however\, two major drawbacks for spatial data problems ot her than RGB images. Firstly\, the algorithm uses the Euclidean distance\, which is adequate in many cases\, such as RGB images. However\, it limits the possible usability for environmental datasets – Euclidean distance is not suitable for many types of spatial raster data (e.g.\, categorical rasters) and has undesirable properties for multi-dimensional data (e.g.\, a set of monthly climate data)\, where the results based on Euclidean dis tance contradict human intuition (Aggarwal\, Hinneburg\, and Keim 2001). Secondly\, the superpixels technique does not result in regions per se but rather over-segmentation – some spatial objects/regions could be repres ented by one superpixel\, while others could consist of many very similar superpixels. \n\nOur preliminary results presented during the GIScience 20 21 conference (Nowosad and Stepinski 2021) provide a basis for using other distance measures to create superpixels. The proposed extension can also be used for various scenarios\, such as creating regions of similar multi- dimensional spatial and temporal patterns or similarly ranked areas. The e xtension is also already available as an open-source software in the form of an R package. The supercells package has extensive documentation in the form of a help file and additional vignettes that can be found\, together with its installation instructions\, at https://jakubnowosad.com/supercel ls/.\n\nThe second issue is\, however\, still not resolved. Many clusterin g methods exist that could be used for merging similar connected superpixe ls\, including traditional ones such as hierarchical clustering and spatia l-aware ones such as SKATER or REDCAP. Wang et al. (2018) developed a REDC AP-based workflow for merging superpixels\, which showed good image result s and outperformed similar techniques\; however\, their work was based on the original superpixels algorithm and thus used Euclidean distance on 3-d imensional RGB images only. Additionally\, it could be worth testing how g ood modern unsupervised machine learning techniques would perform in this task.\n\nOur main goal is to present the work in progress related to devel oping a robust method for merging superpixels and thus creating high-quali ty regionalization. We will test clustering/grouping methods based on thre e main criteria: accuracy\, universality\, and computational performance. Accuracy will be obtained based on the resulting regions’ internal homog eneity and their isolation compared to the neighbors. Universality will be tested on several datasets to check if the method works for various scena rios\, including RGB images\, categorical rasters\, spatial time-series\, etc. The computational performance will be evaluated based on the time nee ded for each method’s calculation and their use of computer resources.\n \nReferences\n\nAchanta\, R.\, A. Shaji\, et al. 2012. “SLIC Superpixels Compared to State-of-the-Art Superpixel Methods.” IEEE Transactions on Pattern Analysis and Machine Intelligence 34 (11): 2274–82. https://doi. org/f39g5f.\n\nAggarwal\, Charu C.\, Alexander Hinneburg\, et al. 2001. “On the Surprising Behavior of Distance Metrics in High Dimensional Spac e.” In Database Theory — ICDT 2001\, edited by Jan Van den Bussche and Victor Vianu\, 1973:420–34. Lecture Notes in Computer Science. Springer Berlin Heidelberg. https://doi.org/10.1007/3-540-44503-X_27.\n\nBlaschke\ , T. 2010. “Object Based Image Analysis for Remote Sensing.” ISPRS Jou rnal of Photogrammetry and Remote Sensing 65 (1): 2–16. https://doi.org/ d4ksqf.\n\nCsillik\, Ovidiu. 2017. “Fast Segmentation and Classification of Very High Resolution Remote Sensing Data Using SLIC Superpixels.” Re mote Sensing 9 (3): 243. https://doi.org/f92zgd.\n\nNowosad\, J.\, and T. Stepinski. 2021. “Generalizing the Simple Linear Iterative Clustering (S LIC) Superpixels.” GIScience 2021 Short Paper Proceedings. 11th Internat ional Conference on Geographic Information Science. September 27-30 2021. Poznań: Poland (Online). https://doi.org/gnw982.\n\nRen\, and Malik. 2003 . “Learning a Classification Model for Segmentation.” In Proceedings N inth IEEE International Conference on Computer Vision\, 10–17 vol.1. Nic e\, France: IEEE. https://doi.org/c6s237.\n\nWang\, Mi\, Zhipeng Dong\, et al. 2018. “Optimal Segmentation of High-Resolution Remote Sensing Image by Combining Superpixels With the Minimum Spanning Tree.” IEEE Transact ions on Geoscience and Remote Sensing 56 (1): 228–38. https://doi.org/gc t8gv. DTSTAMP:20260722T135856Z LOCATION:Room Modulo 3 SUMMARY:A method for universal superpixels-based regionalization (prelimina ry results) - Jakub Nowosad\, Mateusz Iwicki URL:https://talks.osgeo.org/foss4g-2022-academic-track/talk/VUQSVM/ END:VEVENT END:VCALENDAR