Browsing by Subject "Image segmentation"

Sort by: Order: Results:

Now showing items 1-3 of 3
  • Hurskainen, Pekka; Adhikari, Hari; Siljander, Mika; Pellikka, Petri; Hemp, Andreas (2019)
    Classifying land use/land cover (LULC) with sufficient accuracy in heterogeneous landscapes is challenging using only satellite imagery. To improve classification accuracy inclusion of features from auxiliary geospatial datasets in classification models is applied since 1980s. However, the method is mostly limited to pixel-based classifications, and the coverage, accuracy and resolution of free and open-access auxiliary datasets have been poor until recent years. We evaluated how recent global coverage open-access geospatial datasets improve object-based LULC classification accuracy compared to using only spectral and texture features from satellite images. We applied feature sets topography, population, soil, canopy cover, distance to watercourses and spectral-temporal metrics from Landsat-8 time series on the southern foothills and savanna of Mt. Kilimanjaro, Tanzania, where the landscape is characterized by heterogeneous and fragmented mosaic of disturbed savanna vegetation, croplands, and settlements. The classification was based on image objects (groups of spectrally similar pixels) derived from segmentation of four Formosat-2 scenes with 8m spatial resolution using 1370 ground reference points for training, validation, and for defining 17 LULC classes. We built six Random Forest classification models with different sets of object features in each. The baseline model having only spectral and texture features was compared with five other models supplemented with auxiliary features. Inclusion of auxiliary features significantly improved classification overall accuracy (OA). The baseline model gave a median OA of 60.7%, but auxiliary features in other models increased median OA between 6.1 and 16.5 percentage points. The best OA was achieved with a model including all features of which elevation was the most important auxiliary feature followed by Enhanced Vegetation Index temporal range and slope degree. Applying object-based classification to geospatial information on topography, soil, settlement patterns and vegetation phenology, the discriminatory potential of challenging LULC classes can be significantly improved. We demonstrated this for the first time, and the technique shows good potential for improving LULC mapping across a multitude of fragmented landscapes worldwide.
  • Luotamo, Markku Ilkka Juhana; Metsämäki, Sari; Klami, Arto (2021)
    Semantic segmentation by convolutional neural networks (CNN) has advanced the state of the art in pixel-level classification of remote sensing images. However, processing large images typically requires analyzing the image in small patches, and hence, features that have a large spatial extent still cause challenges in tasks, such as cloud masking. To support a wider scale of spatial features while simultaneously reducing computational requirements for large satellite images, we propose an architecture of two cascaded CNN model components successively processing undersampled and full-resolution images. The first component distinguishes between patches in the inner cloud area from patches at the cloud's boundary region. For the cloud-ambiguous edge patches requiring further segmentation, the framework then delegates computation to a fine-grained model component. We apply the architecture to a cloud detection data set of complete Sentinel-2 multispectral images, approximately annotated for minimal false negatives in a land-use application. On this specific task and data, we achieve a 16% relative improvement in pixel accuracy over a CNN baseline based on patching.
  • Bussov, Maarja; Nattila, Joonas (2021)
    Computer vision and machine learning tools offer an exciting new way for automatically analyzing and categorizing information from complex computer simulations. Here we design an ensemble machine learning framework that can independently and robustly categorize and dissect simulation data output contents of turbulent flow patterns into distinct structure catalogs. The segmentation is performed using an unsupervised clustering algorithm, which segments physical structures by grouping together similar pixels in simulation images. The accuracy and robustness of the resulting segment region boundaries are enhanced by combining information from multiple simultaneously-evaluated clustering operations. The stacking of object segmentation evaluations is performed using image mask combination operations. This statistically-combined ensemble (SCE) of different cluster masks allows us to construct cluster reliability metrics for each pixel and for the associated segments without any prior user input. By comparing the similarity of different cluster occurrences in the ensemble, we can also assess the optimal number of clusters needed to describe the data. Furthermore, by relying on ensemble-averaged spatial segment region boundaries, the SCE method enables reconstruction of more accurate and robust region of interest (ROI) boundaries for the different image data clusters. We apply the SCE algorithm to 2-dimensional simulation data snapshots of magnetically-dominated fully-kinetic turbulent plasma flows where accurate ROI boundaries are needed for geometrical measurements of intermittent flow structures known as current sheets.