Guillermo Castilla

Geographic Object-Based Image Analysis (GEOBIA): A new name for a new discipline

What is Geographic Object-Based Image Analysis (GEOBIA)? To answer this we provide a formal definition of GEOBIA, present a brief account of its coining, and propose a key objective for this new discipline. We then, conduct a SWOT1 analysis of its potential, and discuss its main tenets and plausible future. Much still remains to be accomplished.

Image objects and geographic objects

Object-Based Image Analysis (OBIA) has gained considerable impetus over the last decade. However, despite the many newly developed methods and the numerous successful case studies, little effort has been directed towards building the conceptual foundations underlying it. In particular, there are at least two questions that need a clear answer before OBIA can be considered a discipline: i) What is the definition and ontological status of both image objects and geographic objects? And ii) How do they relate to each other? This chapter provides the authors’ tentative response to these questions.

Size-constrained Region Merging (SCRM): An Automated Delineation Tool for Assisted Photointerpretation

The manual delineation of vegetation patches or forest stands is a costly and crucial stage in any land-cover mapping project or forest inventory based upon photointerpretation. Recent computer techniques have eased the task of the interpreter; however, a good deal of craftsmanship is still required in the delineation. In an effort to contribute to the automation of this process, we introduce Size-Constrained Region Merging (SCRM), a recently implemented software tool that provides the interpreter with an initial template of the tobe-mapped area that may reduce the manual digitization portion of the interpretation. In essence, SCRM transforms an ortho-rectified aerial or satellite image (single or multichannel) into a polygon vector layer that resembles the work of a human interpreter, whom with no a priori knowledge of the scene, was given the task of partitioning the image into a number of homogeneous polygons all exceeding a minimum size. We provide background information on SCRM foundations and workflow, and illustrate its application on three different types of satellite images.

Development of a pit filling algorithm for LiDAR canopy height models

LiDAR canopy height models (CHMs) can exhibit unnatural looking holes or pits, i.e., pixels with a much lower digital number than their immediate neighbors. These artifacts may be caused by a combination of factors, from data acquisition to post-processing, that not only result in a noisy appearance to the CHM but may also limit semi-automated tree-crown delineation and lead to errors in biomass estimates. We present a highly effective semi-automated pit filling algorithm that interactively detects data pits based on a simple user-defined threshold, and then fills them with a value derived from their neighborhood. We briefly describe this algorithm and its graphical user interface, and show its result in a LiDAR CHM populated with data pits. This method can be rapidly applied to any CHM with minimal user interaction. Visualization confirms that our method effectively and quickly removes data pits.

A multiscale geographic object-based image analysis to estimate lidar-measured forest canopy height using Quickbird imagery

Lidar (light detection and ranging) has demonstrated the ability to provide highly accurate information on forest vertical structure; however, lidar data collection and processing are still expensive. Very high spatial resolution optical remotely sensed data have also shown promising results to delineate various forest biophysical properties. In this study, our main objective is to examine the potential of Quickbird (QB) imagery to accurately estimate forest canopy heights measured from small-footprint lidar data. To achieve this, we have developed multiscale geographic object-based image analysis (GEOBIA) models from QB data for both deciduous and conifer stands. In addition to the spectral information, these models also included (1) image-texture [i.e., an internal-object variability measure and a new dynamic geographic object-based texture (GEOTEX) measure that quantifies forest variability within neighboring objects] and (2) a canopy shadow fraction measure that acts as a proxy of vertical forest structure. A novel object area-weighted error calculation approach was used to evaluate model performance by considering the importance of object size. To determine the best object scale [i.e., mean object size (MOS)] for defining the most accurate canopy height estimates, we introduce a new perspective, which considers height variability both between- and within-objects at all scales. To better evaluate the improvements resulting from our GEOBIA models, we compared their performance with a traditional pixel-based approach. Our results show that (1) the addition of image-texture and shadow fraction variables increases the model performance versus using spectral information only, especially for deciduous trees, where the average increase of R 2 is approximately 23% with a further 1.47 m decrease of Root Mean Squared Error (RMSE) at all scales using the GEOBIA approach; (2) the best object scale for our study site corresponds to an MOS of 4.00 ha; (3) at most scales, GEOBIA models achieve more accurate results than pixel-based models; however, we note that inappropriately selected object scales may result in poorer height accuracies than those derived from the applied pixel-based approach.

The Land-cover Change Mapper (LCM) and its application to timber harvest monitoring in Western Canada.

We introduce an automated change detection and delineation tool for remote sensing images: the Land-cover Change Mapper (LCM). LCM rapidly generates a polygon vector layer (shapefile) of regions deemed to have undergone significant change in land-cover. In its simplest usage, LCM requires two single band or multi-band co-registered images of the same scene acquired at different dates, and as the only user-defined parameter, the minimum size for change regions. The main advantages of this tool are that (a) it is fully unsupervised, (b) it is exceptionally fast, (c) it is robust to geometric misregistration errors and variations in illumination, and (d) it produces visually pleasing outlines that resemble those obtained through manual digitization. We describe how the tool works, illustrate its application to monitoring forest clear-cuts on a 1,000 km 2 area in Western Canada using SPOT imagery, compare it to a commercial tool, and report on its thematic and spatial accuracy. A freeware LCM version is available on the Internet.

Pixels to objects to information: Spatial context to aid in forest characterization with remote sensing

Forest monitoring information needs span a range of spatial, spectral and temporal scales. Forest management and monitoring are typically enabled through the collection and interpretation of air photos, upon which spatial units are manually delineated representing areas that are homogeneous in attribution and sufficiently distinct from neighboring units. The process of acquiring, processing, and interpreting air photos is well established, understood, and relatively cost effective. As a result, the integration of other data sources or methods into this work-flow must be shown to be of value to those using forest inventory data. For example, new data sources or techniques must provide information that is currently not available from existing data and/or methods, or it must enable cost efficiencies. Traditional forest inventories may be augmented using digital remote sensing and automated approaches to provide timely information within the inventory cycle, such as disturbance or update information. In particular, image segmentation provides meaningful generalizations of image data to assist in isolating within and between stand conditions, for extrapolating sampled information over landscapes, and to reduce the impact of local radiometric and geometric variability when implementing change detection with high spatial resolution imagery. In this Chapter, we present application examples demonstrating the utility of segmentation for producing forest inventory relevant information from remotely sensed data.

The impact of object size on the thematic accuracy of landcover maps

We recently completed the accuracy assessment of a Landsat-derived landcover polygon layer covering the entire province of Alberta (660,000 km2), Canada, for which we gathered reference information for nearly 5000 randomly selected polygons ranging from two hectares to thousands of hectares in size. This gave us the unique opportunity to quantify, for the first time, how the probability of correctly classifying a landcover object varies with its size. Irrespective of whether they are represented as polygons or as sets of connected pixels with the same label, the classification accuracy of landcover objects decreases as their size decreases, steadily for large and medium sizes, and more dramatically when they are within two orders of magnitude of the pixel size of the input image. We show that this size-dependency is bound to occur whenever the size distribution of landcover objects follows an inverse power law. Our results are consistent with previous studies on related issues, confirm the need to account for size when assessing the accuracy of object-based landcover maps, and cast doubts on the validity of (1) recently proposed object-based accuracy estimators, and (2) landscape pattern analyses where the minimum patch size is close to the pixel size.

Size-constrained region merging (SCRM): a new segmentation method to derive a baseline partition for object-oriented classification

Object-oriented analysis of RS images for landcover mapping is based upon the same hierarchical patch model used in modern Landscape Ecology. In such model, each patch is a loosely integrated whole -- an object that can be viewed simultaneously as part of a superobject and as made of subobjects. The focal level, i.e. the level of the nested hierarchy on which the analysis is focused, is indicated by the minimum size that the objects of this level are supposed to have. Based on this framework, we have developed a segmentation method that defines a partition on a multispectral image such that each segment exceeds the minimum size required for patches of the focal level. The segmented image is subsequently used as the baseline for an object-oriented classification in which segments become the basic units. In our contribution we briefly describe the method, focusing on its region merging stage. The most distinctive feature of the latter is that while the merging sequence is ordered by increasing dissimilarity as in conventional methods, there is no need to define a threshold on the dissimilarity measure between adjacent regions. The initial segments are image blobs (defined here as tiny homogeneous regions, darker, brighter or of different hue than their surroundings), contoured by a morphological method (gradient watersheds). The merging process is conducted iteratively, allowing only one mergence per segment and iteration, and not allowing mergence when (1) one of the two segments to be merged has a neighbor that has been merged in current iteration; (2) both segments exceed the minimum size; and (3) one of both segments is smaller than this size but it has a more similar neighbor than the one under consideration. The method is illustrated with an example on a forested region in Spain.

Completion and Updating of a Landsat-Based Land Cover Polygon Layer for Alberta, Canada

Abstract We describe the creation of a GIS vector layer of land cover polygons for the entire province of Alberta, Canada, based upon preexisting, Landsat-derived, land cover raster datasets circa 2000 that were produced by the Canadian federal government. Our novel spatial and semantic generalization algorithm begins with a morphological segmentation of the original Landsat imagery used in the classification, and then assigns classes to the segments based on the land cover labels of pixels inside the segments, using sequential rules that account for contextual and size factors in addition to class preponderance. An object-based accuracy assessment followed, which allowed us to correct issues and refine the polygon map. The enhanced map, which was later updated to circa 2010 conditions using Landsat imagery from that time period and ancillary GIS information on natural and anthropogenic disturbances, consists of 1 million land cover polygons belonging to 11 classes and has an overall accuracy of 75%. This methodology could be employed in other jurisdictions with similar raster datasets to create a more intense spatial generalization than that provided by a conventional raster to vector conversion. Résumé . Nous décrivons la création d’une couche vectorielle SIG de polygones de couverture terrestre pour toute la province de l’Alberta, au Canada, sur la base d’ensembles de données raster de la couverture terrestre produits par le gouvernement fédéral canadien et dérivés de données Landsat acquises vers l’an 2000. Notre nouvel algorithme de généralisation (spatial et sémantique) commence par une segmentation morphologique de l’imagerie Landsat originale utilisée dans le classement. Ensuite, il attribue des classes aux segments basées sur les étiquettes de couverture terrestre des pixels à l’intérieur des segments à l’aide de règles séquentielles qui tiennent compte des facteurs contextuels et des facteurs de taille, en plus de la prépondérance de la classe. Par la suite, une évaluation de la précision basée sur les objets nous a permis de corriger des problèmes et d’affiner la carte de polygones. La carte améliorée, qui a ensuite été mise à jour pour les conditions observées vers 2010 en utilisant l’imagerie Landsat de cette période et des informations connexes du SIG sur les perturbations naturelles et anthropiques, se compose d’un million de polygones de la couverture terrestre appartenant à 11 classes et elle a une précision globale de 75 %. Cette méthode pourrait être utilisée dans d’autres régions avec des jeux de données raster similaires pour créer une généralisation spatiale plus intense que celle fournie par une conversion classique raster-à-vecteur.