Segmentation and Classification of Remotely Sensed Images: Object-Based Image Analysis

Land-use-and-land-cover (LULC) mapping is crucial in precision agriculture, environmental monitoring, disaster response, and military applications. The demand for improved and more accurate LULC maps has led to the emergence of a key methodology known as Geographic Object-Based Image Analysis (GEOBIA). The core idea of the GEOBIA for an object-based classification system (OBC) is to change the unit of analysis from single-pixels to groups-of-pixels called `objects\u27 through segmentation. While this new paradigm solved problems and improved global accuracy, it also raised new challenges such as the loss of accuracy in categories that are less abundant, but potentially important. Although this trade-off may be acceptable in some domains, the consequences of such an accuracy loss could be potentially fatal in others (for instance, landmine detection). This thesis proposes a method to improve OBC performance by eliminating such accuracy losses. Specifically, we examine the two key players of an OBC system : Hierarchical Segmentation and Supervised Classification. Further, we propose a model to understand the source of accuracy errors in minority categories and provide a method called Scale Fusion to eliminate those errors. This proposed fusion method involves two stages. First, the characteristic scale for each category is estimated through a combination of segmentation and supervised classification. Next, these estimated scales (segmentation maps) are fused into one combined-object-map. Classification performance is evaluated by comparing results of the multi-cut-and-fuse approach (proposed) to the traditional single-cut (SC) scale selection strategy. Testing on four different data sets revealed that our proposed algorithm improves accuracy on minority classes while performing just as well on abundant categories. Another active obstacle, presented by today\u27s remotely sensed images, is the volume of information produced by our modern sensors with high spatial and temporal resolution. For instance, over this decade, it is projected that 353 earth observation satellites from 41 countries are to be launched. Timely production of geo-spatial information, from these large volumes, is a challenge. This is because in the traditional methods, the underlying representation and information processing is still primarily pixel-based, which implies that as the number of pixels increases, so does the computational complexity. To overcome this bottleneck, created by pixel-based representation, this thesis proposes a dart-based discrete topological representation (DBTR), where the DBTR differs from pixel-based methods in its use of a reduced boundary based representation. Intuitively, the efficiency gains arise from the observation that, it is lighter to represent a region by its boundary (darts) than by its area (pixels). We found that our implementation of DBTR, not only improved our computational efficiency, but also enhanced our ability to encode and extract spatial information. Overall, this thesis presents solutions to two problems of an object-based classification system: accuracy and efficiency. Our proposed Scale Fusion method demonstrated improvements in accuracy, while our dart-based topology representation (DBTR) showed improved efficiency in the extraction and encoding of spatial information