Multi-scale Geometric Analysis Approaches For Automatic Extraction Of Linear Geospatial Objects From High Resolution Remote Sensing Images

Geospatial information system has become a tool for broad range of applications including complex real world problem analysis.Now days,images derived from remote sensing systems are the primary sources of heterogeneous geospatial data with different levels of detail.The current improvements in sensor technology enabled the acquisition of fine resolution images that can help detection of very fine geospatial objects more than ever before.Wide varieties of geospatial objects are largely associated with lines.Natural or manmade spatial features are rich of lines.The high resolution remote sensing images are novel sources of such objects.The need to extract such linear objects from the images requires effective digital image processing approaches and algorithms.The focus of this dissertation is determining approaches for accurate automatic extraction of linear geospatial objects from remotely sensed high resolution images.Multiscale and perceptual grouping based methods are proposed.Two groups of multi-stage approaches corresponding to two linear geospatial objects(1)the power lines and(2)the roads are prposed.Based on the proposed approaches a review of the theoretical and technical aspects of the multiscale and visually inspiring perceptual grouping approaches with respect to the proposed linear geospatial object are explored.Experimental investigations were conducted on sample images in the MATLAB? environment.For power lines extraction,the complementary multiscale multi-orientation filters: the Gabor filter,the Jerman filter,the matched filter and the Multiscale Geometric Analysis(MGA)tool called the curvelet transform functions are proposed.Theree high resolution image datasets: the unmanned aerial vehicle images,on-board helicopter captured aerial images and Google Earth maps of totally 54 images were used for experimental investigations.The Gabor filter for its multiscale and multi-orientation selective capability was exploited.With the adjustment of its parameters,it was found that it can enhance the power lines structures.The Jerman enhancement tool has also investigated for its elongated object enhancement capability.This has also been found with interesting qualities for its scale and orientation based enhancement qualities.The Gaussian based matched filter has also investigated as an alternative to the Jerman enhancement function and found relevant for its strong noise suppression.The curvelet transform by decomposing the image in to different scales and orientations allowed accessing the objects in the image based on their scale and orientation.By employing selective modification operations to the objects in the image(that is,preserving the target power lines and removing the other unnecessary objects)a well-established power line structures ready for extraction were derived.A method that can implement both connected component analysis and thresholding called the hysteresis thresholding was used for the final segmentation and extraction of the enhanced power lines.The holistic multi-stage experimental investigations proved that the proposed approach is effective.The power lines in the images derived from multiples of sources captured in any orientation even from extremely degraded images were accurately derived.For the extraction of the roads,the integration of multi-stage multiscale geometric analysis and perceptual grouping based approaches were proposed.The Pulsed Coupled Neural Network(PCNN),the curvelet transform and the tensor voting approaches are proposed and used with hysteresis thresholding.The approach is examined experimentally on three high resolution image datasets of 28 total images.The IKONOS and Geo Eye satellite images and the Google Map derived images are used for the experimental investigations.From the experimental analysis,it has been confirmed that the pulsed coupled neural network effectively derived candidate road segments.The pulsed coupled neural network even separately discriminated road groups of different intensity from the same image.The curvelet transform with its multiscale and multi-orientation decomposition capability allowed the road segments to be accessed,preserved and enhanced separately.With preliminary hysteresis extraction,the roads are sufficiently extracted.The tensor voting for its missing data estimation capability used to infer road gaps resulted from the complex nature of the original images and the factors associated with the images that the previously used methods failed to recover.The experimental investigation showed that it is an effective method to recover road network accurately.A topologically integrated and georeferenced road network for geodatabase is accurately derived.For both the approaches qualitative and quantitative performance measurements of the experimental outputs are proposed.Qualitatively,the extraction outputs are adequately overlaying the corresponding ground truth data.For quantitative evaluation,different measurement approaches are investigated.The empirical discrepancy approach method is defined and found valid for linear geospatial object extraction accuracy evaluation.The measurement factors are computed and the corresponding performance scores are obtained.The results reveal that the multiscale geometric analysis integrated approaches are quite robust and effective for accurate and complete extraction of the linear geospatial objects.Therefore,the application of multiscale geometric analysis approach is an effective geospatial data representation and analysis approach.Further investigations of such approaches for spatial information derivation are important.