This thesis presents a case study on the application of deep learning methods for the dense prediction of land cover types in oblique ground-based photography. While deep learning approaches are widely used in land cover classification of remote-sensing data (i.e., aerial and satellite orthoimagery) for change detection analysis, dense classification of oblique landscape imagery used in repeat photography remains undeveloped. A performance evaluation was carried out to test two state-of the-art architectures, U-net and Deeplabv3+, as well as a fully-connected conditional random fields model used to boost segmentation accuracy. The evaluation focuses on the use of a novel threshold-based data augmentation technique, and three multi-loss functions selected to mitigate class imbalance and input noise. The dataset used for this study was sampled from the Mountain Legacy Project (MLP) collection, comprised of high-resolution historic (grayscale) survey photographs of Canada’s Western mountains captured from the 1880s through the 1950s and their corresponding modern (colour) repeat images. Land cover segmentations manually created by MLP researchers were used as ground truth labels. Experimental results showed top overall F1 scores of 0.841 for historic models, and 0.909 for repeat models. Data augmentation showed modest improvements to overall accuracy (+3.0% historic / +1.0% repeat), but much larger gains for under-represented classes. / Graduate

Remote sensing technology is increasingly being used for rapid detection and visualization of changes caused by catastrophic events. This thesis presents a semi-automated feature-based approach to the identification of building conditions especially in affected areas using GIS and remote sensing information. For image analysis, a new ‘Detected Part of Contour’ (DPC) feature is developed for the assessment of building integrity. The DPC calculates a part of the building contour that can be detected in the remotely sensed image. Additional texture features provide information about the area inside the buildings. The effectiveness of the proposed method is proved by high overall classification accuracy for different study cases. The results demonstrate that the ‘map-to-image’ strategy enables extracting valuable information from the remotely sensed image for each individual vector object, thereby being a better choice for change detection within urban areas.

Detecting changes in land cover through time using remotely sensed imagery is a powerful application that has seen increased use as imagery has become more widely available and inexpensive. Before a time series of remotely sensed imagery can be used for change detection, images must first be standardized for effects outside of real surface change. This thesis established a validation protocol to evaluate the effectiveness of an automated technique for normalizing temporally separate but spatially coincident imagery. Using the concept of pseudo-invariant features between master-slave image pairs, spatially coincident dark and bright points are identified from images and a regression equation is calculated to normalize slave images to a master. I used two sets of imagery to test the performance of the standardization process, a spatially coincident, but temporally variable time series, and spatially and temporally variable images. I tested the underlying statistical assumptions of this approach, and performed simple image subtraction to validate the reduction of master-slave differences using invariant locations. In addition I tested the possibility of reducing between-sensor differences by applying simple linear regression to comparable bands of MSS and TM sensors. Image subtraction showed decreases in master-slave differences as a result of the standardization process, and the process behaved appropriately when there should be no difference between master and slave images (adjacent, but temporally identical imagery). I also found that comparable bands between MSS and TM sensors are similar enough that linear regression may not significantly reduce between-sensor differences.

The Mississippi Institute for Forest Inventory (MIFI) is the only cost-effective large-scale forest inventory system in the United States with sufficient precision for producing reliable volume/weight/biomass estimates for small working circle areas (procurement areas). When forest industry is recruited to Mississippi, proposed working circles may overlap existing boundaries of bordering states leaving a gap of inventory information, and a remote sensing-based system for augmenting missing ground inventory data is desirable. The feasibility of obtaining acceptable cubic foot volume estimates from a Landsat-derived volume estimation model (Wilkinson 2011) was assessed by: 1) an initial study to temporally validate Landsat-derived cubic foot volume outside bark to a pulpwood top estimates in comparison with MIFI ground truth inventory plot estimates at two separate time periods, and 2) re-developing a regression model based on remotely sensed imagery in combination with available MIFI plot data. Initial results failed to confirm the relationships shown in past research between radiance values and volume estimation. The complete lack of influence of radiance values in the model led to a re-assessment of volume estimation schemes. Data outlier trimming manipulation was discovered to lead to false relationships with radiance values reported in past research. Two revised volume estimation models using age, average stand height, and trees per-acre and age and height alone as independent variables were found sufficient to explain variation of volume across the image. These results were used to develop a procedure for other remote sensing technologies that could produce data with sufficient precision for volume estimation where inventory data are sparse or non-existent.

The Mississippi and Alabama barrier islands are sensitive landforms that are affected by hurricanes, longshore currents, and available sediment, yet these effects are difficult to quantify with traditional ground-based surveying. In this study, Landsat satellite imagery was used to evaluate changes in barrier island area and centroid position from 1990 and 2005. When hurricanes are infrequent (1999?2003), barrier islands generally increased in total area and showed only moderate repositioning of their centroid locations. However, when hurricanes were frequent (1994?1999 and 2004?2005), barrier islands showed substantial decreases in area and dramatic repositioning of their island centroid locations. This was especially true following Hurricane Katrina (2005). From 1990 to 2005, the general movement of barrier islands was westerly and most islands experienced an overall reduction in area (-18%). The results of this research are similar to findings reported in the literature and illustrate the suitability of using Landsat imagery to study geomorphic changes.

No description available.

Change detection is currently a topic of great interest to theoretic geographic researchers. The necessity to map, monitor, and model land cover change is also important to a variety of applied fields as varied as urban planning and military intelligence. This research compares five algorithms to map urban land cover change in the greater Las Vegas, Nevada metropolitan area. Landsat Thematic Mapper imagery acquired on May 1990 and May 2000 was used as the primary data. The change detection methods yielded simple maps of change vs. no change. These algorithms included image differencing, image ratioing, image regression, vegetation index differencing, and principal components analysis. Each of these techniques accurately identified areas of land cover with moderate levels of accuracy and produced overall change detection accuracy values between 60% and 76% depending on the method. The highest accuracy was obtained by the image ratioing method using the red spectral band (76%). As expected, the determination of change detection thresholds for each technique was critical to the accuracy produced by the algorithm. Moreover, the type of statistic used in optimizing that threshold was also a significant impacting the final accuracy. The approach of using a set of ground points to calibrate the change detection threshold proved to have significant merit.

Deep-Learning (DL) methods have been widely used for Remote Sensing (RS) applications in the last few years, and they allow improving the analysis of the temporal information in bi-temporal and multi-temporal RS images. DL methods use RS data to classify geographical areas or find changes occurring over time. DL methods exploit multi-sensor or multi-temporal data to retrieve results more accurately than single-source or single-date processing. However, the State-of-the-Art DL methods exploit the heterogeneous information provided by these data by focusing the analysis either on the spatial information of multi-sensor multi-resolution images using multi-scale approaches or on the time component of the image time series. Most of the DL RS methods are supervised, so they require a large number of labeled data that is challenging to gather. Nowadays, we have access to many unlabeled RS data, so the creation of long image time series is feasible. However, supervised methods require labeled data that are expensive to gather over image time series. Hence multi-temporal RS methods usually follow unsupervised approaches. In this thesis, we propose DL methodologies that handle these open issues. We propose unsupervised DL methods that exploit multi-resolution deep feature maps derived by a Convolutional Autoencoder (CAE). These DL models automatically learn spatial features from the input during the training phase without any labeled data. We then exploit the high temporal resolution of image time series with the high spatial information of Very-High-Resolution (VHR) images to perform a multi-temporal and multi-scale analysis of the scene. We merge the information provided by the geometrical details of VHR images with the temporal information of the image time series to improve the RS application tasks. We tested the proposed methods to detect changes over bi-temporal RS images acquired by various sensors, such as Landsat-5, Landsat-8, and Sentinel-2, representing burned and deforested areas, and kinds of pasture impurities using VHR orthophotos and Sentinel-2 image time series. The results proved the effectiveness of the proposed methods.

This study looks at the changes to NDVI value in the Weeks Bay Reserve following the impact by Hurricane Katrina. Four Landsat images from March 24, 2005 (Pre-Katrina), September 16, 2005/ April 26, 2006 (Post-Katrina) and August 7, 2002 (Control) were classified into different landcover types and run with the NDVI vegetation index. Those images were compared against each other and showed that the September image had a NDVI value drop of 49% and the April image had a 47% drop as compared to the previous March. The emergent vegetation surrounding the shoreline was most susceptible to changes in NDVI value and recovered the slowest of the tested landcover types. Swift tracks, bay areas, and rivers in the study area where tested and showed that the rivers are the most susceptible change in NDVI value and recovered the slowest.

No description available.