Machine-learning (ML) requires human-labeled “truth” data to train and test. Acquiring and labeling this data can often be the most time-consuming and expensive part of developing trained models of convolutional neural networks (CNN). In this work, we show that an automated workflow using automatically labeled synthetic data can be used to drastically reduce the time and effort required to train a machine learning algorithm for detecting buildings in aerial imagery acquired with low-flying unmanned aerial vehicles. The MSU Autonomous Vehicle Simulator (MAVS) was used in this work, and the process for integrating MAVS into an automated workflow is presented in this work, along with results for building detection with real and simulated images.

Levees are extended embankments naturally or artificially constructed that regulate water levels. The Mississippi Levee system in USA dates from early 1900's, and was designed to keep farmlands safe during the flooding. Today it is considered one of the most extensive engineering infrastructure projects in the world. The Mississippi Levee system comprises over 5,600 km of earthen wall structure designed to protect the agricultural areas of the Mississippi River Valley, thus monitoring the physical condition of the levees is vital for flood control. The dynamics of subsurface water events cause continuous damage on levee structures which are not easily spotted prior to the appearance of slough slides or through-seepage. The traditional method used for detecting these events involves physical surveys, always time consuming and expensive. Synthetic Aperture Radar (SAR) technologies are capable of identifying variations in soil properties that are related to the potential for such events. This paper presents a compendium of methods and results of applying the NASA Unmanned Aerial Vehicle Synthetic Aperture Radar (UAVSAR) to the task of identifying problems on Mississippi's earthen levee system. UAVSAR is a multi-polarized SAR system, and the data exploration in this study included both single-look and multi-look records. The team investigated soil moisture, soil electrical conductivity, sand boils, landslide events, and vegetation backscatter. Different feature detection and classification algorithms were tested, including radiometric and textural methods. The levee structure and flood control system is managed by the United States Army Corps of Engineers (USACE), who would benefit from having a unified tool capable of mapping the weakness of the levee segments. The data and methods presented in this paper are incorporated into the Multi-Tabbed Levee Assessment Tool. The solution provides a user-friendly, open-source geospatial platform to perform visualization, manipulation and analysis of SAR data products.

This article reviews the use of synthetic aperture radar remote sensing data for earthen levee mapping with an emphasis on finding the slump slides on the levees. Earthen levees built on the natural levees parallel to the river channel are designed to protect large areas of populated and cultivated land in the Unites States from flooding. One of the signs of potential impending levee failure is the appearance of slump slides. On-site inspection of levees is expensive and time-consuming; therefore, a need to develop efficient techniques based on remote sensing technologies is mandatory to prevent failures under flood loading. Analysis of multi-polarized radar data is one of the viable tools for detecting the problem areas on the levees. In this study, we develop methods to detect anomalies on the levee, such as slump slides and give levee managers new tools to prioritize their tasks. This paper presents results of applying the National Aeronautics and Space Administration (NASA) Jet Propulsion Lab (JPL)’s Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) quad-polarized L-band data to detect slump slides on earthen levees. The study area encompasses a portion of levees of the lower Mississippi River in the United States. In this paper, we investigate the performance of polarimetric and texture features for efficient levee classification. Texture features derived from the gray level co-occurrence (GLCM) matrix and discrete wavelet transform were computed and analyzed for efficient levee classification. The pixel-based polarimetric decomposition features, such as entropy, anisotropy, and scattering angle were also computed and applied to the support vector machine classifier to characterize the radar imagery and compared the results with texture-based classification. Our experimental results showed that inclusion of textural features derived from the SAR data using the discrete wavelet transform (DWT) features and GLCM features provided higher overall classification accuracies compared to the pixel-based polarimetric features.

There is a critical need to advance the science of levee monitoring to enable the implementation of a cost-effective, multiscale monitoring system that can identify internal weakness and vulnerable points at early stages. In this study, a developed levee monitoring system using polarimetric synthetic aperture radar (PolSAR) data is used. The PolSAR data are used as input in an automated classification system for characterizing areas in the levees facing anomalies with an emphasis on slump slides. In addition to the remotely sensed data, extensive set of in situ soil data are collected to provide detailed information about the soils' properties within the study area. In situ data are collected via geotechnical field exploration to provide information regarding soil texture, moisture content, and penetration resistance. The study area is focused on the mainline levee system of the Mississippi River along the eastern side of the river in the state of Mississippi. While the automated classification system correctly identified the location of actual slide, there were some points, referred as "false positive points," which were classified as anomaly by the classifier but no actual slide was found at those locations. This study focuses on these false positive points by comparing their in situ soil properties with that of true positive points (i.e., actual slide areas) to understand if these false positive points are similar to the slide area, and if they can be vulnerable to future slump slides. The results showed that more than 70% of the false positives have similar soil properties to the actual slide.

Project 1.31 Evaluating Sensitivity of Autonomous Algorithms to Sensor Error and Environmental Conditions While much effort has been devoted to the development of autonomous navigation algorithms in recent years, techniques and procedures for quantifying the performance of these algorithms have been developed on an ad-hoc, case-by-case basis. In the first year of this work, we