Jason A. Tullis

Object-based change detection using correlation image analysis and image segmentation

This study introduces change detection based on object/neighbourhood correlation image analysis and image segmentation techniques. The correlation image analysis is based on the fact that pairs of brightness values from the same geographic area (e.g. an object) between bi‐temporal image datasets tend to be highly correlated when little change occurres, and uncorrelated when change occurs. Five different change detection methods were investigated to determine how new contextual features could improve change classification results, and if an object‐based approach could improve change classification when compared with per‐pixel analysis. The five methods examined include (1) object‐based change classification incorporating object correlation images (OCIs), (2) object‐based change classification incorporating neighbourhood correlation images (NCIs), (3) object‐based change classification without contextual features, (4) per‐pixel change classification incorporating NCIs, and (5) traditional per‐pixel change classification using only bi‐temporal image data. Two different classification algorithms (i.e. a machine‐learning decision tree and nearest‐neighbour) were also investigated. Comparison between the OCI and the NCI variables was evaluated. Object‐based change classifications incorporating the OCIs or the NCIs produced more accurate change detection classes (Kappa approximated 90%) than other change detection results (Kappa ranged from 80 to 85%).

Synergistic use of lidar and color aerial photography for mapping urban parcel imperviousness

The imperviousness of land parcels was mapped and evaluated using high spatial resolution digitized color orthophotography and surface-cover height extracted from multiple-return lidar data. Maximum-likelihood classification, spectral clustering, and expert system approaches were used to extract the impervious information from the datasets. Classified pixels (or segments) were aggregated to parcels. The classification model based on the use of both the orthophotography and lidar-derived surface-cover height yielded impervious surface results for all parcels that were within 15 percent of reference data. The standard error for the rule-based per-pixel model was 7.15 percent with a maximum observed error of 18.94 percent. The maximum-likelihood per-pixel classification yielded a lower standard error of 6.62 percent with a maximum of 14.16 percent. The regression slope (i.e., 0.955) for the maximum-likelihood per-pixel model indicated a near perfect relationship between observed and predicted imperviousness. The additional effort of using a per-segment approach with a rule-based classification resulted in slightly better standard error (5.85 percent) and a near-perfect regression slope (1.016).

An Evaluation of Lidar-derived Elevation and Terrain Slope in Leaf-off Conditions

The effects of land cover and surface slope on lidar-derived elevation data were examined for a watershed in the piedmont of North Carolina. Lidar data were collected over the study area in a winter (leaf-off) overflight. Survey-grade elevation points (1,225) for six different land cover classes were used as reference points. Root mean squared error (RMSE) for land cover classes ranged from 14.5 cm to 36.1 cm. Land cover with taller canopy vegetation exhibited the largest errors. The largest mean error (36.1 cm RMSE) was in the scrub-shrub cover class. Over the small slope range (0° to 10°) in this study area, there was little evidence for an increase in elevation error with increased slopes. However, for low grass land cover, elevation errors do increase in a consistent manner with increasing slope. Slope errors increased with increasing surface slope, under-predicting true slope on surface slopes � 2°. On average, the lidarderived elevation under-predicted true elevation regardless of land cover category. The under-prediction was significant, and ranged up to � 23.6 cm under pine land cover.

Impact of Lidar Nominal Post-spacing on DEM Accuracy and Flood Zone Delineation

Lidar data have become a major source of digital terrain information for use in many applications including hydraulic modeling and flood plane mapping. Based on established relationships between sampling intensity and error, nominal post-spacing likely contributes significantly to the error budget. Post-spacing is also a major cost factor during lidar data collection. This research presents methods for establishing a relationship between nominal post-spacing and its effects on hydraulic modeling for flood zone delineation. Lidar data collected at a low post-spacing (approximately 1 to 2 m) over a piedmont study area in North Carolina was systematically decimated to simulate datasets with sequentially higher post-spacing values. Using extensive first-order ground survey information, the accuracy of each DEM derived from these lidar datasets was assessed and reported. Hydraulic analyses were performed utilizing standard engineering practices and modeling software (HEC-RAS). All input variables were held constant in each model run except for the topographic information from the decimated lidar datasets. The results were compared to a hydraulic analysis performed on the un-decimated reference dataset. The sensitivity of the primary model outputs to the variation in nominal post-spacing is reported. The results indicate that base flood elevation does not statistically change over the post-spacing values tested. Conversely, flood zone boundary mapping was found to be sensitive to variations in post-spacing.

Per-segment Aboveground Forest Biomass Estimation Using LIDAR-Derived Height Percentile Statistics

An outbreak of red oak borer, an insect infesting red oak trees, prompted the need for a biomass model of closed-canopy oak-hickory forests in the rugged terrain of the Arkansas Ozarks. Multiple height percentiles were calculated from small-footprint aerial LIDAR data, and image segmentation was employed to partition the LIDAR-derived surface into structurally homogeneous modeling units. In situ reference data were incorporated into a machine-learning algorithm that produced a regression-tree model for predicting aboveground woody biomass per segment. Model results on training data appear adequate for prediction purposes (mean error 2.38 kg/m2, R 2 = 0.83). Model performance on withheld test data reveals slightly lower accuracy (2.77 kg/m2, R 2 = 0.72).

A Remote Sensing and GIS-assisted Spatial Decision Support System for Hazardous Waste Site Monitoring

Humans produce large amounts of waste that must be processed or stored so that it does not contaminate the environment. When hazardous wastes are stored, waste site monitoring is typically conducted in situ which can lead to a serious time lag between the onset of a problem and detection. A Remote Sensing and GIS-assisted Spatial Decision Support System for Hazardous Waste Site Monitoring was developed to improve hazardous waste site management. The system was designed to be recursive, flexible, and integrative. It is recursive because the system is implemented iteratively until the risk assessment subsystem determines that an event is no longer a problem to the surrounding human population or to the environment. It is flexible in that it can be adapted to monitor a variety of hazardous waste sites. The system is integrative because it incorporates a number of different data types and sources (e.g., multispectral and lidar remote sensor data, numerous type of thematic information, and production rules), modules, and human expert knowledge of the hazardous waste sites. The system was developed for monitoring hazardous wastes on the Savannah River National Laboratory near Aiken, South Carolina.

Spatial scale management experiments using optical aerial imagery and LIDAR data synergy.

Computational trends toward shared services suggest the need to automatically manage spatial scale for overlapping applications. In three experiments using high-spatial-resolution optical imagery and LIDAR data to extract impervious, forest, and herbaceous classes, this study optimized C5.0 rule sets according to: (1) spatial scale within an image tile; (2) spatial scale within spectral clusters; and (3) stability of predicted accuracies based on cross validation. Alteration of the image segmentation scale parameter affected accuracy as did synergy with LIDAR derivatives. Within the tile examined, forest and herbaceous areas benefited more from optical and LIDAR synergy than did impervious surfaces.

Development of a remote sensing change detection system based on neighborhood correlation image analysis and intelligent knowledge-based systems

A remote sensing change detection system utilizing Neighborhood Correlation Images (NCIs) and intelligent knowledge-based systems is introduced in this study. NCIs consist of three variables, correlation, slope, and intercept from correlation analysis between two date images with a specified neighborhood. In certain areas, correlation between two image dates can be closely related to the level of change. Changed areas tend to have very low correlation, while no-change areas usually have high correlation. Slope and intercept values can also provide information related to change. Ideally, a pixel from an unchanged area will have a slope of ~ 1 and an intercept near 0. Unfortunately, such correlation analysis may not provide detailed “from-to” change information. However, if combined with image classification (e.g. using an intelligent knowledgebased decision tree), the magnitude of the three variables can be used to produce detailed “from-to” change information. The amount of change can also be extracted when they are combined with knowledge-based regressions or neural networks. This study summarizes the design and development of a remote sensing change detection system based on neighborhood correlation image analysis and the C5.0/Cubist systems within the ArcGIS 9.0 VBA environment.

Deciduous Forest Structure Estimated with LIDAR-Optimized Spectral Remote Sensing

Coverage and frequency of remotely sensed forest structural information would benefit from single orbital platforms designed to collect sufficient data. We evaluated forest structural information content using single-date Hyperion hyperspectral imagery collected over full-canopy oak-hickory forests in the Ozark National Forest, Arkansas, USA. Hyperion spectral derivatives were used to develop machine learning regression tree rule sets for predicting forest neighborhood percentile heights generated from near-coincident Leica Geosystems ALS50 small footprint light detection and ranging (LIDAR). The most successful spectral predictors of LIDAR-derived forest structure were also tested with basal area measured in situ. Based on the machine learning regression trees developed, Hyperion spectral derivatives were utilized to predict LIDAR forest neighborhood percentile heights with accuracies between 2.1 and 3.7 m RMSE. Understory predictions consistently resulted in the highest accuracy of 2.1 m RMSE. In contrast, hyperspectral prediction of basal area measured in situ was only found to be 6.5 m2/ha RMSE when the average basal area across the study area was ~12 m2/ha. The results suggest, at a spatial resolution of 30 × 30 m, that orbital hyperspectral imagery alone can provide useful structural information related to vegetation height. Rapidly calibrated biophysical remote sensing techniques will facilitate timely assessment of regional forest conditions.

Empirical versus Model‐based Atmospheric Correction of Digital Airborne Imaging Spectrometer Hyperspectral Data

Abstract This paper reports the results of a quantitative comparison of empirical and model based atmospheric correction techniques for the radiometric calibration of a Digital Airborne Imaging Spectrometer (DAIS) 3715 hyperspectral image dataset. Empirical line calibration (ELC) and the radiative transfer based model Atmospheric CORrection Now (ACORN) were applied to transform the hyperspectral dataset from values of radiance to scaled percent reflectance. An additional spectral polishing technique called single spectrum enhancement (SSE) was implemented a posteriori to refine the transformation results. To evaluate the accuracy of the radiometric calibration techniques, spectra extracted from the processed images were analytically compared to spectral measurements collected in situ with a handheld spectroradiometer at 46 sample point locations. Average RMSE errors were as follows: ELC without SSE = 0.1415, ACORN without SSE = 0.0645, ELC with SSE = 0.0345, and ACORN with SSE = 0.0314. Based on the results of this analysis, spectral polishing through the use of SSE appears to introduce the greatest improvement in the removal of deleterious atmospheric effects when compared to in situ data, regardless of the choice of the model (i.e. ACORN or ELC).