Zhang, Caiyun

Reliable geographic intelligence is essential for urban areas; land-cover classification creates the data for urban spatial decision making. This research tested a methodology to create a land-cover map for the main campus of Florida Atlantic University in Boca Raton, Florida. The accuracy of nine separate land-cover classification results were tested; the one with the highest accuracy was chosen for the final map. Object-based image segmentation was applied to fused and LiDAR point cloud (elevation and intensity) data and aerial imagery. These were classified by Random Forest, k-Nearest Neighbor and Support Vector Machines classifiers. Shadow features were reclassified hierarchically in order to create a complete map. The Random Forest classifier used with the fused data set gave the highest overall accuracy at 82.3%, and a Kappa value at 0.77. When combined with the results from the shadow reclassification, the overall accuracy increased to 86.3% and the Kappa value improved to 0.82.

Many minerals, such as calcite and magnetite, show diagnostic overtone and combination bands in the 350-2500 nm window. Sand, though an important unconsolidated material with great abundance on the Earth’s surface, is largely overlooked in spectroscopic studies. Over 100 sand samples were analyzed through traditional microscopic methods and compared to spectral reflectance collected via an ASD Spectroradiometer. Multiple methods were chosen to compare spectroscopic data to sand composition and grain size: 1) existing spectral indices, 2) continuum removal, 3) derivative analysis, and 4) correlation analysis. Particular focus was given to carbonate content. Results from derivative and correlation analysis showed strong correlations in the 2180-2240 nm and 2300-2360 nm windows to carbonate content. Proposed here is the Normalized Difference Carbonate Sand Index (NDCSI), which showed Pearson correlations of r=-0.78 for light-colored samples and r=-0.77 for all samples used. This index is viable for use with carbonate-rich sands.

The canal system of South Florida has become a new distributive focus for the invasive Lionfish (Pterois volitans). Lionfish are considered a pest here, having up to 18 venomous spines that can inflict pain if stepped on or handled. These fish also eat a variety of juvenile species affecting the commercial and recreational fishing industry. The canal system in south Florida is also a center for recreational activities. Water land cover information will aid in species removal by offering species information to areas with a high percent of water land cover and who are more likely to come into contact with Lionfish. This research, comparing classification techniques to map water land cover, is the first step to mitigate the stronghold the lionfish have in South Florida. Once mapped, species information can then be distributed to residents that have close proximity to danger zones.

Quantitative assessment of substrate classification for sand properties is needed for land management and conservation. Establishing a sand spectral library is the first step in this process. Hyperspectal analysis allows for rapid, nondestructive data acquisition. This process uses an ASD spectrometer in a laboratory setting with an artificial light source to collect the spectra. Sand collected worldwide was also analyzed for grain size and composition. Development of spectral libraries of sand is an essential factor to facilitate analytical techniques to monitor coastal problems including erosion and beach nourishment. This in turn can affect various flora and fauna which requires specific substrate to grow, nest, or live. Preliminary results show that each sand sample has a unique signature that can be identified using hyperspectral data.

Thermal remote sensing is a powerful tool for measuring the spatial variability of
evapotranspiration due to the cooling effect of vaporization. The residual method is a
popular technique which calculates evapotranspiration by subtracting sensible heat from
available energy. Estimating sensible heat requires aerodynamic surface temperature
which is difficult to retrieve accurately. Methods such as SEBAL/METRIC correct for
this problem by calibrating the relationship between sensible heat and retrieved surface
temperature. Disadvantage of these calibrations are 1) user must manually identify
extremely dry and wet pixels in image 2) each calibration is only applicable over limited
spatial extent. Producing larger maps is operationally limited due to time required to
manually calibrate multiple spatial extents over multiple days. This dissertation develops
techniques which automatically detect dry and wet pixels. LANDSAT imagery is used
because it resolves dry pixels. Calibrations using 1) only dry pixels and 2) including wet
pixels are developed. Snapshots of retrieved evaporative fraction and actual evapotranspiration are compared to eddy covariance measurements for five study areas in
Florida: 1) Big Cypress 2) Disney Wilderness 3) Everglades 4) near Gainesville, FL. 5)
Kennedy Space Center. The sensitivity of evaporative fraction to temperature, available
energy, roughness length and wind speed is tested. A technique for temporally
interpolating evapotranspiration by fusing LANDSAT and MODIS is developed and
tested.
The automated algorithm is successful at detecting wet and dry pixels (if they
exist). Including wet pixels in calibration and assuming constant atmospheric
conductance significantly improved results for all but Big Cypress and Gainesville.
Evaporative fraction is not very sensitive to instantaneous available energy but it is
sensitive to temperature when wet pixels are included because temperature is required for
estimating wet pixel evapotranspiration. Data fusion techniques only slightly
outperformed linear interpolation. Eddy covariance comparison and temporal
interpolation produced acceptable bias error for most cases suggesting automated
calibration and interpolation could be used to predict monthly or annual ET. Maps
demonstrating spatial patterns of evapotranspiration at field scale were successfully
produced, but only for limited spatial extents. A framework has been established for
producing larger maps by creating a mosaic of smaller individual maps.