Search

Over 50 years ago, the United States Interior Secretary, Stewart Udall, directed space agencies to gather "facts about the natural resources of the earth." Today global climate change and human modification make earth observations from all variety of sensors essential to understand and adapt to environmental change. The Landsat program has been an invaluable source for understanding the history of the land surface, with consistent observations from the Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) sensors since 1982. This dissertation develops and explores methods for enhancing the TM/ETM+ record by fusing other data sources, specifically, Landsat 8 for future continuity, radar data for tropical forest monitoring, and meteorological data for semi-arid vegetation dynamics. Landsat 8 data may be incorporated into existing time series of Landsat 4-7 data for applications like change detection, but vegetation trend analysis requires calibration, especially when using the near-infrared band. The improvements in radiometric quality and cloud masking provided by Landsat 8 data reduce noise compared to previous sensors. Tropical forests are notoriously difficult to monitor with Landsat alone because of clouds. This dissertation developed and compared two approaches for fusing Synthetic Aperture Radar (SAR) data from the Advanced Land Observation Satellite (ALOS-1) with Landsat in Peru, and found that radar data increased accuracy of deforestation. Simulations indicate that the benefit of using radar data increased with higher cloud cover. Time series analysis of vegetation indices from Landsat in semi-arid environments is complicated by the response of vegetation to high variability in timing and amount of precipitation. We found that quantifying dynamics in precipitation and drought index data improved land cover change detection performance compared to more traditional harmonic modeling for grasslands and shrublands in California. This dissertation enhances the value of Landsat data by combining it with other data sources, including other optical sensors, SAR data, and meteorological data. The methods developed here show the potential for data fusion and are especially important in light of recent and upcoming missions, like Sentinel-1, Sentinel-2, and NASA-ISRO Synthetic Aperture Radar (NISAR).

The remote estimation of rainfall rate R is essential for the aviation industry, agriculture, and food warning. Radar, the current means of R estimation, is not available in much of the world. In addition, this measurement involves a level of inaccuracy. Using lightning to detect rain is a relatively inexpensive alternative to radar systems and can be done from existing satellites. Previous research has revealed correlations between lightning and rain, suggesting either that it is possible to estimate R using lightning, or that it is possible to use it to correct for a portion of the radar inaccuracies. These correlations are not only between the amount of lightning and the amount of rain, but also between other parameters, including statistics describing raindrop size. Rain, lightning, and radar data were collected in Central Florida over a two month period in the summer of 2005. Rain data, including raindrop size statistics, were collected from a single point using a disdrometer. Lightning data were collected using the Los Alamos Sferic Array. Radar data were obtained from the WSR-88D radar network. Rain rate R and the raindrop size statistics were compared to lightning statistics to determine which rain/lightning parameter pairs were most correlated. The degree of correlation between rain and lightning parameters was evaluated using the correlation coeffiecint r. Diffrent lightning types (Cloud-to-Ground, Intra-Cloud, Narrow-Bipolar-Event, Total) were considered, and various circular areas were used for lightning collection to optimize the strength of the correlations. Four models using lightning and/or radar for the estimation of R were developed and then compared for accuracy. The rst model is based on the relationship between R and the radar reflectivity factor Z, as is the current practice. Two models using only lightning for the estimation of R were evaluated, and a final model used both radar and lightning data to estimate R. The performance of each model was evaluated using the RMS error. The correlations between rain and lightning parameters were generally weak (r < 0.5), although some pairs clearly produced stronger correlations than others. Results show that the strongest correlations are between lightning density (strokes/km2/hr) and Lambda, a parameter of the raindrop size distribution. This correlation was strongest for Intra-Cloud (IC) lightning measured on a 75 km diameter. Results from the R estimation models indicate that the use of lightning alone is a valid alternative to the use of radar for the conditions studied (R > 0.1 mm/hr, lightning present). The method combining radar and lightning parameters produces more accurate estimations of R than either type alone. Based on these results, lightning data can be used in addition to radar to provide greater accuracy to publicly available rain estimates, and it can be used to provide rain estimation capability to new locations, including greater food warning ability.

A new method is presented for calculating daily rainfall amounts from radar. Radar data from two River Forecast Centers (RFC), and daily rain gauge data from stations around the Northeast U.S. are used to create a radar-level resolution grid of rainfall. The purpose for this method is to produce fields of precipitation estimates in the operational area of the Northeast Regional Climate Center (NRCC), to archive the high-resolution precipitation product, and to use the product as input into a crop modeling program. Considering rain gauge observations as the true values, radar errors are calculated at each rain gauge location every day. Using an interpolation method, the errors are estimated at each radar pixel and added back to the radar grid. Thirty cases were selected from different times of year and different weather types. Three interpolation methods, Inverse Distance Weighting, Multiquadric Interpolation, and Ordinary Kriging, are compared to the Multisensor Precipitation Estimation (MPE), used operationally by River Forecast Centers. Parameters associated with each interpolation method are adjusted daily using cross-validation to produce the best results for each case. By using daily rain gauge data, all three interpolation methods perform similarly and better than MPE, which uses hourly rain gauge data. All methods for estimating precipitation perform best at low values of precipitation and worst at high values of precipitation. Because of the similarity in results between interpolation methods, the simplest method computationally, Inverse Distance Weighting, has been chosen to be used operationally for the Northeast Regional Climate Center.

In this thesis we propose an automated process for the removal of non-precipitation echoes present in weather radar signals and accurate detection of rainfall. The process employs multifractal analysis using directional Gabor wavelets for accurate detection of the rain events. An optoelectronic joint transform correlator is proposed to provide ultra fast processing and wavelet analysis. Computer simulations of the proposed system show that the proposed algorithm is successful in the detecting rainfall accurately in radar images. The accuracy of the algorithms proposed are compared to accurate results that were generated under expert supervision. Results of the proposed system are also compared to results of QC algorithm for the ground validation software (GVS) used by TRMM ground validity Project and a previous QC algorithm. Several statistical measures computed for different reflectivity ranges show that the proposed algorithm gives accuracy as high as 98.95%, which exceed the 97.46% maximum accuracy for the GVS results. Also, the minimum error rate obtained by the proposed algorithm for different dB ranges decreases to 1.09% whereas the GVS results show a minimum error rate of 1.80%. The rain rate accumulation confirms the success of the proposed algorithm in the accurate removal of nonprecipitation echoes and a higher precision in rain accumulation estimates.

Artificial neural networks (ANN) are extremely useful for solving problems without existing algorithmic solutions or with algorithms that are too complex to implement. ANN-based models were developed to study hydrologic processes including the forecasts of watershed runoff, derivation of radar-based rainfall estimation, and development of radar-based flood warning systems. In the first phase of this study, ANN was applied to the study of rainfall-runoff processes on two urbanized watersheds in North Carolina. Rainfall and discharge/gage height data were employed to train and test the ANN-hydrologic models. Stream flows or gage heights can be accurately forecasted, with forecasting periods ranging from 15 minutes to 2 hours, at the watershed outlet and for downstream locations. Based on downstream gauging stations, it was also possible to generate missing historic stream flow data of an upstream station. The second phase of this study involved the comparison of empirical and ANN approaches to radar-based rainfall estimation. The results indicated that although z-R relationships provided fairly accurate rainfall estimates on the average, these relationships tended to overestimate the rainfall amount of low intensity storms and underestimate the rainfall amount of high intensity storms. Two ANN radar-rainfall models have been developed for radar-based rainfall estimation. Weather radar provides detailed, in both time and space, information about precipitation patterns. With proper training and testing, ANN models have exhibited their ability of rainfall estimation from the radar data. In comparison to the empirical z-R relationship, ANN models provide more accurate estimates of rainfall for both low and high intensity storms. The third phase of research included the development of two radar-based stream flow forecast models: a radar-rainfall-runoff (3R) model and a radar-runoff (2R) model. In the 3R-model, radar data was employed to provide rainfall estimates which, in turn, served as input to a hydrologic stream forecast model for stream flow forecasts. The 2R model employs historic radar and stream-flow data as inputs to the ANN model which directly produce predication of stream flow at one-hour lead time. Better performance was obtained from the ANN-based 2R model in terms of accuracy, efficiency and cost.