The feasibility of synthesizing distributed fields of remotely-sensed soil moisture by the novel application of four-dimensional data assimilation applied in a hydrological model was explored in this study. Six Push Broom Microwave Radiometer images gathered over Walnut Gulch, Arizona were assimilated into the TOPLATS hydrological model. Several alternative assimilation procedures were implemented, including a method that adjusted the statistics of the modeled field to match those in the remotely sensed image, and the more sophisticated, traditional methods of statistical interpolation and Newtonian nudging. The high observation density characteristic of remotely-sensed imagery poses a massive computational burden when used with statistical interpolation, necessitating observation reduction through subsampling or averaging. For Newtonian nudging, the high observation density compromises the conventional weighting assumptions, requiring modified weighting procedures. Remotely-sensed soil moisture images were found to contain horizontal correlations that change with time and have length scales of several tens of kilometers, presumably because they are dependent on antecedent precipitation patterns. Such correlation therefore has a horizontal length scale beyond the remotely sensed region that approaches or exceeds the catchment scale. This suggests that remotely-sensed information can be advected beyond the image area and across the whole catchment. The remotely-sensed data was available for a short period providing limited opportunity to investigate the effectiveness of surface-subsurface coupling provided by alternative assimilation procedures. Surface observations were advected into the subsurface using incomplete knowledge of the surface-subsurface correlation measured at only 2 sites. It is perceived that improved vertical correlation specification will be a need for optimal soil moisture assimilation. Based on direct measurement comparisons and the plausibility of synthetic soil moisture patterns, Newtonian nudging assimilation procedures were preferred because they preserved the observed patterns within the sampled region, while also calculating plausible patterns in unmeasured regions. Statistical interpolation reduced to the trivial limit of direct data insertion in the sampled region and gave less plausible patterns outside this region. Matching the statistics of the modeled fields to those observed provided plausible patterns, but the observed patterns within sampled area were largely lost.

Remote sensing is a tool that is increasingly used in agriculture for crop management purposes. A ground-based remote sensing data acquisition system was designed, constructed, and implemented to collect high spatial and temporal resolution data in irrigated agriculture. The system was composed of a rail that mounts on a linear move irrigation machine, and a small cart that runs back and forth on the rail. The cart was equipped with a sensors package that measured reflectance in four discrete wavelengths (550 nm, 660 nm, 720 nm, and 810 nm, all 10 nm bandwidth) and an infrared thermometer. A global positioning system and triggers on the rail indicated cart position. The data was postprocessed in order to generate vegetation maps, N and water status maps and other indices relevant for site-specific crop management. A geographic information system (GIS) was used to generate images of the field on any desired day. The system was named AgIIS (A̲gricultural I̲rrigation I̲maging S̲ystem). This ground based remote sensing acquisition system was developed at the Agricultural and Biosystems Engineering Department at the University of Arizona in conjunction with the U.S. Water Conservation Laboratory in Phoenix, as part of a cooperative study primarily funded by the Idaho National Environmental and Engineering Laboratory. A second phase of the study utilized data acquired with AgIIS during the 1999 cotton growing season to model petiole nitrate (PNO₃⁻) and total leaf N. A latin square experimental design with optimal and low water and optimal and low N was used to evaluate N status under water and no water stress conditions. Multivariable models were generated with neural networks (NN) and multilinear regression (MLR). Single variable models were generated from chlorophyll meter readings (SPAD) and from the Canopy Chlorophyll Content Index (CCCI). All models were evaluated against observed PNO₃⁻ and total leaf N levels. The NN models showed the highest correlation with PNO₃⁻ and total leaf N. AgIIS was a reliable and efficient data acquisition system for research and also showed potential for use in commercial farming systems.

The Remote Sensing Group of the Optical Sciences Center at the University of Arizona performs absolute radiometric calibration of Earth-viewing sensors using vicarious methods. The reflectance and irradiance-based methods require the nadir-view reflectance of a calibration site at sensor overpass. Errors in these reflectance data contribute directly to errors in the retrieved at sensor radiance, and therefore errors in the calibration. This research addresses two areas of improvement for the reflectance retrieval. The discreet laboratory data of the reference panel is spectrally interpolated using the measured hemispherical reflectance rather than a polynomial fit. This interpolation better fits an absorption feature of the reference material near 2200 nm. The desired reflectance is due to the directly-transmitted solar irradiance, but field measurements also include irradiance due to diffuse light. Non-lambertian properties of the reference and surface cause the ratio of the reflected total radiances to differ from the ratio of the reflected solar radiances. This difference can be corrected using additional field measurements, shaded surface/shaded-reference, output from a radiative transfer code, RTC-only, or a combination of both, shaded-reference. For the shaded-reference and RTC-only methods the shape of the bi-directional reflectance factor of the surface must be known to better than 10% to maintain a 2% accuracy for the retrievals, while the shaded-surface/shaded-reference method does not use the surface BRF. All three methods were applied to measurements of calibrated reflectance tarpaulins, and to measurements made at White Sands Missile Range. These data demonstrate that the shaded-surface/shaded-reference and RTC-only methods improve the surface reflectance retrieval, while the shaded-reference method is too sensitive to variations between the actual and modeled diffuse sky irradiance to be useful. This research represents significant improvements in the calculation of surface reflectance for vicarious calibration. The hemispherical reflectance interpolation will reduce uncertainties in the short wave infrared by 1%, and the diffuse corrections will reduce the errors in blue by 2% in some cases.

The in-flight absolute radiometric calibration of the Thematic Mapper (TM) is being conducted using the results of field measurements at White Sands, New Mexico. These measurements are made to characterize the ground and atmosphere at the time the TM is acquiring an image of White Sands. The data are used as input to a radiative transfer code that computes the radiance at the entrance pupil of the TM. The calibration is obtained by comparing the digital counts associated with the TM image of the measured ground site with the radiative transfer code result. The calibrations discussed here are for the first four visible and near-infrared bands of the TM. In this dissertation the data reduction for the first calibration attempts on January 3, 1983, and July 8, 1984, is discussed. Included are a review of radiative transfer theory and a discussion of model atmospheric parameters as defined for the White Sands area. These model parameters are used to assess the errors associated with the calibration procedure. Each input parameter to the radiative transfer code is varied from its model value in proportion to the uncertainty with which it can be determined. The effects of these uncertainties on the predicted radiances are determined. It is thought that the optical depth components τ(Ray), τ(Mie), τ(oz), and τ(H₂O) can be measured to within 10%, 2%, 10%, and 30%, respectively. For the white gypsum sand, surface reflectance uniformity is on the order of 1.5%, and the overall uncertainty in measured reflectance is about 2%. This is due to an uncertainty in the reflectance factor of the calibration plates. The greatest uncertainty in calibration is attributed to our uncertainty in the aerosol parameters, in particular the imaginary component of refractive index. The cumulative effect of these uncertainties is thought to produce an uncertainty in computed radiance of about 5%.

M. Tech. Electronic Engineering. / Explores and compare geospatial techniques to improve the detection and classification of settlement types in QuickBird and SPOT 5 satellite images for the purpose of better environmental assessment and monitoring.

published_or_final_version / Geography and Geology / Master / Master of Philosophy

This dissertation examines the improvement in predicting weather and hydrology in the southwestern United States by assimilating satellite-derived precipitation estimates into a numerical mesoscale model. For this investigation the Regional Atmospheric Model System (RAMS) was used and Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) were assimilated into the RAMS' own land surface scheme; Land Ecosystem Atmosphere Feedback model version 2 (LEAF-2). The simulations were conducted for periods of 36 hours--12 hours of initialization and 24 hours of prediction (from July 8th 0000 UTC to 9th 1200 UTC 1999). The control run underpredicted precipitation over southwestern Arizona and showed an excessive precipitation pattern over northeastern Arizona. This precipitation bias was also responsible for biases in surface fluxes such as soil moisture and evapotranspiration. With a realistic surface shortwave radiation adjustment and the improvement of atmospheric state variables within the central model domains during the assimilation period, there was a slight enhancement for near surface temperature and moisture. However, RAMS still performed poorly and improved only marginally for precipitation prediction. The impact of the assimilation of PERSIANN precipitation estimates on soil moisture was significant however, and this accordingly improved the 2m-high temperature and relative humidity. The general pattern of precipitation showed improvement but was still inaccurate the location and intensity of precipitation. To investigate the soil moisture-precipitation feedback mechanism, RAMS simulations were performed with varying initial soil moisture saturation rates starting from a completely dry condition of 0% to a fully saturated condition of 100%. Analysis showed that with less than 20% of initial soil moisture saturation, more than 70% of the water that precipitated into the analysis domain was due to the indirect effect of soil moisture. This explains in part why initial soil moisture improvements for the southwestern United States during the summer had a limited impact on the prediction of precipitation. Finally, model simulations were performed and analyzed to demonstrate the sensitivity of vegetation parameters in RAMS on land surface and near-surface atmospheric variables in the southwestern United States.

Accurate estimates of soil moisture are necessary to predict evapotranspiration, runoff, infiltration, and groundwater recharge and, through these variables, weather, climate, and fire and flood frequencies. This dissertation is motivated by the need to estimate soil water content from remotely sensed passive microwave emission. Two different approaches are taken: (1) improved modeling of the microwave emission from the land surface; and (2) measurements of the average dielectric properties of the soil media and vegetation canopies. Consequently, the first part of the dissertation describes how a stratified dielectric model of the microwave emission from the soil was extended to take into account the effects of vegetation. The model parameters were calibrated using observed data and a robust optimization routine. However, the availability of measurements of some of these parameters, particularly the profile of dielectric permittivity of the canopy, would facilitate a more complete evaluation of the accuracy of the extended microwave emission model. The second part of this dissertation describes progress towards the development of a technique to measure the dielectric of vegetation canopies. This technique is based on gated time domain transmission measurements through the substance of interest. Preliminary studies carried out using soils with varying salinity and water content and vegetation show (1) an unexpected response of the signal to saline soils, and (2) a possible dielectric signature of the onset of stress in plant stems.

Precipitation estimation from satellite information (VISIBLE , IR, or microwave) is becoming increasingly imperative because of its high spatial/temporal resolution and board coverage unparalleled by ground-based data. After decades' efforts of rainfall estimation using IR imagery as basis, it has been explored and concluded that the limitations/uncertainty of the existing techniques are: (1) pixel-based local-scale feature extraction; (2) IR temperature threshold to define rain/no-rain clouds; (3) indirect relationship between rain rate and cloud-top temperature; (4) lumped techniques to model high variability of cloud-precipitation processes; (5) coarse scales of rainfall products. As continuing studies, a new version of Precipitation Estimation from Remotely Sensed Information using Artificial Neural Network (PERSIANN), called Cloud Classification System (CCS), has been developed to cope with these limitations in this dissertation. CCS includes three consecutive components: (1) a hybrid segmentation algorithm, namely Hierarchically Topographical Thresholding and Stepwise Seeded Region Growing (HTH-SSRG), to segment satellite IR images into separated cloud patches; (2) a 3D feature extraction procedure to retrieve both pixel-based local-scale and patch-based large-scale features of cloud patch at various heights; (3) an ANN model, Self-Organizing Nonlinear Output (SONO) network, to classify cloud patches into similarity-based clusters, using Self-Organizing Feature Map (SOFM), and then calibrate hundreds of multi-parameter nonlinear functions to identify the relationship between every cloud types and their underneath precipitation characteristics using Probability Matching Method and Multi-Start Downhill Simplex optimization techniques. The model was calibrated over the Southwest of United States (100°--130°W and 25°--45°N) first and then adaptively adjusted to the study region of North America Monsoon Experiment (65°--135°W and 10°--50°N) using observations from Geostationary Operational Environmental Satellite (GOES) IR imagery, Next Generation Radar (NEXRAD) rainfall network, and Tropical Rainfall Measurement Mission (TRMM) microwave rain rate estimates. CCS functions as a distributed model that first identifies cloud patches and then dispatches different but the best matching cloud-precipitation function for each cloud patch to estimate instantaneous rain rate at high spatial resolution (4km) and full temporal resolution of GOES IR images (every 30-minute). Evaluated over a range of spatial and temporal scales, the performance of CCS compared favorably with GOES Precipitation Index (GPI), Universal Adjusted GPI (UAGPI), PERSIANN, and Auto-Estimator (AE) algorithms, consistently. Particularly, the large number of nonlinear functions and optimum IR-rain rate thresholds of CCS model are highly variable, reflecting the complexity of dominant cloud-precipitation processes from cloud patch to cloud patch over various regions. As a result, CCS can more successfully capture variability in rain rate at small scales than existing algorithms and potentially provides rainfall product from GOES IR-NEXARD-TRMM TMI (SSM/I) at 0.12° x 0.12° and 3-hour resolution with relative low standard error (∼=3.0mm/hr) and high correlation coefficient (∼=0.65).

This dissertation deals with modeling the radiative regime in vegetation canopies and the possible remote sensing applications derived by solving the forward and inverse canopy transport equation. The aim of the research is to develop a methodology (called "end-to-end problem solution") that, starting from first principles describing the interaction between light and vegetation, constructs, as the final product, a tool that analyzes remote sensing data for precision agriculture (ripeness prediction). The procedure begins by defining the equations that describe the transport of photons inside the leaf and within the canopy. The resulting integro-differential equations are numerically integrated by adapting the conventional discrete-ordinate methods to compute the reflectance at the top of the canopy. The canopy transport equation is also analyzed to explore its spectral properties. The goal here is to apply Case's method to determine eigenvalues and eigenfunctions and to prove completeness. A model inversion is attempted by using neural network algorithms. Using input-outputs generated by running the forward model, a neural network is trained to learn the inverse map. The model-based neural network represents the end product of the overall procedure. During Oct 2002, an Unmanned Aerial Vehicles (UAVs) equipped with a camera system, flew over Kauai to take images of coffee field plantations. Our goal is to predict the amount of ripe coffee cherries for optimal harvesting. The Leaf-Canopy model was modified to include cherries as absorbing and scattering elements and two classes of neural networks were trained on the model to learn the relationship between reflectance and percentage of ripe, over-ripe and under-ripe cherries. The neural networks are interfaced with images coming from Kauai to predict ripeness percentage. Both ground and airborne images are considered. The latter were taken from the on-board Helios UAV camera system flying over the Kauai coffee field. The results are compared against hand counts and parchment data to evaluate the network performances on real applications. In ground images, the error is always less than 11%. In airborne image, the error bound is 20%. The results are certainly adequate and show the tremendous potential of the methodology.