SOIL MOISTURE RETRIEVAL USING MICROWAVE REMOTE SENSING DATA

Abstract

Soil moisture state of a river catchment affects many important hydrological and meteorological processes at various scales and plays significant role in partitioning of precipitation during the storm events into runoff and ground water storage. Thus, it is very important to know the spatial distribution of soil moisture. The soil moisture state of a catchment antecedent to a storm event is the most important input required for the Physics-based distributed Hydrological models of rainfall - runoffprocesses. Soil moisture can directly be measured at selected locations by using gravimetric and other conventional methods, which being point measurements, can not be used to estimate soil moisture at spatial scales. However, microwave remote sensing data, particularly that acquired by Synthetic Aperture Radar (SAR) have demonstrated their potential to make frequent and spatially comprehensive measurements of the near-surface soil moisture. Currently, there are several operational satellite systems that carry SAR sensors at frequencies suitable for soil moisture estimation. These include ESA ERS-1 and 2, ESA ENVISAT, Japanese JERS and the Canadian RADARSAT-1/2. The digital number (DN) in SAR image represents the intensity of the microwave signal reflected back from the target and is recorded by the sensor, from which, backscatter coefficient (a0) is estimated. Backscatter coefficient is considered to be an important sensor parameter that exhibits sensitivity towards soil moisture. For a soil, an increase in soil moisture leads to a corresponding increase in its dielectric constant thereby, causing an increase in the backscatter coefficient. However, direct backscatter coefficient (termed as a°D) exhibits a weak relationship with the volumetric soil moisture due to a numberof factors that contribute towards the attenuation of crD. n The main factors that have been analysed to be responsible for causing their effect on <JD are topography, absorption and scattering characteristics of the vegetation cover and the soil surface roughness. The degree of effect of these factors on <JD forms a determinative basis for the correct estimation of <JD. Understanding and analysing the effect of these factors to eliminate their effect on <JD to accurately and effectively estimate the soil moisture was the main focus of this research. Two different approaches namely physically-based modelling approach and the ANNbased modelling approach have been developed and implemented. Catchment of Solani river in and around the town of Roorkee , in the province of Uttaranchal, India has been selected as study area to accomplish the major objectives of this research (i.e., estimation of soil moisture) using the mentioned two approaches. Three ERS-2 SAR PRI (precision image) images acquired on 28th July 2003 (start of autumn season), 29th March 2004 (end of spring season) and 3rd May 2004 (start of summer season) are analysed. Two images from IRS-IB and 1C Linear Imaging and Self Scanning sensors (LISSII and LISS III) ofdates 23rd June 2002 (autumn season) and 12th March 2004 (spring season) have also been utilised to produce land cover classification of the study area so as to identify the class allocation of the sampled pixels in SAR data pertaining to three different seasons. Extensive field data on volumetric soil moisture measured at a large number of locations concurrently to the period of satellite pass, soil surface roughness as well as vegetation characteristics were collected for use in the modelling approaches developed herein. The first approach makes use of the physical characteristics of the study area, therefore it is termed herein as the physically based approach for the estimation of the soil in moisture. This approach considers the effects of physical factors related to the topography of the terrain and land cover on the backscatter coefficient. The influence of the physical factors viz. topography and vegetation on crD was accounted for through the use of mathematical formulations available in the literature. The effect of topography has been examined by the use of model proposed in Robinson (1966). The semi-empirical water cloud model has been used to conduct a comparative study between three important parameters that describes the vegetation in terms of their bulk characteristics (e.g., leaf area index; LAI, plant water content; PWC and crop height '/i') to identify a vegetation descriptor that had themaximum influence on backscatter coefficient. Anempirical model namely Dubois model that incorporates soil dielectric constant and root mean square (rms) surface roughness (s) has been used to estimate the backscatter coefficient required as an input to least square method (LSM) for estimating the volumetric soil moisture. The rms surface roughness heights were measured in the field by using a mechanical device named as "Surface Roughness Profiler" that was indigenously designed and built inthe laboratory at Roorkee during the present study. The relationship ofERS-2 SAR backscatter coefficient with the concurrent soil moisture was studied. A weak relationship was observed between the two for all the three seasons, which was mainly attributed to the influences of terrain related factors stated above. An insignificant improvement in the relationship between <JD and the concurrent soil moisture was observed when the topographic effects on the backscatter coefficient were accounted for through the use of incidence angle based model. This result is understandable as the land slope within the study area is small (i.e. less than 4°). The effect of vegetation on <T°D was taken into account by utilising the semiempirical water cloud model. Asignificant improvement in the sensitivity of backscatter coefficient towards soil moisture was observed when LAI was used as a vegetation iv descriptor in the model. The volumetric soil moisture showed a strong correlation with the <JD corrected for the influences due to vegetation and topography. The relation between these has also been used for generating the soil moisture maps of the study area depicting its spatial variation corresponding to the three seasons. The validation data set when superimposed on these maps indicated a maximum error of ± 10% in the estimation of the soil moisture for all the land cover classes of the three seasons. The second approach involves the use of a non-parametric ANN technique based on feed forward back propagation neural network (BPNN) for establishing relationship between volumetric soil moisture and terrain as well as sensor related parameters (i.e., cr0D or DN). Anumber ofANN architectures have been designed so as to obtain the optimal architecture by employing a few parameters defining the characteristics of surface roughness and vegetation. The results from ANNapproach have been compared with those derived from physically-based modelling approach as well as from the conventional multiple regression approach. Finally, soil moisture maps of the study area have been generated based on the results obtained from this approach. The best estimation of soil moisture through the ANN based modelling was achieved when seven variables (i.e.,a°D, land cover, sensor incidence angle; a,, s, LAI, PWC and h) representing the scene and sensor parameters were supplied to the network as the inputs. It was also observed that the neural networks trained even with five input variables namely a° ,land cover, a,, sand LAI also produced soil moisture estimates with higher values of the coefficient of determination R2 .The ANNs could also be trained with sufficient accuracy by only using four input variables viz. a°D or DN, incidence angle, land cover and LAI so as to generate soil moisture maps based on ANN approach, since the pixel wise data for the remaining input variable namely rms surface roughness height (s) were not available. The validation data set when supplied to the trained ANNs showed a close agreement between neural network estimated soil moisture and the in-situ observed soil moisture. However, combining data sets of all the three seasons in training the ANN reduced the accuracies of soil moisture estimation for all neural networks due to a large variation in the weather conditions of the seasons in which data were acquired. The results from ANNapproach were found to be significantly more accurate than those obtained from the traditional multiple regression approach attempted only for the purpose of comparative analysis. Thus, it confirms the hypothesis that the ANN can better simulate the non-linear relationships between the inputs and the output variables. Also, the results from both physically-based as well as ANN based approach were found to be in close agreement with each other (correlation between in-situ observed soil moisture with physically based model and ANNestimated soil moisture showed R2 ~ 0.95 to 0.97) for all land cover classes of three seasons. VI