SPATIAL MAPPING OF SOIL MOISTURE USING RADARSAT-1 DATA INTRODUCTION

Transcription

1 SPATIAL MAPPING OF SOIL MOISTURE USING RADARSAT-1 DATA Tarendra Lakhankar, PhD Student Hosni Ghedira, Asst. Professor Reza Khanbilvardi, Professor NOAA-CREST, City University of New York New York ABSTRACT In this research, a back-propagation neural network was used to retrieve and map the surface soil moisture in Oklahoma (97d35'W, 36d15'N) from Synthetic Aperture Radar data acquired by RADARSAT-1 satellite. In addition to SAR backscattering, different vegetation-related information (vegetation optical depth and Normalized Difference Vegetation Index) have been added as additional inputs to the neural network algorithm. The soil moisture data measured by Electronically Scanned Thinned Array Radiometer during the SGP97 campaign were used as truth data in the training and the validation processes. All the training and validation pixels have been labeled as either homogeneous or heterogeneous based on land cover type and number of sub-pixels of 25m resolution. The results showed that homogeneous pixels are more likely to have better accuracy than heterogeneous pixels in soil moisture classification. A better correlation between soil moisture and SAR backscattering was found in areas with high soil moisture content. The modeling results have shown that the retrieval of soil moisture in highly vegetated areas was less accurate than bare soil areas. Further, the same results have shown that the additions of vegetation optical depth and Normalized Difference Vegetation Index as additional input to the SAR data had a significant effect on the overall classification accuracy. INTRODUCTION The spatial distribution of soil moisture is a key parameter for many applications such as irrigation management and vegetation growth monitoring, flash flood forecasting, hydrology and stream flow prediction. The soil moisture mapping with the actual field measurement techniques is expensive; time consuming; and very difficult to have an accurate estimate, as it varies in space and in time and its value is generally affected by the variation in soil surface characteristics such as land cover, vegetation density, soil texture etc. Different hydrological study missions such as: FIFE 87-89, Mansoon 90, OXSOME 90, MACHYDRO 90, HAPEX 90-92, WASHITA 92, SGP 97, SGP 99, SMEX 02, SMEX03, and SMEX 04 have been carried out to explore the potential of microwave remote sensing for soil moisture mapping and for the retrieval of other hydrological parameters. Active and passive remote sensing systems and especially those operating in the microwave region of the electromagnetic spectrum have been widely used to retrieve soil moisture from space. The brightness temperature from the passive sensor and backscatter coefficient from the active sensor are strongly related to volumetric soil moisture. The passive sensors measure the natural thermal emission in the form of brightness temperature from the land surface. However, the active microwave systems generate their own radiation, which is transmitted toward the earth surface and measures the reflected energy called backscatter coefficient. However, the most notable difference between active and passive microwave remote sensing systems is the spatial resolution. Active sensors have the capability to provide high spatial resolution, but are more sensitive to the parameters such as surface roughness, topography and vegetation than passive systems. In the other hand, the passive microwave systems can provide low spatial resolutions but with a higher temporal resolution. The active microwave sensors onboard ERS-1, ERS-2, JERS-1, RADARSAT-1, and SIR- C/X-SAR satellite, produce fine resolution data and have high capabilities in estimating the spatial variation of soil moisture in hydrological studies. The accuracy of satellite-derived soil moisture is usually affected by the land cover characteristics such as vegetation, which significantly modifies and attenuates the outgoing microwave radiation of the soil and make it difficult and inaccurate the retrieval of realistic soil moisture from satellite-based sensors. Soil moisture estimation by active remote sensing involves the measurement of backscattering from the soil, which may be affected by vegetation

2 canopy and soil moisture. The vegetation canopy may affect the backscattered energy by contributing to the volume backscatter of the observed scene and by attenuating the soil component of the total backscatter (Ulaby et al. 1981; Kasischke et al. 2003). The total amount of attenuation and backscatter depends on several vegetation parameters, such as vegetation optical depth, vegetation height, leaf area index, and vegetation water content. Indeed, the presence of high and dense vegetation decreases the correlation between the backscattering and the soil moisture. Indeed, under the same land cover conditions, the stronger radar backscattering values are observed when soil moisture is high. However, soil moisture estimation based on SAR data only may face several challenges since the microwave sensors are sensitive to other land cover characteristic such as vegetation density, surface roughness, and soil texture (Ulaby et al. 1981, 1986; Engman and Chauhan, 1995). This paper deals with the application of neural network to estimate the spatial soil moisture by using active microwave data. An attempt has been made to: 1) Improve the overall accuracy by using additional information such as soil texture, land cover use, vegetation optical depth and NDVI. 2) Evaluate how the vegetation cover (type, density, spatial distribution) may affect the soil moisture dynamics under different land cover conditions. STUDY AREA The study selected for this research is located in Oklahoma, USA (97d35 W, 36d15 N). This area has been selected based on the data availability from Southern Great plain mission conducted by NASA in 1997 (SGP97). This study area is a spotlight for large number of remote sensing and soil moisture experiments carried out by various government agencies during the last 15 years. The study area has a sub-humid climate with annual average rainfall of 75 cm. The topography of study area is moderately rolling with maximum relief of 200 m. The predominant vegetation covers are Pasture/Rangeland (approx 49%), wheat (35%), summer legume (5%), Alfalfa (3%), and Forage (4%). The SGP97 experiment was a large, interdisciplinary experiment performed over one-month period (18 June 17 July) with the objective to test formerly established soil-moisture retrieval algorithms for the ESTAR Instrument (Electronically Scanned Thinned Array Radiometer) L-band passive microwave radiometer at 800-meter resolution (Jackson et al. 1999). The ESTAR instruments were operated at a frequency of GHz with horizontal polarization N N A Soil Moisture Data 165 km x 495 km (Res. 800 m) Study Area (A and B) A: 26.4 km x 96 km B: 31.2 km x km N B SAR Image 350 km x 300 km (Res. 25 m) N W W W W W W Figure 1. Details of study area location with reference to SAR images.

3 PCI Geomatica software has been used to convert RADARSAT raw data into calibrated radar backscatter (σ ) in db (Decibel) using CDSAR, SARINCD and SARSIGM commands. Two subset scenes covered by all available data (Radarsat data, soil moisture, vegetation, land-use data, soil texture) have been selected for algorithm development, calibration and validation. As shown in figure 1, the study area was divided in two parts A and B. The study area A covers 26.4 km x 96 km ( km2) and B covers 31.2 km x km ( km2). The soil moisture and other vegetation data such as NDVI and vegetation optical depth, collected during SGP97 experiment, have been compiled and used in different steps of this research. PARAMETERS USED FOR SOIL MOISTURE MAPPING Active Microwave Data The active microwave Synthetic Aperture Radar (SAR) data acquired by RADARSAT-1 satellite were used in this study. With its unique C-band channel, the effective penetration depth of RADARSAT beam is shallower than 5 cm for highly wet soil and deeper than 5 cm for dry soil (Ulaby et al. 1986). Two RADARSAT-1 images have been acquired on July 2 nd and July 12 th, 1997 in SCANSAR Narrow Mode. The characteristics of these images are summarized in Table 1. In order to match the backscattering data to the other input data, such as NDVI, Vegetation optical depth and soil moisture, the images have been geo-coded and co-registered using standard ground control points by maintaining root mean square error (rmse) smaller than one pixel size. Two subset scenes covered by all available data have been selected for analysis. Table 1. Main characteristics of used RADARSAT-1 scenes Parameter Scene-1 Scene-2 Date and time July 02, 1997; 12:27:11 July 12, 1997; 12:36:29 Pass Descending Descending Beam Mode (B) W2 (31-39), S5 (36-42), S6 (41-46) (A) W1 (20-31) W2 (31-39) Product type SCANSAR Narrow SCANSAR Narrow Incidence angle Resolution 25 m 25 m Scene centre N W N W Normalized Difference Vegetation Index (NDVI) The NDVI is defined as the normalized difference between the reflectance in the visible (red) and the near infrared band. The visible (RED) band represents the absorption band of chlorophyll and NIR represents a maximum of vegetation reflectance related to the mesophyll structure. The contrast between vegetation and soil is large in NIR and visible (red) band. The NDVI values are related to the optical properties of vegetation and mainly sensitive to leaf water and chlorophyll content. The NDVI gives an estimation of the health of vegetation. The values close to zero represent areas with little or no vegetation and values close to one indicates high density of green leaves. The vegetation effect on the total backscatter received by the sensor is mainly due to the macrostructure of vegetation canopy such as height of canopy and number of plants or trees per unit area; and the microstructure, which refers to geometry, moisture contents, and vegetation volume fraction of canopies. The penetration depth of radar beam is lower at higher soil moisture contents in vegetation canopies. The total backscattering is composed of backscatter from vegetation and soil, and attenuation caused by the vegetation canopy. The vegetation canopy affects the backscattered energy in two ways: first, the vegetation layer attenuates the soil backscatter contribution and second, the vegetation canopy contributes a backscatter component of its own. The NDVI data have been derived from one Landsat TM image acquired on July 25, Vegetation Optical Depth (VOD) The vegetation optical depth is directly related to the vegetation water content and nature of vegetation cover, and

4 given by the product of vegetation water content and b parameter. The b parameter is function of vegetation dielectric properties, the plant shape or structure, the wavelength, polarization, and look angle (θ) measured from nadir (Jackson and Schmugge 1991). Jackson et al. (1999) specifies the parameter b using land cover classification from published data. The emissivity from ground surface is depend upon vegetation optical depth, is acting as an attenuating layer with transmissivity (γ) and incidence angle. Soil Texture The soil texture is a relative composition of the three major soil classes: sand, silt and clay. The dependence of radar backscatter on soil moisture for different soil textures has been studied by Ulaby et al. (1981). Previous studies showed a strong linear correlation between the backscatter and soil moisture at individual soil texture. However, the slope of linear relation decreases with the increase in clay content of the soil. The reliance of the dielectric constant on soil texture is a function of variation of dielectric constant and water molecule pressure at which it held between soil particles (Ulaby et al. 1986). The sensitivity of soil texture to dielectric constant is lower at dry soil, and higher in wet soil condition (Bindlish and Barros, 2002). At temporal and spatial scales, the soil texture is closely related to soil moisture and radar backscatter. Different soil texture have distinct pattern of soil moisture content and soil moisture drainage (Mattikalli et al. 1998). Soil Moisture Data The soil moisture data measured by ESTAR Instrument (Electronically Scanned Thinned Array Radiometer) during the SGP97 campaign (operated by NASA) were used as truth data in the training and the validation processes. The details about soil moisture retrieval from passive radiometer can be found in (Jackson et al. 1999). The soil moisture data available at 800m x 800m resolution is acquired from NASA website. In order to match the resolution of the measured soil moisture, the SAR spatial resolution was degraded to 800 meters using an averaging algorithm. For the first processing steps, we classified the soil data into 3 classes based on the water content: class 1 (dry soil 0-10%), class 2 (slightly wet soil 11-20%), and class 3 (wet soil +21%). SAR NDVI VOD OSM PSM SAR - Synthetic Aperature Radar (Radarsat data) NDVI - Normalized Difference Vegetation Index VOD Vegetation optical depth OSM Truth soil moisture data PSM - Simulated soil moisture data Figure 2. Soil Moisture mapping. Class 1: SM < 10% Class 2: 10% < SM < 20% Class 3: 20% < SM Unclassified Pixels

5 METHODOLOGY Artificial neural networks have been applied to a wide range of problems in many disciplines. The rapid increase of neural network applications in remote sensing is due mainly to their ability to perform more accurately than other parametric classification techniques especially when we are dealing with non-gaussian classes. Multi-layer perceptron trained with backpropagation algorithm is the most common neural network used for image classification. This type of neural network has been successfully applied to image processing and has shown a great potential in the classification of different remotely sensed data. A useful review of the application of neural networks in remote sensing may be found in (Benediktsson et al. 1990; Paola and Schowengerdt 1995). A multi-layer neural network (or perceptron) consists of a number of interconnected nodes. The nodes are organized into layers where each node transforms the inputs received from other nodes. The adjacent layers are fully interconnected. The input to one node is the weighted sum of the outputs of the previous layer nodes. This sum is then passed through an activation function to produce the final output. The activation function is usually a sigmoid or hyperbolic tangent, which are non-linear functions that have an asymptotic behavior. Further details concerning the training algorithm can be found in (Rumelhart et al. 1986). The training stage consists in adjusting the connection weights (randomly initialized) in order to decrease the difference between the network output and the desired outputs (truth data). The training data were presented to the input layer and propagated through the hidden layer to the output layer. The differences between the computed and the desired outputs were computed and fed backwards to adjust the network connections. This iterative process continued until the mean square error reached a preset threshold or when the validation criteria were reached. When one of the two criterions is met, the training is stopped and the weight values saved. The trained network may now be used as a classifier. In this study, a neural network (NN) based methodology was used to retrieve soil moisture from SAR data and land-cover related information. The NN model was trained by using different combinations of SAR, soil texture, vegetation optical depth and NDVI data. For each combination of inputs, a confusion matrix was generated from the comparison between the real soil moisture values and those predicted by the neural network for different soil moisture classes. The output was evaluated using confusion matrices and the effect of NDVI and vegetation optical depth was analyzed and described in following sections. These results show that the major confusion is between the immediate classes of soil moisture. The vegetation data was shown to have a significant role in reducing confusion between pixels with low soil water contents. Further, these results have clearly showed the effect of vegetation optical depth and NDVI on the accuracy of the retrieved soil moisture. EFFECT OF SUB-PIXEL VARIABILITY OF LAND COVER The effect of sub-pixel variability of a land-cover class on the overall accuracy has been tested. The vegetation classification data for the study area was available in 25m x 25m resolution. The soil moisture and aggregated SAR pixel having resolution of 800m x 800m contains total 1024 (32x32) pixels. A pixel is considered homogeneous if all 1024 sub-pixels are of single vegetation class otherwise the pixel is considered as heterogeneous. The pixel heterogeneity has been defined in three categories H1, H2 and H3, is function of sub-pixel present the big pixel shown in Figure 3. H1: P i 75% H2: 75% > P i 50% H3: 50% > P i 25% (a) (b) (c) Figure 3. Example of heterogeneity of pixels class.

6 The H1 category is assigned to pixels having more 75% of the sub-pixels of a single vegetation cover (see Figure 3-a). Similarly, H2 class contains the sub-pixels ranging from 50% to 75%, and H3 class contains sub-pixels ranging from 25% to 50% (Figure 3-b and Figure 3-c). If a pixel of vegetation class representing less that 25% of sub-pixels is considered as most heterogeneous pixel. The results showed that homogeneous pixels (H1 category) have better accuracy in soil moisture classification. The effect of heterogeneity of pixels for all categories in soil moisture estimation is illustrated in Table 2. Based on above results, we noted that the retrieval of soil moisture from a heterogeneous land surface is difficult and needs more attention to understand the individual patches in a big pixel containing different land areas. Table 2. Effect of heterogeneity of pixel on accuracy of classification Type of Land Mode class H1: P i 75% H2: 75% > P i 50% H3: 50% > P i 25% TP CC % TP CC % TP CC % TP CC % Forage % % % % Pasture/Rangeland % % % % Wheat % % % % summer Legume % % % % Total pixel % % % % * TP = Total pixels, CC = correctly classified pixels. CONCLUSION This study demonstrates promising capabilities in mapping soil moisture from active microwave images. The influence of various parameters such as NDVI, vegetation optical depth, soil texture, and land-cover use can be better understood by the classifiers like neural network, where more than one input parameter can be used to improve the final classification. The additions of vegetation optical depth and NDVI information to NN model have significant effect on the final soil moisture accuracy. Indeed, the pixels having lower NDVI and vegetation optical depth values have a good chance to be accurately classified. However, soil moisture retrieval from pixels with high vegetation optical depth and/or high NDVI values was shown to be less accurate. Further, the results showed that homogeneous area pixels have better accuracy in soil moisture classification. ACKNOWLEDGEMENT This research was funded by the NOAA Cooperative Remote Sensing Science & Technology Center (NOAA- CREST) and the City University of New York. REFERENCES Benediktsson J., P. Swain, and O.K. Ersoy, (1990), Neural network approaches versus statistical methods in classification of multi-source remote sensing data, IEEE Transactions on Geoscience and Remote Sensing, vol. 28-4, pp Bindlish R. and A.P. Barros, (2002). Subpixel variability of remotely sensed soil moisture: an inter-comparison study of SAR and ESTAR. Remote Sensing of Environment, vol. 40-2, pp Engman E.T., and N. Chauhan, (1995). Status of microwave soil moisture measurements with remote sensing. Remote Sensing of Environment, vol. 51, pp Jackson T.J. and T. Schmugge (1991). Vegetation effects on the microwave emission of soils. Remote Sensing Environment, vol. 36, pp

7 Jackson T.J., D. Le Vine, Hsu A.Y., Oldak A., Starks P., Swift C., Isham J., and M. Haken (1999). Soil moisture mapping at regional scales using microwave radiometry: the Southern Great Plains Hydrology Experiment. IEEE Transactions on Geoscience and Remote Sensing, vol. 37, pp Kasischke E., Smith K., L. Bourgeau-Chavez, Romanowicz E., Brunzell S., and C. Richardson (2003). Effects of seasonal hydrologic patterns in south Florida wetlands on radar backscatter measured from ERS-2 SAR imagery. Remote sensing of environment, Vol. 88 (4), pp Mattikalli N., Engman E.T., Ahuja L., and T.J. Jackson (1998). Microwave remote sensing of soil moisture for estimation of profile soil property. International Journal of Remote Sensing, vol. 19-9, pp Paola J., and R.A. Schowengerdt, (1995). A review and analysis of back-propagation neural networks for classification of remotely-sensed multi-spectral imagery. International Journal of Remote Sensing, vol , pp Rumelhart D.E., Hinton G.E., and R.J. Williams (1986). Learning internal representation by error propagation, Parallel distributed processing: Exploration in the microstructure of cognition, M.I.T., Cambridge Press (MA), pp Ulaby F.T., Dobson M., and G. Bradley (1981). Radar reflectivity of bare and vegetation covered soil. Advanced Space Research, vol. 1, pp Ulaby F.T., Dubois, P.C., and J.J. Van Zyl, (1996). Radar mapping of surface soil moisture. Journal of Hydrology, vol. 184, pp Ulaby F.T., R. Moore, and, A. Fung, (1986), Microwave Remote Sensing, Active and Passive, From, Theory to Applications, Artech House, Norwood, MA.