1Remote Sensing Laboratory Dept. of Information Engineering and Computer Science University of Trento Via Sommarive, 14, I-38123 Povo, Trento, Italy 2Institute for Applied Remote Sensing for Alpine Environment Eurac Research Viale Druso, 1, I-39100 Bolzano, Italy 3Institute 4Department of Computer, System and Production Engineering Tor Vergata University Via del Politecnico, 1, I-00133 Rome Italy Spatial and Temporal Mapping of Soil Moisture Content with Polarimetric RADARSAT2 SAR Imagery in the Alpine Area E-mail: luca.pasolli@disi.unitn.it luca.pasolli@eurac.edu Web: http://rslab.disi.unitn.it http://www.eurac.edu Luca Pasolli1,2 Claudia Notarnicola2 Lorenzo Bruzzone1 Giacomo Bertoldi3 Georg Niedriest3 Ulrike Tappeiner3 Marc Zebisch2 Fabio Del Frate4 Gaia Vaglio Laurin4 Outline 1 Introduction 2 Aim of the Work 3 Study Area and Dataset 4 Estimation System Description 5 Analysis of Results 6 Conclusion IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 2 Introduction SOFIA: SOil and Forest Information retrieval by using RADARSAT2 images • ESA AO-SOAR 6820 • Supported in the framework of the IRKIS project (Civil Protection Department, Province of Bolzano) Main Innovative Aspects: • Fully-polarimetric RADARSAT2 satellite SAR data • Mountain landscape (Alpine area) • Advanced estimation methods Objectives: • Estimation of soil moisture content on bare and vegetated areas (alpine meadows and pastures) • Estimation of vegetation biomass (forest) • Investigation on the influence of soil and vegetation parameters in connection to natural hazard in Alpine regions. IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 3 Introduction Soil moisture estimation supports various application domains: • drought monitoring • flood and landslide prediction • climate change analysis Challenges: • non-linearity of the relationship between microwave signals and soil moisture • sensitivity of microwave signals on different target properties (moisture content, roughness, vegetation, land use) • influence of topography on the microwave signal acquired by the sensor In a previous study (Pasolli et al., 2010) RADARSAT2 SAR images have shown to be promising for the retrieval of soil moisture in Alpine areas: • by integrating the information coming from ancillary data • by exploiting an advanced retrieval algorithm based on the Support Vector Regression (SVR) method L. Pasolli, C. Notarnicola, L. Bruzzone, G. Bertoldi, S. Della Chiesa, V. Hell, G. Niedrist, U. Tappeiner, M. Zebisch, F. Del Frate, G.V. Laurin, “Estimagion of Soil Moisture in an Alpine catchment with RADARSAT2 images”, Applied and Environmental Soil Science, in press IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 4 Aim of the Work To Further Investigate the Retrieval of Soil Moisture from RADARSAT2 SAR Images in Alpine Areas 1.By exploiting the fully-polarimetric capability of RADARSAT2 in combination with standard and advanced feature extraction/selection methods 2.By extending the analysis in time and space with the available images IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 5 Study Area Mazia Valley, Alto Adige, Italy Well known and monitored area Well representative in terms of • Topography • Land use • Soil moisture content conditions IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 6 Dataset 1.Satellite SAR images: • 4 RADATSAT2 quad-pol standard mode images (3rd June, 21st July, 14th August, 5th October 2010) • DEM geocoded, filtered (Frost 7x7) • Final pixel size 20 m 2.Field measurements: • 77 soil dielectric constant measurements on meadows (blue) and pasture (red) acquired contemporary to satellite overpasses (3rd June and 21st July) RADARSAT2, 21° July 2010 (R=HH, G=HV, B=VV) 3.Ancillary data: Meadow • DEM (pixel size 2.5 m) • NDVI map extracted from MODIS Terra images (pixel size 250 m) • Land use map (meadows, pasture); Pasture June July June July Min Diel 6.7 3.8 6.4 3.2 Max Diel 21.8 27 16.2 25.63 Average Diel 16.5 14.2 11.2 7.7 IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 7 Estimation System Polarimetric RADARSAT2 SAR image Data Pre-processing Feature Extraction & Selection Ancillary Data Retrieval Algorithm Estimated Soil Moisture Content Map IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 8 Estimation System: Retrieval Algorithm Polarimetric RADARSAT2 SAR image Aim: to define the mapping between the input features and the target biophysical variable • Support Vector Regression (SVR) technique trained on Field Reference Samples • Multi-objective Model Selection Approach Data Pre-processing Training Phase Features from Ancillary Data Features from Remotely Sensed Image Ground Truth Reference Feature Samples Extraction & Selection Sub-Sample K-Fold Cross Validation Validation Set Training Set Generator Ancillary Data SVR Learning Retrieval Algorithm Performance Evaluation SVR Estimation SVR Parameters Config. Multi-Objective Model Model Selection Selection Estimation Perform. (MSE, R2) Estimation Operational Phase Input Features Estimated Soil Moisture (Image Content + Ancillary) Map SVR Estimator IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 Output SMC Value 9 Estimation System: Features Extraction and Selection Polarimetric RADARSAT2 SAR image Aim: to extract and select from the remotely sensed data the most relevant information for the estimation problem considered Features Extraction Data Pre-processing Feature Extraction & Selection Ancillary Data • Standard Intensity&Phase SAR processing • Polarimetric backscattering coefficients • Polarimetric Combinations: Span (HH+HV+2HV), Polarization Ratio (HH/VV) and Linear Depolarization Ratio (HV/VV) • Polarimetric phase difference (PPD) and interferometric coherence • Polarimetric Decompositions Retrieval Algorithm • H/A/α decomposition • General purpose feature extraction techniques • Independent Component Analysis (ICA) Estimated Soil Moisture Content Map Features Selection • Sequential Forward Selection (SFS) strategy with performance evaluation on a subset of reference samples IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 10 Experimental Setup Experiment 1: Assessment of the Estimation System with the proposed Feature Extraction & Selection strategies • 60 reference samples for training/tuning the estimation system according to a 5-fold cross validation procedure • Retrieval Algorithm Settings: • SVR with Gaussian RBF kernel function • Hyper-parameters ranges: 10-3 < γ < 103 , 10-3< C < 103 , 10-3 < ε < 10 • Multi-objectives model selection according to RMSE and R2 quality metrics • Performance assessment on 17 independent test reference samples according to: • Root Mean Squared Error (RMSE) • Determination coefficient (R2) • Slope and Intercept of the linear tendency line between estimated and measured target values Experiment 2: Assessment of Spatially and Temporally Distributed Soil Moisture Estimates in the Alpine Area • Exploitiation of the estimation system configuration identified in Experiment 1 • Generation of soil moisture content maps associated with RADARSAT 2 SAR images time series acquired during summer 2010 • Qualitative and quantitative assessment with prior knowledge on the area and field station measurements IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 11 Results: Experiment 1 Selected Features RMSE R2 Slope Intercept 0.77 2.13 0.8 2.37 0.86 1.53 Reference HH 2.79 0.79 Intensity & Phase Features HH HV/VV 2.55 0.82 ICA Features ICA1 ICA4 2.66 0.81 Cloude Decomposition Features αA 3.1 HH HV/VV features 0.73 0.76 3.09 ICA1 ICA4 features IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 HH feature α A features 11 Results: Experiment 2 Estimated Soil Moisture Content Map, June 2010 IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 11 Results: Experiment 2 Estimated dielectric constant map, July 2010 Estimated dielectric constant Map, August 2010 Estimated Dielectric constant Map, June 2010 Estimated dielectric constant Map, October 2010 IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 14 Conclusion The potential of fully-polarimetric RADARSAT 2 SAR images in combination with an advanced retrieval algorithm has been investigated for the mapping in space and time of soil moisture in the Alpine area 1. Polarimetric features are effective for improving the retrieval of soil moisture in the challenging Alpine environment • Generally, they allow one to reduce the ambiguity in the data and increase the accuracy of the estimation • The HH HV/VV configuration has shown to be the most suitable in this specific operative conditions 2. The achieved results suggest the potential of the proposed estimation system in combination with RADARSAT 2 SAR data for the retrieval of soil moisture in Alpine areas • Good capability to reproduce the spatial patterns of the desired target parameter • Good agreement with the measured temporal trends of soil moisture Future work • Investigation of the proposed estimation system in combination with higher geometrical resolution polarimetric SAR data • Integration of data from different sensors (e.g., L-Band SAR images) IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 15 Thank you for the Attention!! Questions? luca.pasolli@disi.unitn.it luca.pasolli@eurac.edu IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011 16