A synergetic use of observations from MODIS SEVIRI

A synergetic use of observations from MODIS, SEVIRI MSG, ASAR and AMSR-E to infer a daily soil moisture index C. Notarnicola 1, F. Di Giuseppe 2, K. Lewinska 1, L. Pasolli 1, 3, M. Temimi 4, B. Ventura 1, M. Zebisch 1 1 EURAC-Institute for Applied Remote Sensing, Viale Druso 1, Bolzano, Italy. 2 ARPA-Servizio. Idro. Meteo. Clima, 3 Dep. Viale Silvani 6, Bologna, Italy of Information Engineering and Computer Science, University of Trento, Via Sommarive, 14, Trento, Italy. 4 NOAA-CREST/The City University of New York, The City College, 140 th St @ Convent Ave. Steinman Hall (T-109), New York, NY 10031. IEEE International Geoscience and Remote Sensing Symposium - IGARSS 2011 Vancouver, Canada - July 24 -29, 2011

Outline ØIntroduction ØMain concept: multi-sensor approach ØTest sites and EO data ØApparent Thermal Inertia (ATI) approach ØSynergy with SAR images and AMSR-E data ØExperimental results ØConclusions and future steps 2

Introduction Techniques and methods to evaluate soil and vegetation water content includes as main instruments passive and active microwave methods but also some indirect measurements based on radiometric techniques in the opticalthermal range. Due to large data set availability from different sensors, in the last years the synergy among sensors and exploitation of multi-sensor approaches has increased notably. The objective of this study is to infer a soil moisture index (soil moisture classes) from an approach mainly based on the concept of apparent thermal inertia (ATI) by the following steps: ØExploitation of the Apparent Thermal Inertia (ATI) from optical sensors (MODIS) ØContinuous calibration with SAR images (when available) and AMSR-E data ØSynergy with SEVIRI MSG acquisitions 3

Main concept: multi-sensor approach Downscaling and cloud cover reduction Check spatial distribution and calibration steps ATI from MODIS images (or equivalent sensors) Soil moisture estimates from SAR sensors ATI from MSG images (or equivalent sensors) Check temporal trends and regularization Soil moisture estimates from AMSR-E sensors Daily Soil Moisture Index (3 -4 classes)

Test sites and EO data ST-IT, test site located in Italy : Emilia Romagna region-red dot (ARPA Emilia Romagna) ST-FR, test site located in France (belonging to SMOSMANIA network), near the Pyrenees-yellow dot (after Google Earth©). Data availability: -MODIS images for year 2007, 2008 -2009, 2010 -AMSR-E/Aqua level 3 global daily surface soil moisture images -SAR images (when available on Emilia Romagna test sites) -SEVIRI MSG acquisition contemporary to MODIS images ST-IT, test site located in Italy : South Tyrol-green dot (EURAC-Institute for Alpine Environment)

Apparent Thermal inertia (ATI) algorithm vegetation Tempe rature Physical Thermal Inertia (TI) (Watson et. al. , 1971; Price 1977) • Response to temperature change • Physical TI = √(density*thermal conductivity*heat capacity) bare soil Apparent Thermal Inertia (ATI) (Price 1985; Mitra & Majumdar 2004; Claps & Laguardia, 2004) • ATI = (1 -albedo) / (Temperature max - Temperature min) 00: 00 • Thermal image pair solar noon and pre-dawn MYD L 1 B Georeferencing/ Radiometric cal. BT 11 m night Re-projection Albedo calculation Night acq. MYD 09 12: 00 24: 00 Time BT 11 m day Day acq. MYD L 1 B water Cloud screening (MYD 035) ATI

ATI error analysis If the limit of T =10 K is considered in order to have reliable ATI estimates [Cai et al. , 2007], the error on ATI is around 0. 002 which corresponds to 5% of the lowest values detected in this analysis.

Cross-comparison with SMC retrieved from SAR images ATI-Ground measurements ATI-SMC_SAR ATI=1. 8*10 -3*SMC+0. 0154 ATI=1. 3*10 -3*SMC+0. 019 R 2=0. 76 R 2=0. 78 (SMC expressed in %) Three classes (Emilia Romagna Test site): ATI < 0. 04, SMC-SAR < 10% 0. 04 < ATI < 0. 05, 10% < SMC-SAR < 15% ATI (K-1) Urban ATI > 0. 05, Forest SMC-SAR > 15%, Water OA=72% ATI SAR bodies

Temporal filtering with AMSR-E data Even under similar soil moisture conditions and acquisition time, the ATI values show a high variability. It is therefore necessary to introduce a filtering technique to reduce the noise in the observed data. The use of microwave based time series of soil moisture to refine the ATI based product should perform better than any other stand-alone signal analysis technique like moving average as the microwave estimates are intrinsically consistent with ATI estimates. The main assumption of this study is the agreement between soil moisture estimates from microwave and ATI. This expected agreement fosters using the microwave time series to filer and refine the ATI product. A temporal moving window has been considered by using the following expression: • If there is no significant changes: the average values of SMCAMSRE(ti+m) are equal to SMCAMSRE(ti), the weight is equal to 1 and the filter performs only a simple mean over the days considered; • If there is an increase in soil moisture values: the average values of SMCAMSRE(ti+m) are higher than SMCAMSRE(ti), the variations are enhanced and so the effect of averaging of the filter is reduced; • If there is a decrease in soil moisture values: the average values of SMCAMSRE(ti+m) are lower than SMCAMSRE(ti), the variations are reduced and then smoothed by the filter, thus reducing the noise.

Analysis of temporal trends: Emilia Romagna Comparison of the temporal trend among SMC (cm 3/cm 3), ATI originally calculated and ATI filtered. H stands for High NDVI values (> 0. 4) and L stand for Low NDVI values (< 0. 4). NDVI < 0. 4 NDVI > 0. 4 No filter 0. 58 0. 45 Simple filter 0. 59 0. 45 Filter using AMSRE data 0. 72 0. 56 Comparison between ATI and measured soil moisture values (SMC) over 1 year period for Emilia Romagna test site. The values represent the determination coefficients between ATI values and SMC in the different cases considered.

SMC classes from ATI From the error analysis on ATI, we need to consider: • For theoretical error the lowest value 0. 015 K-1 has been considered, thus assuming that the filtering notably reduces the effect due to acquisition time; • The standard deviation values of the different ROI considered have a median value around 0. 003 K-1. Considering these two independent sources of errors, the total error is: 0. 0153 K -1. As the ATI values range from 0. 04 K-1 to 0. 10/0. 12 K-1, the total error determines the possibility to detect at least four classes. In order to determine the class boundaries and to verify their consistency with the error on SMC measurements (around 5% of the measured value), the clustering tool of Maltlab has been considered. 4 - classes ATI/SMC 1 (<0. 17) 2 (0. 17 -0. 25) 3 (0. 25 -0. 3) 4 ( >0. 3) 1 (< 0. 05) 0. 57 0. 24 - 2 (0. 05 -0. 07) 0. 43 0. 62 0. 25 3 (0. 07 -0. 085) 0. 09 0. 33 0. 25 4 (>0. 085) 0. 05 0. 44 0. 50 The overall accuracy with four classes is around 51%. If we exclude the values of ATI within the confidence interval corresponding to the error measurements of SMC values that can be misclassified, the accuracy raises to 81%. 3 - classes SMC/ ATI 2 (<0. 20) 3 (0. 20 -0. 30) 4 ( >0. 30) 2 (<0. 055) 0. 76 0. 12 - 3 (0. 055 -0. 085) 0. 24 0. 60 0. 50 4 (>0. 085) 0. 28 0. 50 In this case, the overall accuracy is around 65%, and rises to 88% considering the misclassified values due to their position very close to the class boundaries.

Analysis of temporal trends: France Comparison of the temporal trend among SMC (cm 3/cm 3), ATI originally calculated and ATI filtered. H stands for High NDVI values (> 0. 4) and L stand for Low NDVI values (< 0. 4). NDVI < 0. 4 NDVI > 0. 4 No filter 0. 61 0. 23 Simple filter 0. 68 0. 24 Filter using AMSRE data 0. 70 0. 43 Temporal comparison between ATI and measured soil moisture values (SMC) over 1 year period for France test site. The values represent the determination coefficients between ATI values and SMC in the different cases considered.

SMC classes from ATI 3 - classes ATI/SMC 2 (<0. 25) 3 (0. 25 -0. 35) 4 ( >0. 35) 2 (<0. 06) 0. 78 0. 37 0. 17 3 (0. 06 -0. 085) 0. 22 0. 53 0. 75 4 (>0. 085) 0. 10 0. 08 The overall accuracy with three classes is around 58%. If we exclude the values of ATI within the confidence interval corresponding to the error measurements of SMC values that can be misclassified, the accuracy raises to 73%. The ranges adopted are slightly different from the previous test site, because the SMC values were in general higher while the corresponding ATI did not change due to the presence of vegetation detected with high NDVI values. In fact in the confusion matrix most of the values of SMC higher than 0. 35 cm 3/cm 3 are in class 3 instead of 4. This happens because all the highest SMC values were in the period with the highest values of NDVI.

Example of ATI and derived SMC classes

Time series on the South Tyrol test site Matchertal watershed – One of the driest valley in South Tyrol ATI (K-1) The mismatch between ATI values and SMC measurements can be due to: -High level of vegetation: NDVI is generally higher than 0. 6 -Difficulty in eliminating completely the cloud presence (the completely cloud free images were only 20% over 2010 acquisitions) -Topography effect 15 -Time of the acquistions

Time of acquisitions

Spatial analysis ATI spatial distribution has been analyzed through ANOVA with respect to: - Land cover (1 – urban areas; 21 – agriculture, arable land; 22 – agriculture, vineyards and other ‘standing crops’ 31 – natural vegetation, trees, forest; 32 – natural vegetation, grassland; 41 – no vegetation, rocks; 42 – no vegetation, debris 43 – no vegatation, glacier; 5 – water; 0 – no data) -Elevation classes (>500 m, 500 -1000 m, 1000 -1500 m, 1500 -2000 m, > 2000 m) -NDVI classes (8 classes with border values of 0. 1, 0. 2, 0. 3, 0. 4, 0. 5, 0. 6, 0. 7 and 0. 8) Tests of of Between-Subjects Effects Dependent Variable: 145 Variable: 156 Type. IIISum Sumof of Source Squares Noncent. df df Mean Square FF Sig. Parameter Observed Powerbb Corrected Model . 274. 234 aa 81 51 . 003. 005 12. 780 24. 856 . 000 1035. 194 1267. 661 1. 000 Intercept 1. 284. 509 11 1. 284. 509 1924. 258 6955. 441 . 000 1924. 258 6955. 441 1. 000 LC El . 023. 011 105 . 002 11. 751 8. 819 . 000 88. 192 58. 754 1. 000 NDVI_145 NDVI_156 . 014. 043 88 . 002. 005 29. 079 6. 651 . 000 232. 630 53. 206 1. 000 LC El ** NDVI_156 NDVI_145 . 024. 042 63 38 . 000. 001 1. 424 6. 035 . 016. 000 229. 344 89. 743 1. 000 Error 1. 920 1. 353 7256 7327 . 000 Total 26. 702 20. 566 7338 7379 2. 194 1. 587 7337 7378 Corrected Total a. R Squared =. 125 (Adjusted R Squared =. 115) b. Computed using alpha =. 05 NDVI and Land cover classes are significant in all the analyzed days. This impact (value of F-statistic) depends on the day, so probably on NDVI value – phenology; BUT weather conditions (cloud cover, fog etc) also might play here important role!

Spatial analysis cont’d

Example of ATI maps

Cross-comparison with SEVIRI- MSG data A cross-comparison with METEOSAT MSG data is under evaluation. Some MSG images contemporary to MODIS acquisitions have been processed analyzed over the Emilia Romagna test sites. One of the major problems is the use of a correct cloud mask in order to not introduce not corrected ranges of ATI values.

Conclusions and future steps This work proposes a multi-sensor analysis in order to determine soil moisture classes based on ATI derived from MODIS images. 3 -4 classes of SMC were detectable. For the approach, main limitation are the cloud coverage, the acquisition time of day and night images and the presence of vegetation. The ATI values have been compared with ground measurements and with soil moisture maps estimated from SAR images in order to verified the spatial pattern consistency. The main limitation remains the low repetition time of SAR images The temporal trends have been filtered by using the AMSR-E data in order to reduce the effect due to outliers (time of acquisition, clouds, etc. . ). In this case, the main limitation still remains the presence of vegetation. First comparisons between MODIS and MSG images indicate a good correlation between these data. For MSG a good cloud cover is highly recommanded. Future developments will include: -Further comparison with MSG data, including improvement of the cloud mask -Definition of a calibration procedure between MODIS and MSG data -Integration of MODIS time series with MSG data -Considering other data such as SMOS, Aquarius

Thank you for the attention! Comments/questions? 22