Read-me-first note for the release of the SMOS Level 2 Soil Moisture Near Real Time Neural Network (L2-SM-NRT-NN) data product

Transcription

1 Read-me-first note for the release of the SMOS Level 2 Soil Moisture Near Real Time Neural Network (L2-SM-NRT-NN) data product Processor version Level 2 Soil Moisture Near Real Time Neural Network V100 Release date by ESA 16 March 2016 Further information Details on the neural network design, training and the processor can be found in: Rodríguez-Fernández, N. J., et al., 2015: Soil moisture retrieval using neural networks: application to SMOS. IEEE Transactions on Geoscience and Remote Sensing, Volume:53, NO. 11 November 2015 Rodríguez-Fernández, N. J., et al., 2015: SMOS Near-Real- Time soil moisture processor. Part 1: Neural network evaluation and algorithm description, CESBIO-SMOS report SO-TN-CB-GS-0049 v1.5, Muñoz-Sabater, J., et al., 2016: SMOS Near-Real-Time soil moisture processor: Operational chain and evaluation, ECMWF TR2, WP 4020, 01/2016 Rodríguez-Fernández N.J. et al. 2016: Evaluation of the SMOS Near-Real-Time soil moisture (in prep) L2 soil moisture operational product V620 read-me-first note can be found here ( Contact for helpline Comments on the NRT SM Product For all issues related to data access, formats and read/write, processors, please contact ESA s HelpDesk at eohelp@esa.int. Feedback on the NRT soil moisture product can be provided directly to ESA s HelpDesk at eohelp@esa.int.

2 1. Rationale The SMOS Near-Real-Time (NRT) soil moisture data (SM-NRT-NN) answer the need of operational agencies to get highly accurate information with minimum data latency. To meet the NRT requirement (i.e. data provision within 3 hours of sensing) the number of auxiliary data files needed in the processor must be kept at a minimum and the processing must be fast. At the same time, the product quality must remain high and be comparable with the product obtained for the operational (geophysical) Level 2 soil moisture (L2) processor, run as part of the ESA ground segment. It is desirable to keep the NRT and operational L2 soil moisture products closely coupled to allow for an efficient product evolution and a common validation procedure. It was decided to use the SMOS NRT brightness temperature product provided in BUFR format as an input to a Neural Network (NN) that was trained using the operational Level 2 soil moisture (L2 SM) product. Essentially, the NN replaces the nominal L2 SM processor and NRT brightness temperatures are used instead of the L1C brightness temperatures. Consequently, the two soil moisture data sets, NRT and L2, exhibit the same global climatology but can show local differences. Overall, the NRT product quality is as good as or better than the one of the L2 SM data set. In this note, we briefly describe the processor, the neural network training and focus on the verification of the data product. 2. Product generation and limitations The Soil Moisture Near Real Time Processor (SM-NRT-OP) provides fast retrieval of soil moisture measurements from the multi-angular brightness temperature as available inside the SMOS level 1 Near Real Time BUFR product. The soil moisture retrieval is based on a neural network schema trained using the level 2 soil moisture data set generated by the operational L2 soil moisture processor V620. The input configuration for the neural network is a vector of 13 elements including: six SMOS brightness temperature from 30 to 45 degrees incidence angles in 5 degrees-width bins for both H and V polarization, the six associated local normalization indices reflecting the dynamic range of soil moisture, and the 0-7 cm soil temperature forecast from ECMWF. The neural network architecture has two layers with a hidden layer containing 5 neurons. The processor consists of two parts: The NRT product generator operated at ECMWF which provides the Soil Moisture NRT product within about four hours from sensing and the offline processor developed and maintained at CESBIO. The offline processor is the essential part for the training of the NN, the quality control, and the definition of the minimum and maximum measurements obtained locally. In the product generator the NRT TBs are pre-processed and quality controlled, the weights are applied, and the data product is generated. The data files are then disseminated through EUMETCAST, GTS via ESA and SMOS data portal. The filename convention adopted is the following: Filename convention for the SM-NRT-NN product Filename: W_XX-ESA,SMOS,NRTNN_C_LEMM_time1_time2_time3_o_v100_l2sm.nc Fields description W_XX-ESA,SMOS,NRTNN_C_LEMM Fixed WMO product identifier time1 Product generation time. Format YYYYMMDDhhmmss time2 Acquisition time of the first observation into the product. Format YYYYMMDDhhmmss

3 time3 o v100 l2sm nc Acquisition time of the last observation into the product. Format YYYYMMDDhhmmss Fixed one character for operational data. Processor version used to generate the product Fixed SMOS product identifier Fixed product format identifier for netcdf Time format: YYYYMMDDhhmmss, where YYYY corresponds to the year, MM to the month, DD to the day, hh to the hour, mm to the minutes, and ss to the seconds Filename example: W_XX-ESA,SMOS,NRTNN_C_LEMM_ _ _ _o_v100_l2sm.nc To access the SMOS data see here ( The design of the NN and the processing result in some limitations with respect to the operational Level 2 soil moisture product: - Reduced swath of about 915 Km due to the usage of brightness temperature from 30 to 45 degrees incidence angle - Circular gaps in case a full set of measurements is not available - Reduced set of parameters contained in the netcdf structure if compared with the operational Level 2 soil moisture product: List of parameters available inside the SM-NRT-NN product SM-NRT-NN NetCDF Field Units Description Latitude Longitude degrees Geographic latitude of the retrieval degrees Geographic longitude of the retrieval Soil moisture Soil moisture error m 3 m 3 Soil moisture retrieval value m 3 m 3 Estimated uncertainty of the retrieval RFI probability % Probability that the retrieval is affected by RFI. The RFI probability is computed per grid point, and it is defined as the number of BUFR brightness temperature observations flagged as affected by RFI with respect to the total number of observations remaining after filtering. Number of days since days Acquisition day of the retrieval computed as number of days since 01 January 2001 Number of seconds since midnight UTC seconds Acquisition time of the retrieval computed as number of seconds since 00UTC SMOS DGG id - Iidentification number for the latitude, longitude in the SMOS ISEAS 4H9 grid. The SM-NRT-NN is an only land product.

4 3. Training of the Neural Network (NN) for the generation of the SM-NRT-NN product Using reprocessed NRT brightness temperatures in BUFR format from 01/06/2010 to 30/06/2012, the SM-NRT-OP processor has been used to compute a training data base with brightness temperatures in H and V polarizations and three incidence angle bins from 30 o to 45 o. This configuration is the best trade-off of retrieval accuracy and swath width. The NRT-NN SM is retrieved in swaths of about 915 km (the standard L2 algorithm retrieves soil moisture in swaths of _ about 1150 km). From more than 1 Mio vectors available, one fifth (i.e ) have been selected as the training data base vectors. A subset of 60% of those is used for the actual training, 20% is used for evaluation of the NN performances during the training and to avoid over-training, and the final 20% is used to test the performances of the trained NN a posteriori. Gradient back-propagation and minimization with the Levemberg-Marquard algorithm has been used. One single hidden layer with 5 neurons has been used, as it has been shown in that it is enough to capture the relationship in between the input data and the reference SM and one wants to keep the NN as simple as possible. No signs of overtraining have been found and the training has been stopped after 50 iterations when the mean squared difference is asymptotically approaching to a minimum. Figure 1 shows an example of a retrieval over the same orbit with L2-SM and SM-NRT-NN. 4. Validation of SM-NRT-NN product The performances of the SM-NRT-NN product, generated by the version V100 of the Soil Moisture Near Real Time Processor (SM-NRT-OP), have been evaluated against the operational level 2 soil moisture products V620 for two different acquisition periods: i) from May 2015 to November 2015, ii) from June 2010 to July 2012 in order to fully verify seasonal and long term data performances. Verification activities have been also complemented by in situ measurements comparison. The next three paragraphs summarize the results obtained.

5 Figure 1:comparison between SMOS SM-L2 and SM-NRT-NN 4.1 SM-NRT-NN comparison to level 2 soil moisture operational product V620 (2015) The NRT SM processor has been applied to NRT TBs and compared to SMOS L2 SM obtained with version 6.20 of the SMOS operational processor from 15/May/2015 to 25/November/2015. Figures 2 to 4 show maps and histograms computed from local (all DGGs) statistical metrics obtained over that period. The typical number of points with both NRT-SM-NN and SM-L2 in that period is about 110 (upper panel of Fig. 2). The correlation of both products is very high (> 0.7) over a large part of North-America, the southernmost part of South-America, the Iberian peninsula, the Sahel and South- Africa, Australia and parts of central Eurasia. The correlation is significantly lower over forest (both tropical and boreal) and over deserts such as the Sahara, where the variance is low and driven by the noise. In conclusion, both products show similar dynamics over large parts of the Globe. The bias map (upper panel of Fig. 3) shows that the SM-NRT-NN product has a tendency to underestimate the L2-SM dataset, which is an expected behaviour as it has been obtained using a regression technique and extreme values are under-represented in the reference dataset. The most significant effect of the bias is to increase the RMSD with respect to the STDD in parts of Europe and Canada. However, one should note that both the RMSD and the STDD are lower than 0.04 m3/m3 over most of the Globe

6 (all except the reddish regions in the two lower panels of Fig. 3, see also the positions of the peak of the histograms.) Finally, Fig. 4 shows the mean of the SM-NRT-NN and SM-L2 over the period of the study. Both maps show an overall excellent agreement, although it is possible to appreciate the negative bias in the NRT-SM-NN product in some regions. In any case, note that it has been preferred to compare the mean of both products instead of the median to be more sensitive to the possible underestimation of extreme values in the SM-L2 dataset. Figure 2: From top to bottom: maps (left) and histograms (right) of the number of points (with both SM-NRT-NN and SM L2), the Pearson correlation of SM-NRT-NN with respect to L2 SM and the probability of have a given correlation by chance (in logarithmic scale) for each DGG point.

7 Figure 3: From top to bottom: maps (left) and histograms (right) of the bias (mean SM-NRT-NN minus mean SM L2), the RMS and the STD of the difference of SM-NRT-NN with respect to L2 SM for each DGG point.

8 Figure 4: From top to bottom: maps (left) and histograms (right) of the mean SM-NRT-NN and mean L2 SM for each DGG point SM-NRT-NN comparison to level 2 soil moisture product V620 ( ) Since at the time of developing and testing the NRT processor the period available with operational v6.20 data was relatively short, covering not even a whole year, the same kind of global analysis discussed in the previous section has been done applying the SM-NRT-OP processor to NRT brightness temperatures reprocessed with the same parameters of current operational NRT. The period of the study was June June The NRT-SM-NN products have been compared to v6.20 operational Level 2 soil moisture (L2 SM). Figures 5 and 6 show the results. The results are comparable to those obtained with operational data in a shorter period discussed above. However, it is noteworthy that the Pearson correlation map (upper panel of Fig. 5 shows higher values over larger regions of the Globe, which shows that the temporal dynamics of both products are actually in better agreement when a longer time period is considered.

9 Figure 5: From top to bottom: (i) Pearson correlation (adimensional) in between SM-NRT-NN and SM-L2. (ii) Local mean value of SM-L2 (m 3 /m 3 ). (iii) Local mean value of SM-NRT-NN (m 3 /m 3 ). These maps have been computed comparing the SM-NRT-NN products and the SM-L2 reprocessed v6.20 data from June 2010 to June 2012.

10 Figure 6: From top to bottom: (i) Bias of the SM-NRT-NN product with respect to the SM-L2 product: mean of SM-NRT- NN minus mean SM-L2 (m 3 /m 3 ). (ii) Root mean square of the difference (m 3 /m 3 ). (iii) Standard deviation of the difference (m 3 /m 3 ). These maps have been computed comparing the SM-NRT-NN products and the SM-L2 reprocessed v6.20 data from June 2010 to June 2012.

11 4.3. SM-NRT-NN comparison to in situ measurements The SM-NRT-NN product has been evaluated against in situ measurements from the SCAN 1 and USCRN 2 networks. These networks of in situ measurements have been extensively used for the validation of remote sensing data 3. The in situ data have been obtained directly from the teams operating both networks. After a quality inspection of the data, the sites listed in Table 1 have been selected for the current comparison. Table 1: List of the sites used for the simultaneous evaluation of SM-L2 and SM-NRT-NN (127 sites). lat deg lon deg lat deg lon deg lat deg lon deg G.L. Schaefer, M.H. Cosh, and T.J. Jackson. The USDA natural resources conservation service soil climate analysis network (SCAN). Journal of Atmospheric and Oceanic Technology, 24(12): , J.E. Bell, M.A. Palecki, C.B. Baker, W.G. Collins, J.H. Lawrimore, R.D. Leeper, M.E. Hall, J. Kochen- dorfer, T.P. Meyers, T. Wilson, and H.J. Diamond. US. climate reference network soil moisture and temperature observations. J. Hydrometeorol., 14: , C. Albergel, P. de Rosnay, C. Gruhier, J. Muñoz-Sabater, S. Hasenauer, L. Isaksen, Y. Kerr, and W. Wagner. Evaluation of remotely sensed and modelled soil moisture products using global ground-based in situ observations. Remote Sensing of Environment, 118: , 2012.

12 Several quality metrics have been computed site per site independently for the SM-NRT-NN and the SM-L2 products. Table 2 summarizes the results in the form of averages over all the sites (for the Pearson correlation also the median value is given). Table 2: Comparison to in situ measurements over the USCRN and SCAN networks. The columns are: the SM product, the mean number of points in the time series, the mean and median Pearson correlation with respect to in situ measurements, the mean bias (mean in situ SM minus mean SMOS SM), the RMS and STD of the difference time series averaged over all sites, and the Anomaly Correlation Coefficient 4. The statistics have been computed independently for the SM-NRT-NN and the SM-L2 product. The number of SM retrievals is, on average, larger for the SM-L2. The number of sites with available data for the evaluation is 144 for SM-L2 and 155 for SM-NRT-NN. Table 1 shows the coordinates of all the sites SM Mean Npts Mean R Median R mean Bias mean RMSD Mean STDD mean ACC L NRT The mean number of points in the time series from May 2015 to November 2015 is 186 for the SM-L2 product while is only half of that value for the SM-NRT-NN product. The reason is a lower revisit time for the SM- NRT-NN product due to the somewhat narrower swath of and the fact that retrievals are only available when all six 6 TBs are well defined (for both polarizations and the three angle bins from 30 o to 45 o ). It is noteworthy that the current implementation of the SM-NRT-NN processor does not perform any interpolation of the TB versus incidence angle profiles. Both SMOS products show a very similar mean bias with respect to the in situ measurements, while the mean STDD and RMSD are slightly lower for the SM-NRT-NN product. In order to get further insight into the intrinsic quality differences of both datasets, the same statistics have been computed but only keeping times for which both SMOS products are retrieved. The results are shown in Table 3. The differences in the evaluation of both products decreases, but the NRT product still shows a larger correlation and lower STDD with respect to in situ measurements than the L2 product. This result is in perfect agreement with the results obtained with the NRT prototype implemented in the first phase of this project 5. 4 C. Albergel, C. Rüdiger, D. Carrer, J.C. Calvet, N. Fritz, V. Naeimi, Z. Bartalis, and S. Hasenauer. An evaluation of ASCAT surface soil moisture products with in-situ observations in South western France. Hydrology and Earth System Sciences, 13(2): , February N. J. Rodríguez-Fernández, P. Richaume, J. Muñoz-Sabater, P. de Rosnay, and Y. H. Kerr. SMOS Near- Real-Time Soil Moisture processor. Part1: Neural Networks evaluation and algorithm description. Technical Report SMOS Ground Segment SO-TN-CB-GS-049, CESBIO, Toulouse, France, 2015.

13 Table 3: Same as Table 2 but the statistics have been computed only with times for which both the SM-NRT-NN and the SM-L2 products are available. The total number of sites selected is 127. Table 1 shows the coordinates of all the sites. SM Mean Npts Mean R Median R mean Bias mean RMSD Mean STDD mean ACC L NRT Since the mean or median values alone do not show the full picture of the evaluation for more than 100 sites, Fig. 7 shows boxplots for the Pearson correlation coefficient R ( CC in the figure), bias, RMSD, STDD ( ubrmsd in the figure), ACC [Anomaly Correlation Coefficient] and the number of points in the time series (equal for SM-L2 and SM-NRT-NN in this case). As expected, there is a large variation from one site to another. The Bias and STDD distribution are similar for both products, while higher values of RMSD are found for some NRT time series. In contrast, the correlation is as high as almost 1 for some sites both for the SM-NRT-NN and SM-L2 (the maximum is slightly higher for the later). Interestingly, the lower values of the distribution of the correlation are higher for the NRT product. Figure 7: Boxplots for the Pearson correlation coefficient (CC), Bias (mean in situ minus mean SMOS SM), RMSD, STDD (ubrmsd), anomaly correlation coefficient (ACC) and number of points per time series (sampling) the SM-L2 (SMOS L2) and SM-NRT-NN (SMOS NN) products in comparison to in situ measurements. The box contains the middle 50% of the data, the central bar represents the median value of the distribution. The upper edge (hinge) of the box indicates the 75th percentile of the data set (q 3), and the lower hinge indicates the 25th percentile (q 1). The mean values are also shown as black crosses. The upper and lower bars represent the minimum and maximum values of the distribution excluding outliers. Points are considered as outliers if they are larger than q (q 3 q 1) or smaller than q 1 1.5(q 3 q 1).

14 Finally, Fig. 8 shows scatter plots of the correlation and the anomaly correlation coefficient for both products taking into account the respective confidence intervals. For most of the sites, both products show the same statistics with respect to the in situ measurements and globally, the scatter plot points lie close to the 1:1 line. Figure 8: Scatter plots showing the Pearson correlation coefficient (R) and correlations on anomaly time-series (ACC) for the SM-NRT-NN (SMOS NN) dataset against in situ measurements versus R and ACC for the L2-SM (SMOS L2) dataset against in situ measurements. The error bars account for the 95% confidence intervals. The red symbol represents an averaged value Figures 9 and 10 show examples of the three time series for some sites selected randomly. As expected, for some sites the two SMOS products are very different to the in situ measurements. This is most likely due to the different resolution of the remote sensing measurement with respect to an in situ point measurement. Different sensing depths can also explain the differences for some sites. It is also possible to see sites for which the L2 product seem to be closer to the in situ measurement and sites for which the NRT product is closer.

15 Figure 9: Soil moisture time series for a selection of in situ measurement sites. Yellow dots: SM-NRT-NN. Green crosses: L2. Black dot: in situ measurements.

16 Figure 10: Soil moisture time series for a selection of in situ measurement sites. Yellow dots: SM-NRT-NN. Green crosses: SM L2. Black dot: in situ measurement