Classification and retrieval of dry snow parameters

0 downloads 0 Views 247KB Size Report
[3] A.T.C. Chang and J.L. Foster and D.K. Hall,Nimbus 7 SM derived global snow cover patterns,Annals of Glaciology,Vol. 9, 1987,pp 39-44. [4] M. Tedesco, “Dry ...
Classification and retrieval of dry snow parameters by means of SMM/I data and Artificial Neural Networks M. Tedesco, P. Pampaloni

J. Pulliainen, M. Hallikainen

IFAC - CNR Firenze – Italy [email protected]

Lab. of Space Technology – HUT Espoo - Finland [email protected]

Abstract— Dry snow temperature, snow water equivalent (SWE) and snow depth have been retrieved by using the 19 and 37 GHz SSM/I brightness temperatures and Artificial Neural Networks (ANN’s). The results obtained have been compared with those obtained using other approaches such as the Spectral Polarization Difference, the HUT Model-based iterative inversion, the Chang algorithm and linear regressions. In general, it has been noted that the ANN based technique gives better results than the other approaches, which tend to underestimate the unknown parameters. Keywords: dry snow, passive microwave remote sensing, artificial neural networks

I.

INTRODUCTION

Mapping and retrieval of snow depth, snow water equivalent and temperature are the major objectives of passive remote sensing of snow. In general, the unknown parameters can be obtained by performing the inversion of theoretical equations relating snow parameters to the corresponding brightness temperatures, or by employing empirical algorithms based on experimental data. In the first case, non-linear integro-differential equations are involved, and the inversion problem may be ill-posed. In the second case, the retrieval algorithm can be valid only locally, and a huge amount of data may be necessary. In this paper, the retrieval of dry snow temperature (DST), snow water equivalent (SWE) and dry snow depth (SD) has been performed by using the 19 and 37 GHz SSM/I brightness temperatures and supervised artificial neural networks (ANN’s). The neural network training sets were generated by using either simulated or measured brightness temperatures. Simulations were obtained by using the HUT Snow Emission Model. Once trained, the ANN’s were interrogated by employing the SSM/I brightness temperatures as inputs. The results were then compared with those obtained using other approaches, such as the Spectral Polarization Difference (SPD) [1], the HUT Model-based iterative inversion [2], the Chang algorithm [3], and linear regressions [4]. The detection of dry snow and the classification of different snow types were performed by using unsupervised neural networks. The latter were set up in order to classify the input patterns into two different classes, receiving as input only the 19 and 37 GHz SSM/I brightness temperatures. The

This work was supported in part by the CIMO (Centre for International Mobility), Laboratory of Space Technology HUT, and in part by the EU project ENVISNOW EVG2-2001-00018

classification of different types of snow areas was also achieved by using a competitive neural network, receiving as input a combination of the four SSM/I measured Tb(19 and 37 GHz, V and H pol.). II.

SSM/I AND GROUND TRUTH DATA

The radiometric data set consisted of four brightness temperatures (19 and 37 GHz, V and H polarizations) from the SSM/I Pathfinder Daily EASE-Grid (Equal Area SSM/I Earth Grid) supplied by the National Snow and Ice Data Centre (NSIDC) [5]. The snow water equivalent (SWE) data set was based on national operational snow observations from the Finnish Environment Institute (FEI), while daily maximum and minimum air temperatures were obtained by the Finnish Meteorological Institute (FMI) from weather stations located less than 50 km from the test sites. The latter, with a forestcover fraction ranging from 78.5 % to 94.5%, were distributed over all of Finland. To guarantee dry snow conditions, we selected only those data collected after three successive days with a daily maximum temperature of below -5 °C. III.

RETRIEVAL OF DRY SNOW PARAMETERS

The retrieval of snow parameters was performed by inverting the brightness temperatures by means of supervised artificial neural networks (ANN), and specifically multi-layer perceptrons. The ANNs were trained either with data generated by the HUT snow emission model or with ground measured data. A. Retrieval of SWE and dry snow depth Artificial Neural Networks trained with the HUT snow emission model have shown a good accuracy in retrieving the SWE and the SD. Figure 1. shows a comparison of the measured SWE with the one retrieved using a single ANN trained over all the test sites. Figure 2. shows the measured and retrieved SD as a function of time. It can be seen that the retrieval performances were very satisfactory for SWE values up to 150-200 mm (or for data collected until the beginning of springtime), but were bad for data collected when SWE was

Figure 1. Measured versus retrieved SWE using a single ANN trained over all the test sites with the HUT model

In the case of ANN’s trained with experimental data, the networks were trained with SWE measured in some specified locations, and the retrieval was performed for those areas not included in the training set. Several networks were trained, by using data measured in northern, southern and, in some cases, central Finland. The ANN’s trained with northern and southern data were denoted with ANN N/S, those ones with only northern and southern, respectively, ANN N and ANN S; and those containing data also from the central Finland with ANN N/C/S. In all these cases, the retrieval of SWE was achieved with high accuracy, which pointed out the considerable ability of this technique to retrieve unknown parameters. The best performances were obtained by training the net with southern data only. B. Retrieval of dry snow temperature In the case of ANN trained with the HUT model, the accuracy in the retrieval of snow temperature was rather modest because of the assumptions that snow, vegetation, air and soil had the same temperature and that data collected in the descending orbit of the satellite corresponded to the minimum daily temperature. Moreover, also in this case, the stratification effects could not be taken into account. The results obtained suggested that the retrieval of snow temperature with the HUT model and ANN’s could be improved by considering more detailed relationships between air and snow temperatures.

Figure 2. Snow depth as a function of time (cont. line = measured, crosses = retrieved with net trained with the HUT model)

The retrieval of dry snow temperature was also performed by training the ANN with temporal sequences of data (19961998) and interrogating it for another sequence of SSM/I data (1998-99). In this case, performances showed good accuracy. Figure 3. shows the measured versus the retrieved dry snow temperatures obtained with the ANN trained with experimental data for the 1998-1999 period. IV.

 Figure 3. Measured vs retrieved dry snow temperatures obtained with the ANN, trained with experimental data for the 1998-1999 period

higher than the cited values. Indeed, since the HUT model was a single-layer model, it was not able to take into account the stratification effects that significantly influence the radiometric responses. We also compared the performances of a single network trained over all the test sites with those obtained by training a different network for each single test site (using the same training set with a different validation set). The best retrieval was obtained for networks trained locally. However, the generalized network gave good results as well.

DETECTION AND CLASSIFICATION OF DRY SNOW

Separationof dry snow from bare soil or wet snow was performed by using the SSM/I brightness temperatures only (19 and 37 GHz channels) as input to an unsupervised ANN, which was set up to classify the input patterns into two different classes. The basic idea was that the two classes corresponding to the two conditions of dry-snow presence or absence (wet snow or soil) could be separated because of volume scattering effects introduced by the dry snow layers. Obtained results showed that the dry snow had been clearly detected. Figure 4. shows the result for the northern test site located near Sodankylä. In the figure, the SWE has been assumed as indicator of the presence of snow. Note that for high SWE values (snow strongly stratified) or very low values (low snow depth and density), the two classes were mixed, as expected, since the volume scattering effects were weak for low snow depth and affected by the stratification during the melting cycles.

TABLE I - COMPARISON OF RESULTS Applied Technique SWE

R2

RMSE

R

SPD (all)

32.27 mm

57.32 %

HUT iter. Inv.

30.72 mm

38.54 %

0.957

ANN and HUT

24.1 mm

44.89 %

0.85

ANN and exp.

19.53 mm

80.44 %

0.938

SPD

18.34 cm

16.31 %

0.558

ANN and HUT

14.64 cm

47.78 %

0.854

Chang

18.97 cm

12.38 %

0.502

Linear

5.68 C

40.88 %

0.908

ANN and HUT

6.66 C

18.44 %

0.860

ANN and exp

4.82 C

57.35 %

0.954

Snow Depth

Figure 4. SWE as a function of time for the test site at Sodankylä: circles for dry snow, crosses for wet snow or no snow

Again, a competitive neural network was trained to classify different types of snow from a combination of the four SSM/I (19 V,H and 37V, GHz) brightness temperatures. We used a combination of brightness temperatures, since both local snow/air temperature and SWE had to be considered in the analysis. Indeed, the separation of the different regions may have been due not only to the differences in snow type, but also to the differences in other parameters of the observed scene (such as snow depth). For this reason, the input vector employed for the classification was made of polarization indexes (PI) at 19- and 37 GHz, and of the spectral polarization difference (SPD) index. In this case, only data collected in January were employed. The distinction between the different snow type areas (north/south, internal/costal areas) was performed successfully. The results showed that the northern test site data were classified with an accuracy (number of correct cases/number of total cases) of 98.3 % ,while those of southern areas had an accuracy of 90.86 %. The distinction between internal and costal areas was performed by using data of test sites 1 and 2 for the first type, and of sites 11 and 12 for the second one. In this case, the accuracy was 96.55 % for the internal area and 85.71 % for the costal area.

V.

Temperature

As for the ability of the ANN’s to interpolate spatial data, the best results were obtained by training the ANN’s with data collected on southern sites. The retrieval of snow temperature by using both linear regression or the ANN trained with the model was affected by the assumptions that snow, vegetation, air and soil had the same temperatures and that minimum daily temperature was reached in the descending orbit. The best results were obtained for the 1998-99 period with the ANN trained by using ground data obtained in 1996-97. Detection of dry snow covered area was performed with good accuracy using unsupervised artificial neural networks. ACKNOWLEDGMENT The authors wish to thank Matias Takala for processing and furnishing the data. REFERENCES

CONCLUSIONS

The retrieval performances obtained by using ANN’s were compared with the SPD algorithm, the iterative inversion of HUT snow emission model, and linear regressions. In general, the ANN based techniques gave better results than the other approaches. Table I shows a comparison of the performances for the different methods applied. The SPD algorithm and the iterative inversion of the HUT model provided good results for SWE values up to about 150 mm. The formula proposed by Chang for the retrieval of snow depth also gave best results by taking into account only measurements performed until mid February for relatively low values of SWE. Neural networks trained either with the HUT model or experimental data showed the best accuracy for both parameters.

[1] [2]

[3] [4]

[5]

J. Aschbacher, Land surface studies and atmospheric effects by satellite microwave ra-diometry, PhD thesis , University of Innsbruck, 1989. J. T. Pulliainen and J. Grandell and M. Hallikainen, “HUT snow emission model and its applicability to snow water equivalent retrieval”,IEEE Trans. on Geosc. and Rem. Sens.,May 1999,vol. 37,No. 3,1378-1390 A.T.C. Chang and J.L. Foster and D.K. Hall,Nimbus 7 SM derived global snow cover patterns,Annals of Glaciology,Vol. 9, 1987,pp 39-44 M. Tedesco, “Dry snow mapping in Finland employing space-borne passive microwave observations and artificial neural networks”, Helsinki University of Technology , Laboratory of Space Technology, Report No. 50, Espoo, September 2002 J. Maslanik and HJ. Stroeve, DMSP SSM/I daily polar gridded brightness temperatures,National and Snow Ice Data Center CD-ROM, Boulder, CO, USA