THE EFFECT OF VEGETATION COVER ON NOW COVER MAPPING FROM PAIVE MICROWAVE DATA Hosni Ghedira Juan Carlos Arevalo Tarendra Lakhankar Reza Khanbilvardi NOAA-CRET, City University of New York Convent Ave at 138th t, New York, NY 10031 Peter Romanov NOAA-NEDI 1335 East-West Highway, ilver pring, MD 20910 ABTRACT now-cover parameters are being increasingly used as input to hydrological models. Having an accurate estimation of the snow cover characteristics during the snowmelt season is indispensable for an efficient hydrological modeling and for an improved snowmelt runoff forecasts. Direct measurements of snow depth at a single station are generally not very useful in making estimates of distribution over large areas, since the measured depth may be highly unrepresentative of the neighboring area because of drifting or blowing. Additionally, the traditional field sampling methods and the ground-based data collection are often very sparse, time consuming, and expensive compared to the coverage provided by remote sensing techniques. In this paper, we used an adaptive neural network system to generate the spatial distribution of snow accumulation from multi-channel M/I data in the Northern Midwest of the United tates. Five M/I channels were used in this experiment (19H, 19V, 22V, 37V, and 85V). now depth measurements have been collected from the National Climatic Data Center (NCDC) through the Cooperative Observer Network for the U.. snow Monitoring. The snow depths have been compiled and gridded into 25 km x 25 km grid to match the final M/I resolution. Different vegetation-related parameters (NDVI, Optical Depth, homogeneity) have been collected and gridded over the study area. The current results have shown a significant effect of vegetation cover properties on the mapping accuracy. Furthermore, the addition of vegetation related information to the mapping process has shown to have a positive impact on mapping performance, especially for areas with shallow snow cover (less than 5 cm). INTRODUCTION now cover is one of important and highly variable component of the Earth s weather and climate system. atellite observations of snow have been conducted for more than three decades and currently present one of the major components of the global snow cover monitoring system. Due to a high spatial resolution and frequent widearea coverage, satellites can supplement ground-based measurements and provide timely information on snow spatial distribution and temporal variations. It is important however, that currently available satellite-based snow datasets do not completely satisfy the needs of hydrological and numerical weather prediction models, which require timely reliable, high resolution (~1 km) spatially-continuous information on both the snow cover extent and its bulk physical properties, snow water equivalent and/or the snow depth. Passive microwave remote sensing techniques have been widely investigated over the last three decades and have been demonstrated to be effective for monitoring snow pack parameters such as snow water equivalent (WE), snow depth, and snow (wet/dry condition) (Walker and Goodison, 1993). Most of these researches are underpinned by the hypothesis that brightness temperature measured by the actual passive microwave sensors (e.g. M/I and AMR-E) can be linked to the physical properties of the snow cover with different correlation degrees. These properties include but are not limited to: snow fractional volume (or density), snow grain size, snow depth, and, vertical stratification by ice layers. However, the snow products derived from passive microwave sensors are usually limited by the relatively low resolution, especially when the purpose is to use the snow-related information as input for hydrological models which generally require higher resolution (4 km for the advanced hydrologic prediction system (AHP) operated by NOAA-NW).
Traditionally, snowpack physical properties were determined by using gauge measurements and snow coring for specific sampling points. However, this in situ snow measurement procedure is both expensive and ineffective for vast areas especially inhabited and inaccessible areas. In addition, direct measurements of snow depth at a single station are generally not very useful in making estimates of distribution over large areas since the measured depth may be highly unrepresentative of the surrounding areas even under the same snowfall conditions. The determination of snow depth and snow water equivalent from passive microwave measurements is based on the snow scattering property of microwave radiation defined as brightness temperature (Tb). The brightness temperature of the surface is function of various parameters: (1) the ground (humidity, surface roughness, and temperature), (2) now cover (density, depth, temperature, average grain diameter, and humidity), and (3) vegetation (temperature and stem volume). Many empirical models were developed to estimate WE, most of them assuming a linear relationship between WE and brightness temperature differences (generally 19 or 18 and 37 GHz). ome of these methods progressed to a practical level such Meteorological ervice of Canada (MC) model which was used to produce real-time WE maps over the Canadian Prairies (Walker and Goodison, 1989; Thirkettle et al., 1991). In forest environments WE retrieval becomes more complicated due to the attenuation of the ground microwave signal propagating through the canopy as well as the vegetation contribution to the brightness temperature (Goita et al., 1997 and 2003). Artificial neural network is the most widely-used non-linear and non-parametric model in the last two decades. It has been successfully applied to a wide range of non-linear problems in several disciplines. Multi-layer perceptron trained by the backpropagation algorithm is the most commonly used neural network 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 types of remotely sensed data. In contrary to traditional techniques such as regression analysis, neural network uses its complex configuration to find the best nonlinear function between the input and the output data without any constraint of linearity or pre-specified non-linearity. A useful review of the application of neural networks in remote sensing can be found in (Benediktsson et al., 1990; Paola and chowengerdt, 1995). In this paper, a neural-network-based model has been developed and has shown a great potential in identifying snow pixels from M/I data. This algorithm has been applied in the Northern Midwest tates and compared to a filtering algorithm developed by Grody and Basist (Grody and Basist, 1996). The preliminary results indicate that the neural-network-based model provides a significant improvement in snow mapping accuracy over the filtering algorithm. The effects of vegetation-related parameters on classification performance have been also investigated. TUDY AREA AND DATA ACQUIITION The study area is located in the Northern Midwest of the United tates within 110 63 W - 102 04 W and 48 71 N - 40 73 N. The selection of the study area was based on the existence of a large number of meteorological stations and the high snow accumulation rate. The passive microwave data from the pecial ensor Microwave/Imager (M/I) Level 3 EAE-Grid Brightness Temperatures was used in both ascending and descending orbits. These images provide measurements of the brightness temperature in seven channels with different frequencies and polarizations. In this project, the same five M/I channels used by Grody and Basist for the filtering algorithm have been selected to train and validate the neural network algorithm. These include: 19H, 19V, 22V, 37V, and 85V. Three non-precipitating days with the highest snow accumulation rate have been selected during the 2001/2002 winter season (01/23, 01/24, and 01/25). A total of 185 ground stations covering the study area have been identified for this experiment as well as 195 additional stations in the surrounding area. The snow depth collected from these ground-stations was linearly interpolated to a regular grid over the study area to serve as truth data. The study area contains 34 x 30 pixels with spatial resolution of 25 km. DECIION TREE A filtering algorithm for global snow cover identification was applied and evaluated over the study area. The algorithm consists of a decision tree, which establishes sensitive thresholds to filter out precipitation, cold desert and frozen surfaces (Fig 1). This filtering algorithm uses the antenna temperature retrieved from five M/I channels (19V, 19H, 22V, 37V, and 85H). More details about this technique can be found in (Grody and Basist, 1996).
cattering Materials T b22v T b85v > 0 or T b19v T b37v > 0 YE NO No Classification The scattering material represents the signatures of snow using vertically polarized antenna temperatures Precipitation T b22v =258 or 258 = T b22v =254 and T b22v T b85v =2 or T b22v =165 + 0.49 T b85v YE Precipitation Pixel NO Cold Desert T b19v - T b19h =18 and T b19v - T b37v =10 and T b37v T b85v =10 YE Cold Desert NO Frozen Ground T Tb19V - Tb19H T =8 and T b19v - T b37v =2 and T b22v T b85v =6 YE Frozen Ground NO now Cover Adapted from Grody and Basist, 1996 Figure 1. Decision tree flowchart. OPTIMIZATION OF THE NEURAL NETWORK PARAMETER A multi-layer neural network consists of a number of interconnected nodes. The nodes are organized into layers where each node transforms the inputs received from other nodes. The input layer serves as an entry for the vector of data presented to the network (M/I channels). The output layer serves to produce the neural network decision (snow or non- snow) for the pixel presented at the input layer. All layers between the input and output layers are referred to hidden layers. The best neural network architecture can only be determined experimentally for each particular problem. The number of hidden nodes should be large enough to ensure a sufficient number of degrees of freedom for the network function and small enough to minimize the problem of loss in generalization ability of the network. The training process consists in adjusting the connection weights in order to decrease the difference between the network output and the desired outputs. The training data were presented to the input layer and propagated through the hidden layers to the output layer. The differences between the neural network output and the desired outputs were computed and feed-backward to adjust the network connections. In this project, the available training data have been divided into three subsets: 1. The learning set is used for computing and updating the network weights.
2. The validation set is used to reduce the risk of overtraining by monitoring the training error. 3. The test set is an independent set used only to assess the classification accuracy and to compare between different classifiers and different network configurations. For each vector of five brightness temperatures presented in the input layer, a value equal to one will be assigned in the output layer if the presented vector corresponds to a snow pixel. Otherwise, a value equal to zero will be assigned. However, due to the asymptotic behavior of the activation function, a continuous range from zero to one was produced by the output neuron during the simulation process. This variability can be explained by the fact that the neural network could not be trained to produce a zero error on the training data, or also the data being classified can be more diversified than the data used in the training. 1.0 Overall accuracy (%). 0.8 0.6 0.4 0.2 now correctly classified as snow Non-snow correctly classified as non-snow 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Threshold Value Figure 2. Effect of threshold on classification accuracy. To transform the continuous output format into a categorical format, a threshold value between 0 and 1 have been introduced to decide if the pixel will be labeled as snow or non-snow. The optimal threshold value cannot be identified with certainty without measuring its effect on the overall accuracy of the neural network classification. In this project, the threshold value has been varied from 0.2 to 0.8. The effect of the decision threshold on classification accuracy of each class is illustrated in figure 2. 1.0 Neural Network Output. 0.8 0.6 0.4 0.2 0.0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 now Depth (inch) Figure 3. now depth vs. neural network output. The figure 2 shows that the increase of the threshold value affects the overall classification significantly. pecifically, the increase of the threshold value results in a simultaneous decrease of the percentage of correctly classified snow pixels and an increase in the percentage of correctly classified non-snow pixels. For this specific training, a threshold of 0.6 has been retained providing an overall accuracy of 82 % on the test set. The scatter plot of figure 3 shows more precisely the crucial role of the threshold selection in providing an accurate classification
and how the neural network performs better for the pixels with high snow accumulations. Indeed, for all the pixels with snow depth higher than 6 in, the neural network output was higher than 0.8. Thus, if we select a threshold equal to 0.8, all the pixels with snow accumulation higher than 6 in will be correctly classified and only 3 non-snow pixels will be misclassified. TRAINING APPROACHE The proper selection of training data is a crucial step in achieving best results. To ensure an accurate selection of training pixels, four approaches have been tested by varying the selection criteria of snow pixels. 1. In the first approach, all the pixels with one inch or more of snow accumulation were considered as snow pixels. 2. In the second approach, only the pixels with two inches or more of snow depth were considered as snow pixels. This approach reduces the risk of overestimating the ground snow depth during the interpolation (or gridding) of the snow gauge measurements. In this approach, the neural network was trained to classify the one-inch snow pixels as no snow pixels. 3. In the third approach, all the pixels with one inch of snow depth have been removed from the training process to reduce the risk of mislabeling them as snow or non-snow pixels. 4. In the fourth approach, only the pixels with ground stations inside their boundaries were used for the training. For these pixels, only those with accumulation larger than one-inch were considered as snow pixels. By comparing the four approaches, we find that the third and the fourth give the best performances by reducing the misclassification of non-snow pixels by about 10% and increasing the accuracy of correctly classified snow pixels by about 5%. COMPARION BETWEEN NEURAL NETWORK AND FILTERING TECHNIQUE A comparison between neural network technique and the filtering algorithm has been made. The same test set (third set: 200 pixels) has been used to calculate the confusion matrices for both techniques and for the three selected days. The confusion matrices presented in table 1 show that the neural network technique (trained using approach 3) provides a significant improvement (up to 30%) in snow mapping accuracy over the filtering algorithm. However, the filtering algorithm slightly outperforms the neural network by 2 % only on one day (Jan 25). Table 1. Confusion Matrices Jan 23 Jan 24 Jan 25 Decision Tree N 0.86 0.69 N 0.14 0.31 Accuracy = 43 Kappa = 9.2 N 0.70 0.48 N 0.30 0.52 Accuracy = 56 Kappa = 13 N 0.56 0.15 N 0.44 0.85 Accuracy = 80 Kappa = 38.1 ANN N 0.57 0.20 N 0.43 0.80 Accuracy = 75 Kappa = 33.3 N 0.38 0.16 N 0.62 0.84 Accuracy = 76 Kappa = 21.6 N 0.57 0.17 N 0.43 0.83 Accuracy = 78 Kappa = 35.9 The snow maps illustrated in figure 3 represent the gridded gauge measurements and the output of each technique with the same inputs for the three selected days. The decision tree maps represent the output of the filtering algorithm and the neural network maps represent the simulation results of the selected neural network (threshold = 0.6 and training with approach 3) using the same five channels used in the filtering algorithm.
THE EFFECT OF VEGETATION ON CLAIFICATION ACCURACY Different combinations of M/I data (5 channels) and vegetation-related information (NDVI + tandard deviation) have been tested in the neural network input layer. The standard deviation was calculated for each group of NDVI pixels (1 km resolution) within one M/I pixel (25 km resolution). The results of these combinations have shown that the addition of the standard deviation of NDVI to the five M/I channels gives the best accuracy by increasing the classifier performance by about 5% (figure 4). % 95 90 85 80 75 70 65 60 55 50 45 40 35 30 25 20 5Tb NDVI -NDVI tdv Highest Overall accuracy Average overall accuracy Average Kappa coefficient Figure 4. the effect of vegetation on classification accuracy. REULT AND CONCLUION The purpose of this study was to explore the ability of neural networks to improve the mapping of snow cover using M/I data. The results indicate that neural networks can be considered as better alternative to retrieve snow cover information from passive microwave satellite measurements. The performance of neural network reaches 82%, 10% higher than the accuracy obtained with the filtering algorithm for the test data sets (not used in the neural network training process). Furthermore, for shallow snow cover, misleading results have been reported when the one-inch snow pixels was considered as snow pixels. It was found that for deeper snow depth (>2 inches), the snow and non-snow pixels are more likely to be correctly classified when a threshold of 0.4 or 0.6 was used.
Ground Data Decision Tree Artificial Neural Network Jan 23 Jan 24 Jan 25 No coverage now No now Figure 5. Comparison of the gridded gauge measurement data with the decision tree and neural network outputs. ACKNOWLEDGEMENT This research was funded by the NOAA Cooperative Remote ensing cience & Technology Center (NOAA- CRET) and the City University of New York. REFERENCE Benediktsson J., P. wain, and O. K. Ersoy, (1990). Neural Network Approaches Versus tatistical Methods in Classification of Multisource Remote ensing Data, IEEE Transactions on Geoscience and Remote ensing, vol. 28-4, pp. 540-552. Goita K., A. C. Walker, B. F. Goodison, and A. 1. C. Chang, (1997). Estimation of snow water equivalent in the boreal forest using passive microwave data. in Proc. Int. ymp. Geoniatics Era of Radarsat, Ottawa, ON. Canada, May 24 30. Goita K., A. F. Walker, and B. F. Goodison, (2003). Algorithm development for the estimation of snow water equivalent in the boreal forest using passive microwave data. Int. J. Remote ensing, vol. 24, pp. 1097-1102. Grody N. and A. Basist, (1996). Global identification of snow cover using M/I measurements, IEEE Transactions on Geoscience and Remote ensing, vol. 34, pp. 237-249. Paola J., and R. A. chowengerdt, (1995). A review and analysis of backpropagation neural networks for classification of remotely-sensed multi-spectral imagery, International Journal of Remote ensing, vol. 16-16, pp. 3033-3058. Thirkettle F., A. E. Walker. 13. 13 Goodison, and D. Graham, (1991). Canadian prairies snow cover maps from realtime passive microwave data: From satellite data to user information, in Proc. 14th Can.ymp. Remote sensing, Calgary, AB. Canada, May 6 10, 1991, PP 172 176.
Walker A. and B. Goodison, (1993). Discrimination of a wet snow cover using passive microwave satellite data, Annals of Glaciology, vol. 17, pp. 307-311. Walker A.E. and B. E. Goodison, (1989). Determination of areal snow water equivalent on tile Canadian prairies using passive microwave satellite data, in Proc. IG.4R, sol. 3, Vancouver. BC, Canada. pp. 1243 1246