25 Km – St Dv. NDVI Image 25 Km – NDVI Image Snow No Cover age No Snow 5 Channels5 Channels + St Dev NDVI Jan 16 Jan 18 Jan 17 Capabilities and Limitations.

Slides:



Advertisements
Similar presentations
Applications of one-class classification
Advertisements

Page 1 of 50 Optimization of Artificial Neural Networks in Remote Sensing Data Analysis Tiegeng Ren Dept. of Natural Resource Science in URI (401)
Characterization of ATMS Bias Using GPSRO Observations Lin Lin 1,2, Fuzhong Weng 2 and Xiaolei Zou 3 1 Earth Resources Technology, Inc.
MODIS Collection 6 BRDF/Albedo: Status and Updates Zhuosen Wang 1, Crysal Schaaf 1, Miguel Román 2 1 University of Massachusetts-Boston 2 NASA Goddard.
Multiple Criteria for Evaluating Land Cover Classification Algorithms Summary of a paper by R.S. DeFries and Jonathan Cheung-Wai Chan April, 2000 Remote.
Neural Network Based Approach for Short-Term Load Forecasting
Green Vegetation Fraction (GVF) derived from the Visible Infrared Imaging Radiometer Suite (VIIRS) sensor onboard the SNPP satellite Zhangyan Jiang 1,2,
Prénom Nom Document Analysis: Artificial Neural Networks Prof. Rolf Ingold, University of Fribourg Master course, spring semester 2008.
Monitoring the Arctic and Antarctic By: Amanda Kamenitz.
Detecting SWE peak time from passive microwave data Naoki Mizukami GEOG6130 Advanced Remote Sensing.
An artificial neural networks system is used as model to estimate precipitation at 0.25° x 0.25° resolution. Two different networks are being developed,
Radial-Basis Function Networks
Ranga Rodrigo April 5, 2014 Most of the sides are from the Matlab tutorial. 1.
Accuracy Assessment. 2 Because it is not practical to test every pixel in the classification image, a representative sample of reference points in the.
A new prototype AMSR-E SWE operational algorithm M. Tedesco The City College of New York, CUNY, NYC With contributions from : Chris Derksen, Jouni Pulliainen,
Co-authors: Maryam Altaf & Intikhab Ulfat
Microwave Remote Sensing Group 1 P. Pampaloni Microwave Remote Sensing Group (MRSG) Institute of Applied Physics -CNR, Florence, Italy Microwave remote.
Experiments with the microwave emissivity model concerning the brightness temperature observation error & SSM/I evaluation Henning Wilker, MIUB Matthias.
Snow Cover: Current Capabilities, Gaps and Issues (Canadian Perspective) Anne Walker Climate Research Branch, Meteorological Service of Canada IGOS-Cryosphere.
Fractional snow cover estimation in the complex alpine-forested areas using MODIS and Landsat Elzbieta Czyzowska – Wisniewski research conducted under.
Comparison of SSM/I Sea Ice Concentration with Kompsat-1 EOC Images of the Arctic and Antarctic Hyangsun Han and Hoonyol Lee Department of Geophysics,
Applications and Limitations of Satellite Data Professor Ming-Dah Chou January 3, 2005 Department of Atmospheric Sciences National Taiwan University.
Near-real-time transition from the DMSP SSM/I to SSMIS sensor for NSIDC near-real-time snow and ice climate records Peter R. Gibbons, Walt Meier, and Donna.
Assessment of Regional Vegetation Productivity: Using NDVI Temporal Profile Metrics Background NOAA satellite AVHRR data archive NDVI temporal profile.
Recent advances in remote sensing in hydrology
Winter precipitation and snow water equivalent estimation and reconstruction for the Salt-Verde-Tonto River Basin for the Salt-Verde-Tonto River Basin.
Princeton University Development of Improved Forward Models for Retrievals of Snow Properties Eric. F. Wood, Princeton University Dennis. P. Lettenmaier,
Workbook DAAC Product Review Passive Microwave Data Sets Walt Meier.
Retrieving Snowpack Properties From Land Surface Microwave Emissivities Based on Artificial Neural Network Techniques Narges Shahroudi William Rossow NOAA-CREST.
Passive Microwave Remote Sensing
Image Classification 영상분류
Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami.
Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) Kuolin Hsu, Yang Hong, Dan Braithwaite, Xiaogang.
AMSR-E Cryosphere Science Data Product Metrics Prepared by the ESDIS SOO Metrics Team for the Cryosphere Science Data Review January 11-12, 2006.
Estimating the Spatial Distribution of Snow Water Equivalent and Snowmelt in Mountainous Watersheds of Semi-arid Regions Noah Molotch Department of Hydrology.
5. Accumulation Rate Over Antarctica The combination of the space-borne passive microwave brightness temperature dataset and the AVHRR surface temperature.
Recent increases in the growing season length at high northern latitudes Nicole Smith-Downey* James T. Randerson Harvard University UC Irvine Sassan S.
DEVELOPING HIGH RESOLUTION AOD IMAGING COMPATIBLE WITH WEATHER FORECAST MODEL OUTPUTS FOR PM2.5 ESTIMATION Daniel Vidal, Lina Cordero, Dr. Barry Gross.
NASA Snow and Ice Products NASA Remote Sensing Training Geo Latin America and Caribbean Water Cycle capacity Building Workshop Colombia, November 28-December.
Merging of microwave rainfall retrieval swaths in preparation for GPM A presentation, describing the Merging of microwave rainfall retrieval swaths in.
INNOVATIVE SOLUTIONS for a safer, better world Capability of passive microwave and SNODAS SWE estimates for hydrologic predictions in selected U.S. watersheds.
Data was collected from various instruments. AOD values come from our ground Radiometer (AERONET) The Planetary Boundary Layer (PBL) height is collected.
MSRD FA Continuous overlapping period: Comparison spatial extention: Northern Emisphere 2. METHODS GLOBAL SNOW COVER: COMPARISON OF MODELING.
AN OVERVIEW OF THE CURRENT NASA OPERATIONAL AMSR-E/AMSR2 SNOW SCIENCE TEAM ACTIVITIES M. Tedesco*, J. Jeyaratnam, M. Sartori The Cryospheric Processes.
The passive microwave sea ice products…. ….oh well…
Introduction GOES-R ABI will be the first GOES imaging instrument providing observations in both the visible and the near infrared spectral bands. Therefore.
Technical Details of Network Assessment Methodology: Concentration Estimation Uncertainty Area of Station Sampling Zone Population in Station Sampling.
AMSR-E Swath to Grid Toolkit Ken Knowles, Mary J. Brodzik, Matthew H. Savoie University of Colorado, Boulder, CO Introduction.
SeaWiFS Views Equatorial Pacific Waves Gene Feldman NASA Goddard Space Flight Center, Lab. For Hydrospheric Processes, This.
Snow and Ice Products from the Aqua, Terra, and ICESat Satellites at the National Snow and Ice Data Center Introduction Sensors on the NASA Earth Observing.
Downscaling Global Climate Model Forecasts by Using Neural Networks Mark Bailey, Becca Latto, Dr. Nabin Malakar, Dr. Barry Gross, Pedro Placido The City.
SeaWiFS Highlights July 2002 SeaWiFS Celebrates 5th Anniversary with the Fourth Global Reprocessing The SeaWiFS Project has just completed the reprocessing.
Assessment on Phytoplankton Quantity in Coastal Area by Using Remote Sensing Data RI Songgun Marine Environment Monitoring and Forecasting Division State.
The Derivation of Snow-Cover "Normals" Over the Canadian Prairies from Passive Microwave Satellite Imagery Joseph M. Piwowar Laura E. Chasmer Waterloo.
Over 30% of Earth’s land surface has seasonal snow. On average, 60% of Northern Hemisphere has snow cover in midwinter. About 10% of Earth’s land surface.
CPH Dr. Charnigo Chap. 11 Notes Figure 11.2 provides a diagram which shows, at a glance, what a neural network does. Inputs X 1, X 2,.., X P are.
Gofamodimo Mashame*,a, Felicia Akinyemia
Temporal Classification and Change Detection
HIERARCHICAL CLASSIFICATION OF DIFFERENT CROPS USING
Passive Microwave Systems & Products
NSIDC DAAC UWG Meeting August 9-10 Boulder, CO
Evaluating Land-Use Classification Methodology Using Landsat Imagery
REMOTE SENSING Multispectral Image Classification
REMOTE SENSING Multispectral Image Classification
Kostas M. Andreadis1, Dennis P. Lettenmaier1
network of simple neuron-like computing elements
Igor Appel Alexander Kokhanovsky
Neural Networks II Chen Gao Virginia Tech ECE-5424G / CS-5824
Neural Networks II Chen Gao Virginia Tech ECE-5424G / CS-5824
Random Neural Network Texture Model
Presentation transcript:

25 Km – St Dv. NDVI Image 25 Km – NDVI Image Snow No Cover age No Snow 5 Channels5 Channels + St Dev NDVI Jan 16 Jan 18 Jan 17 Capabilities and Limitations of Neural Networks in Snow Cover Mapping from Passive Microwave Data 1. Introduction Snow-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. Passive microwave remote sensing techniques have been investigated by numerous researchers using various sensors and have been demonstrated to be effective for monitoring snow pack parameters such as spatial and temporal distribution, snow water equivalent (SWE), depth, and snow condition (wet/dry state). However, the snow products derived from passive microwave sensors are usually limited by the relatively low resolution, especially when the purpose is to use this product as input for hydrological models. Moreover, the accuracy of passive-microwave-based maps is usually affected by the presence of vegetation. In this project, we used an adaptive neural network system to generate the spatial distribution of snow accumulation from multi-channel SSM/I data in the Northern Midwest of the United States. Five SSM/I channels were used in this experiment (19H, 19V, 22V, 37V, and 85V). The normalized difference vegetation index (NDVI) is used from the AVHRR to quantify the vegetation dynamic in the snow mapping process. Six snow days with high snow accumulation have been selected during the 2001/2002 winter season to train and test the neural network system. The snow depths and NDVI values have been compiled and gridded into 25 km x 25 km grid to match the final SSM/I resolution. To ensure an accurate selection of training pixels, different approaches have been tested by varying the selection criteria of snow pixels. The final results have shown the importance of these selection criterions on the neural network performance. Juan C. Arevalo ( NOAA/CREST Graduate student) Hosni Ghedira (Assistant Professor) Reza Khanbilvardi (Professor) 2. Study Area, Data Acquisition and Remote Sensing Data The City College of the City University of New York. Convent Avenue at 140th St, Steinman Hall, New York, NY SSM/I Images Ground data distribution Jan23 19HJan24 19HJan25 19H Input layer2 hidden layersOutput layer Network Architecture : [5:10:10:1] 19 H 19 V 22 V 37 V 85 V Six days have been selected during the 2001/2002 winter season (01/16, 01/17, 01/18, 01/23, 01/24, and 01/25). The following illustrations show the channel 19H of SSM/I satellite with 25 km resolution for the six selected days. A total of 195 ground stations covering the study area have been identified for this experiment. The figure below shows the distribution of the ground stations over the study area (red rectangle). 4. Artificial Neural Network 5. Neural Network Approaches 8. Snow Cover Maps Original gridded data in Northern Hemisphere projection with coordinates: 119˚44’W - 99˚57’W and 49˚36’W - 34˚34’W The study area is located in the Northern Midwest of the United States within 110˚37’48’’W - 102˚02’24’’W and 48˚42’36’’N - 40˚43’48’’N.. The passive microwave data from the NOAA/NASA Pathfinder Program Special Sensor Microwave/Imager (SSM/I) Level 3 Equal Area Scalable Earth-Grid (EASE-Grid) Brightness Temperatures F13 satellite is used in both ascending and descending orbits. These images provide measurements of the brightness temperature in seven channels with different frequencies and polarizations (19 V, 19 H, 22 V, 37V, 37 H, 85 V, and 85 H). The available training data, has been divided into three subsets: The 190 training pixels have been set as follow: Learning set 90 pixels Validation set 45 pixels Test set 50 pixels The first one is the learning set, whish is used for computing and updating the network weights. The second subset is the validation set, which is used for stopping the training by monitoring the validation error during the training process. The third subset is the test set that is not used during the training process, and it is only used to assess the classification accuracy and to compare between different classifiers and different network configurations. The following color images represent two snow-cover maps for each selected day generated from the artificial neural network output. Each map contains 34 X 30 pixels with spatial resolution of 25 km. 6. Neural Network Results Highest Overall accuracy Average overall accuracy and standard deviation Average Kappa coefficient and its standard deviation AVHRR Image The AVHRR data was acquired from the Distributed Active Archive Center (DAAC) located at Goddard Space Flight Center, NASA. Jan16 19HJan17 19HJan18 19H 5TbNDVIS-NDVI Threshold 0.4 StDv Threshold 6 5TbNDVIS-NDVIStDv Run % Approach 4. Five channels,Tb + standard deviation of NDVI. Threshold 0.6 Jan 23 Jan 24 Jan 25 Ground snow map 3. Data Acquisition: Vegetation The input layer size may vary depending on the approach used; from 5 to 7 neurons. Threshold % Graph shows the pattern and difference from the truth data; snow/no-snow and the ANN result 01 Threshold ANN Truth data Actual snow depth from the truth data and the ANN result Snow - Snow Non snow – Non snow Overall accuracy Kappa coefficient Effect of the decision threshold on classification Truth data – snow depth [in] Threshold ANN During the simulation process, a continuous range from zero to one will be produced by the output neuron. We have introduced a threshold value (between 0 and 1) to decide if the pixel will be classified as snow or no- snow pixel. 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 the following figure: The original 8 Km NDVI values have been gridded into a regular grid of 25 Km over the study area. The standard deviation values have been measured during the gridding process to quantify the vegetation homogeneity for each pixel. 7. Confusion matrices 5 Channels + NDVI5 Channels 5 Channels+St Dev. NDVI 5 Channels+NDVI+St Dev. NDVI As a part of the assessment of the capabilities and limitations of neural networks for snow mapping, the neural network output has been evaluated with a confusion matrix that was computed for each approach. The overall accuracy and Kappa coefficient were measured. The following matrices correspond to the net giving the highest accuracy out of 100 runs. Source of Data: Ground data from National Oceanic and Atmospheric Administration, NOAA SSM/I data from DMSP SSM/I Pathfinder daily EASE-Grid brightness temperatures, January Boulder, CO: National Snow and Ice Data Center, NSIDC (Armstrong, R.L., K.W. Knowles, M.J. Brodzik and M.A. Hardman. 1994, updated current year). The Normalized Difference Vegetation Index (NDVI) has been derived from the visible and the near-infrared channels of NOAA-AVHRR Sensor over our study area. 8-Km, 10-days composite NDVI image, January Only pixels with ground stations inside their boundaries are considered for the training and validation of Neural Network. A total of 165 pixels satisfy this criterion. Results obtained for the four approaches (number of input channels): 1. the 5 brightness temperature (Tb) channels; 2. the 5 Tb plus NDVI; 3. the 5 Tb plus NDVI and standard deviation, and 4. the 5 Tb plus the standard deviation of NDVI. A total of 100 runs of the neural network for each approach have been performed. The following graphs show for each approach the average accuracy and the Kappa coefficient for the 100-runs and their corresponding standard deviation. These results have shown that the addition of the NDVI standard deviation (homogeneity factor) improves the snow identification accuracy. The graph below shows the accuracy variation of 100 neural network trained with different initial configurations. Threshold 0.6 for the approach 4, which yields the net with the highest accuracy. That net was used to simulate the corresponding snow maps. The overall accuracy varies between 60 and 80 %, having a fairly stable pattern. However, some high and low peaks have been observed; 86 % and 52% being the highest and lowest accuracy. The Kappa coefficient, which assess the agreement in the classification between the snow and non-snow pixels, has a very irregular pattern. For each vector of five brightness temperatures presented to the input layer, a value equal to one will be assigned in the output layer if the presented vector correspond to a snow pixel. Otherwise, a value equal to zero will be assigned to the corresponding vector. The 5 channels maps (first column) represent the simulation generated by simulating the trained neural network in the approach 1, (threshold = 0.4). The other maps (2 nd column) represent the simulation results of neural network trained with the approach 4 by using 5 SSM/I channels plus the standard deviation of the NDVI as input. For this configuration, the best performance was obtained with a threshold equal to 0.6.