MIT Lincoln Laboratory MIS IGARSS11-1 RVL 9/15/2010 An Atmospheric Algorithm Suite Based on Neural Networks for Microwave Imager/Sounders W. J. Blackwell,

MIT Lincoln Laboratory

MIS IGARSS11-1RVL 9/15/2010

An Atmospheric Algorithm Suite Based on Neural Networks for Microwave Imager/Sounders

W. J. Blackwell, R. Czerwinski, R. V. Leslie, M. Pieper, P. Rosenkranz*, J. Samra, D. H. Staelin*, C. Surrussavadee ‡, K.

Wallenstein, & D. Zhang

MIT Lincoln Laboratory* Research Laboratory of Electronics at MIT

‡ Prince of Songkla University

IGARSS 2011: Vancouver, Canada

28 July 2011

This work was sponsored by the National Oceanic and Atmospheric Administration under contract FA8721-05-C-0002. Opinions, interpretations, conclusions, and recommendations are those of the author and are not necessarily endorsed by the United States Government.



Outline

• Overview

• Physics

• Retrieval Approach– Neural Networks– Radiative Transfer

• Training Datasets

• Expected performance

• Summary



Atmosphere EDR Suite

• Atmospheric Vertical Temperature Profile (AVTP) – Kelvin– Lower Atmospheric Sounding (Surface to 10 mb)– Upper Atmospheric Sounding (10 mb to ~0.01 mb)

• Atmospheric Vertical Moisture Profile (AVMP) – MMR g/kg• Atmospheric Pressure Profile (APP) – millibar• Total Water Content (TWC) - kg/m2 or mm in a 3-km vertical segment

• Total Integrated Water Vapor (TIWV) - kg/m2 or mm (a.k.a., precipitable water)

• Precipitation Rate/Type (PRT) – mm/hr and types: rain or ice• Cloud Liquid Water Content (CLWC) – kg/m2 or mm• Cloud Ice Water Path (CIWP) - kg/m2 or mm

Profile Subset

2-D Field Subset



Algorithm Simulation Methodology



MIS Atmospheric Algorithm Methodology

• Cloud/precipitation products derived from cloud-resolving NWP models combined with multi-stream scattering models

– Global NWP runs over ~5M pixels– Multi-phase microphysical modeling

• Profile products derived from global high-resolution analysis fields

– Performance validated over many years (millions of pixels) for similar AMSU/AIRS algorithm

– Framework allows for optimization of product spatial resolution

• Neural network estimators offer accuracy/robustness/speed– Very easy to code (large infrastructure currently available)– Very easy to upgrade (simply replace coefficient file)– Very low computational burden – can run on mobile terminals

Physical Models + Stochastic Processing



PHYSICS AND

PHENOMENOLOGY



Microwave Scattering and Absorption

Atmospheric Transmission

Hydrometeor Mie Scattering and Absorption

Liquid water

Ice

Frequency [GHz]

Frequency [GHz]

Frequency [GHz]



Passive Microwave Sensing of Precipitation

35 km

45 k

m



Overview of SSMIS Channel Setand Spatial Resolutions

V = vertical pol.H = horizontal pol.R = right-hand circ.* subset in precipitation algorithm

km



SSMIS UAS Channel Characteristics



Temperature and Water VaporWeighting Functions

Temperature Water Vapor

45° off-nadir angle



Upper Air TemperatureWeighting Functions

26 uT 90 deg. (tropical)

65 uT 53 deg. (polar)



MULTILAYER FEEDFORWARD

NEURAL NETWORKS



Neural NetworksNonlinear, Parameterized Function Approximators



Example: Temperature Profile RetrievalAdvantages Relative to Linear Regression (LLSE)



Advantages Relative to Linear RegressionBetter Noise Immunity and Physical Representation

Noise contribution: Component of retrieval error due only to sensor noiseAtmosphere contribution: Retrieval error in the absence of sensor noise



RADIATIVE TRANSFER AND

SIMULATION METHODOLOGY



Algorithm Simulation Methodology



Radiative Transfer / NWP Interface Issues

Each level requires hydrometeor densityper drop radius

MM5

Pre

ssur

e [m

b]

Mas

s D

ensi

ty [

g/m

3 ]

Radius [mm]

Mass Density [g/m3]

graupel

snow

rain

10 mb

Sekhon-Srivastava

Marshall-Palmer

Image courtesy of Colorado State University

SSMIS(NGES)



PERFORMANCE

VERIFICATION DATASETS



Geographical locations of the pixels in the MM5 and NOAA88b data sets



Mean and Standard Deviation of NOAA/MM5 Data Sets

Temperature Water Vapor



MM5 Cloudy Data Set



PERFORMANCE

VERIFICATION



Precipitation Rate Retrieval Performance



Summary of Cloud Water/Ice Retrieval Performance



AVTP Retrieval PerformanceCloudy (40 km)

MM5 not valid at these high altitudes



Upper Air Sounding Performance

• SSMIS UAS channels (CH20-24)

• No Doppler effects

• IGRF-11 geomagnetic model

• Multi-layer Feedforward Neural Network

• NOAA88b dataset

• SSMIS Spec:– 7-1 mb: 5 K– 0.4 mb: 5.5 K– 0.2-0.03 mb: 8 K



AVMP Retrieval PerformanceCloudy (40 km)

• SSMIS: Greater of 1.5 g/kg or 20%

• IORDII: • 10% objective• Greater of 0.2 g/kg or 20%

(surf. to 600 mb)



Clear-Air Atmospheric Pressure Profile Performance (40 km)

APP derived using AVTP and AVMP retrievals and surface pressure (assumed perfect)Quality-controlled global radiosondes used for ground truth

Land Ocean



Summary

• Comprehensive, end-to-end performance assessment capability in place for all products in the Atmosphere EDR Suite

– Minimal retrieval optimization performed at this point– Clear path to requirement compliance for all products

• Flexible, modular algorithm architecture easily accommodates changes to sensor characteristics and performance



Backup Slides



Simulated SSMIS Pass Over CONUS

• 50.3-GHz brightness temperature

• 40-km Spatial resolution

• 2/3 CONUS HRRR – 3 km

• CCA antenna pattern



SSMIS and AMSU Precipitation Rate Retrievals

8



Structure of the SSMIS Precipitation Algorithm

Pixel Longitude/Latitude Brightness Temperatures

Bias correction Interpolate to fine retrieval grid

Surface classification

PCA Transform

Channel Selection Channel Selection

Spatial Perturbations

Specialized Neural Network

Surface-Classification-Dependent Weighting

Retrieved Precipitation Parameters

Channel Selection



Radiance Simulation Methodology

• CRM = MM5 1-km saved every 15 min• RTM = multiple-stream radiative transfer

solution (TBSCAT† or TBSOI*) • Simulated NAST-M radiances• Developed and adapted MIT software to

LLGrid parallel computing facility

MM5 grid levels

Cloud Resolving Model (CRM)

Radiative Transfer Model (RTM)

Simulated Radiances

SP

AT

IAL

FIL

TE

RIN

G

“Satellite Geometry”Toolbox (MATLAB)

* Successive Order of Interaction: Heidinger A. K., et al., J. Appl. Meteor. Climatol., 2006† TBSCAT: Rosenkranz, P. W., IEEE Trans. Geosci. Remote Sens. 2002



Histogram of Surface Pressures for the Synoptic Radiosonde Data Set



Geographical Locations of the Pixels in the Synoptic Radiosonde Data Sets

~200,000 quality-controlled radiosondes from 2009-2010 representing all seasons



Precipitation Rate Performance



Precipitation Type Retrieval



Precipitation Rate Performance Stratified by Precipitation Type



Cloud Water/Ice Retrieval Performance



Total Integrated Water Vapor Performance (25 km)



AVMP Retrieval PerformanceClear-air (40 km)

• Black = Ocean• Green = Land• Blue = Global

• SSMIS: Greater of 1.5 g/kg or 20%

• IORDII: • 10% objective• Greater of 0.2 g/kg or 20%

(surf. to 600 mb)



Total Water Content Performance

Altitude Ocean Land Global Spec. (IORDII)

surface 1.20 2.00 1.44 2.0 kg/m2

5 km 0.80 1.40 1.10 2.0 kg/m2

7.5 km 0.45 0.48 0.46 2.0 kg/m2

10 km 0.10 0.12 0.11 2.0 kg/m2

• 3-km “slabs”• 25 km resolution• cloudy MM5 dataset



Limitations and Degradation

• Precipitation– Effects all atmos. EDRs except PRT– Nominally, atmos. EDRs will be retrieved under 1 mm/hr– Difficult to quantify 1 mm/hr, will use status flags to classify the

precipitation (e.g., “no precip.”, “stratiform”, “light convective”)– Status flags must determine if a CFOV has even one precipitation-

impacted EFOV

• Land emissivity– Properly classifying land conditions (e.g., flooded or snow-covered) will

make stratifications (i.e., a condition specific NN) more difficult to implement

– Difficult to obtain a statistically-adequate sample set

• Land elevation– Difficult to obtain a statistically significant sample set to train on– Must evaluate whether training many altitude stratifications is worth the

effort and cost

Documents

MIT Lincoln Laboratory MIS IGARSS11-1 RVL 9/15/2010 An Atmospheric Algorithm Suite Based on Neural Networks for Microwave Imager/Sounders W. J. Blackwell,