Beijing June05 HE

Slide1: 

Herman Eerens & Dong Qinghan Vlaamse Instelling voor Technologisch Onderzoek (VITO) Mol, Belgium The Use of Satellite Data for Crop State Monitoring and Yield Forecasting CONTENTS INTRODUCTION IMG PRE-PROCESSING: Raw → 10-Daily Syntheses (S10) IMG POST-PROCESSING: S10 → Value-Added Products CROP AREAS CROP YIELDS CONCLUSIONS

Slide2: 

INTRODUCTION VITO VITO: Vlaamse Instelling voor Technologisch Onderzoek Flemish Institute for Technological Research Research Institute of Flemish Government (Flanders=Northern Region of Belgium) 500 staff members 8 “Centres of Expertise” dealing with new materials, environmental issues, … TAP: Centrum voor Teledetectie & Aardobservatie-Processen Centre for Remote Sensing Applications 60 staff members 5 Units: SPOT-VEGETATION processing unit Hyperspectral (APEX) Unmanned Aerial Vehicles Long-term Land Use Changes AGRO-UNIT: Remote sensing for large-scale agricultural applications 8 staff members

Slide3: 

CGMS Cali-bration INTRODUCTION REMOTE SENSING in CGMS

Slide4: 

INTRODUCTION REGIONS OF INTEREST (ROI’s)

Slide5: 

INTRODUCTION SENSORS

Slide6: 

The Use of Satellite Data for Crop State Monitoring and Yield Forecasting CONTENTS INTRODUCTION IMG PRE-PROCESSING: Raw → 10-Daily Syntheses (S10) IMG POST-PROCESSING: S10 → Value-Added Products CROP AREAS CROP YIELDS CONCLUSIONS Herman Eerens & Dong Qinghan Vlaamse Instelling voor Technologisch Onderzoek (VITO) Mol, Belgium

Slide7: 

PRE-PROCESSING GENERAL SCHEME (HiRes + LoRes) Compositing = Select per pixel “best available” measurement

Slide8: 

PRE-PROCESSING SPOT-VGT

Slide9: 

PRE-PROCESSING SPOT-VGT Standard Products Global VGT-S10: - Geographic Lon/Lat at 1°/112 resolution ( 1 km) - 11 “layers” (Refl., NDVI, angles, status map, registration time) - Available after 2-4 days  Purchased for all projects/ROI’s

Slide10: 

Problem: 1 VGT-S10 = 10 Gb  1 year = 360 Gb  10 years (‘98-’07) = 3.6 Tb Solution: Conversion to Pseudo-Image format (PI) Radiometric compression (e.g. Reflectances from 16-bit to 8-bit) Elimination of sea-pixels (no information: all values are 0) Result: Data reduction to 22% of original level  10 years = 792 Gb (feasible) All info retained: Global, 1 km, 11 layers (spectral, angles, SM, TG) Same as ZIP, but PI's are stored as normal images (ENVI) All existing pixel-based procedures directly applicable on the PI's ! NORMAL IMAGES PSEUDO-IMAGES Reconversion possible + Selection of specific ROI + Change of Map-projection A. GTOPO30-DTM B. VGT-S10, 98/5/21, NIR PRE-PROCESSING: Data Compression of Global VGT-S10

Slide11: 

Reconvert all Global PI’s to Normal Images with Resolution of 21km x21km Systematic selection of central pixel of 21 x 21 window  Original “signatures” ! Data Set Size: - IMG-Size: 1920 x 698 = 1.3 Mb pixels - 10 year x 36 dek. x 11 layer = 3960 IMG’s - Total: 5 Gb instead of 3.6 Tb (1km-resol.) Excellent Data Set: - for global vegetation monitoring - to study behaviour of SPOT-VGT - to test new procedures on global scale - to train global classification algorithms PRE-PROCESSING: Global VGT-S10 at 21 Km-Resolution Animation of NDVI April 98 → March 99

Slide12: 

PRE-PROCESSING NOAA-AVHRR over EUROPE SpacePC = AVHRR processing chain developed by EU-JRC INPUT = Raw AVHRR-registrations (SHARP1, Level1B, GAC) + TBUS OUTPUT = Fully processed Daily Syntheses (S1) over Europe (all bands) = More cloudfree, 10-Daily MVC Composites (S10) MARSOP = AVHRR-processing since 2000 (Antenna Berlin) = Re-processing of JRC-archive (1989-2000)

Slide13: 

PRE-PROCESSING S10 for EUROPE (MARSOP2) NOAA-AVHRR & SPOT-VGT: After Unification Same IMG-format, Contents (as far as possible – see Table) & Radiometric Scaling Map System: ETRS89-Lambert Azimuthal Equal Area (INSPIRE), 1 km resolution NOAA AVHRR SPOT VGT

Slide14: 

The Use of Satellite Data for Crop State Monitoring and Yield Forecasting CONTENTS INTRODUCTION IMG PRE-PROCESSING: Raw → 10-Daily Syntheses (S10) IMG POST-PROCESSING: S10 → Value-Added Products CROP AREAS CROP YIELDS CONCLUSIONS Herman Eerens & Dong Qinghan Vlaamse Instelling voor Technologisch Onderzoek (VITO) Mol, Belgium

Slide15: 

1. 10-daily Vegetation State Indicators: LST, NDVI, DMP,… 2. Within-Year Time Domain Operations: - Within Year: Phenology - Multi-annual: Comparison with Long-Term Trends 3. QuickLook Maps (for all above-mentioned Images): - Reduced resolution, colour maps - Added: Legend, grid, vectors,… POST-PROCESSING NEW IMAGES DERIVED FROM S10

Slide16: 

MONTEITH-Approach: DMP = R . 0.48 . fAPAR .  (Tmin/Tmax) Symbol Meaning Units Source DMP Dry Matter Productivity kgDM/ha/day R Incoming solar radiation (0.2-3.0µm) J/ha/day Meteo 0.48 Fraction of PAR (0.4-0.7µm) in R - fAPAR PAR-fraction absorbed by Vegetation - Remote Sensing Tmin/Tmax Daily min/max temperature °C Meteo  Efficiency ≈ Autotrophic respiration: kgDM/J - conversion of absorbed PAR-Energy to carbohydrates - maintenance respiration POST-PROCESSING DRY MATTER PRODUCTIVITY (DMP)

Slide17: 

NDVI = (NIR-RED)/(NIR+RED) SAVI = (1+L).(NIR-RED)/(NIR+RED+L) with L=0.5 About fAPAR: fAPAR is most relevant vegetation state variable at 1km-level (better than LAI, fCover,…) Correlation with fAPAR: better for SAVI than for NDVI (less influence from soil background) IMAGE PROCESSING NDVI, SAVI & fAPAR

Slide18: 

SAVI End of March 2004

Slide19: 

Meteo-France: Daily, Global Meteo-Data, interpolated to Grid of 1.5°-resolution 1 July 2000: Mean Daily Temperature [°C] 1 July 2000: Solar Irradiation [MJ/m²/day] POST-PROCESSING DRY MATTER PRODUCTIVITY (DMP)

Slide20: 

Free Download (Global, S10, DMP): http://www.geosuccess.net POST-PROCESSING DRY MATTER PRODUCTIVITY (DMP)

Slide21: 

IMG Archive: X = NDVI, SAVI, FAPAR, DMP,… Np = 36 dekads, 12 months TIME DOMAIN TEMPORAL MONITORING Within-Year Operations  Short-term Phenology Amongst-Year/Multi-annual Operations  Detection/Delineation of Calamities and Changes

Slide22: 

POST-PROCESSING WITHIN-YEAR TIME-OPERATIONS 1. Monthly Composites: Three S10 → S30 (less clouds, less temporal detail) 3. Phenological Indicators 2. Cumulative values over specific periods in Growing Season

Slide23: 

IMG Archive: X = NDVI, SAVI, FAPAR, DMP,… Np = 36 dekads, 12 months TIME DOMAIN TEMPORAL MONITORING Within-Year Operations  Short-term Phenology Amongst-Year/Multi-annual Operations  Detection/Delineation of Calamities and Changes

Slide24: 

IMG Archive: X=NDVI,DMP,… Np=36 dekads 12 months Historical IMGs: Long-Term Statistics (+ Deciles) Difference IMGs with regard to Historical Year POST-PROCESSING MULTI-ANNUAL TIME-OPERATIONS

Slide25: 

INPUTS: for variable X in period p of year y: X(y,p) = Actual Values Minx(p), x(p), x(p), …= Long-term statistics (Historical Year) OPERATORS for computation of Dx(y,p) = Difference w.r.t. “Historical Year”: Absolute Difference: ADVIx(y, p) = X(y,p) - x(p) Relative Difference: RDVIx(y, p) = [X(y,p) - x(p)]/ x(p) Standardized Diff.: SDVIx(y, p) = [X(y,p) - x(p)]/ x(p) Relative Range: RRVIx(y, p) = [X(y,p) - Minx(p)]/[Maxx(y,p) - Minx(p)] Historical Probab.: HPVIx(y,p) = Cumul. Probability of X(y,p) REMARKS: Strong correlation between all Operators If X=NDVI  RRVIx(y,p) =Vegetation Condition Index VCI (Kogan, Viau,…) If X=NDVI  HPVIx(y,p) =Vegetation Productivity Index VPI (Sannier et al.) Probability of occurence of X(y,p) in historical context Applied for NDVI in Southern Africa  Good correlation with Yields Probably best “historical” approach POST-PROCESSING MULTI-ANNUAL TIME-OPERATIONS

Slide26: 

Vegetation Productivity Index VPI - or HPVIx(y,p): Example for a given pixel in a fixed period p Cumulative histogram derived from Ny historical values Approximate cumulative histogram (P0, P10,…, P100 = Deciles): -Store in separ. IMGs -P0 =Minx(p) -P50 =Medianx(p) -P100=Maxx(p) Specific value X(y,p)  Find corresponding Historical Cumul. Probability  Store in HPVI-IMG HPVI-Interpretation: < 50%: bad, below normal 50%: normal > 50%: good, above normal POST-PROCESSING MULTI-ANNUAL TIME-OPERATIONS

Slide27: 

POST-PROCESSING MULTI-ANNUAL OPERATIONS Long Term Average for AUGUST months (1998-2004) August 2003 compared to Long Term Average

Slide28: 

POST-PROCESSING MULTI-ANNUAL OPERATIONS Example: VGT-S30, August 2003 Actual NDVI VPI Objective of IMG-Differencing: Quick Delineation of Good/Bad Zones Crop state monitoring

Slide29: 

The Use of Satellite Data for Crop State Monitoring and Yield Forecasting CONTENTS INTRODUCTION IMG PRE-PROCESSING: Raw → 10-Daily Syntheses (S10) IMG POST-PROCESSING: S10 → Value-Added Products CROP AREA CROP YIELDS CONCLUSIONS Herman Eerens & Dong Qinghan Vlaamse Instelling voor Technologisch Onderzoek (VITO) Mol, Belgium

Slide30: 

MAPPING & AREAS: Biome maps from LR-imagery

Slide31: 

MAPPING & AREAS: Biome maps from LR-imagery

Slide32: 

LAND USE MAP OF SANJIANG PLAIN (2002) from Multitemporal Landsat-TM/ETM+ (30m resol.) RUM-DATABASES AREA FRACTION IMAGES

Slide33: 

RUM-DATABASES NEURAL NETWORKS IN/OUT = LoRes (1km) values per pixel Direct Application  Sub-pixel Classification (derive AFI’s) I/O: IN: pixel image values (e.g. 36 dekadal NDVI) OUT: pixel area fractions of C classes (AFI’s) Calibration I/O data from HiRes Test Areas Application I/O results for complete LoRes area Inverse Application  Unmixing (vector p with pure class signatures) Calibration: IN: pixel area fractions of C classes (AFI’s) OUT: pixel Image values (e.g. 36 dekadal NDVI) Application: IN: area fraction 100% for class c, 0% for other classes OUT: pure signature of class c

Slide34: 

Broadleaf Forest Wetland Soybean Rice INPUTS: - Monthly NDVI of VGT, 2002 - AFI’s from HiRes ETM classif. over Sanjiang Plain (Calibration) OUTPUTS: - AFI’s for entire Province (15 clas) HEILONGJIANG GOAL: Extrapolation of expensive HiRes Land Use Data over large areas with cheap LoRes Imagery RUM-DATABASES AFI’s from LoRes IMAGERY (Soft Class.)

Slide35: 

The Use of Satellite Data for Crop State Monitoring and Yield Forecasting Herman Eerens & Dong Qinghan Vlaamse Instelling voor Technologisch Onderzoek (VITO) Mol, Belgium CONTENTS INTRODUCTION IMG PRE-PROCESSING: Raw → 10-Daily Syntheses (S10) IMG POST-PROCESSING: S10 → Value-Added Products CROP AREA CROP YIELDS CONCLUSIONS

Slide36: 

CROP YIELDS RUM = Regional Unmixed Means PROBLEM = Incompatibility OFFICIAL STATISTICS + CGMS  IMAGES

Slide37: 

CROP YIELDS EUROPE (MARS-STAT) AREA FRACTION IMAGE Cropland area % (from CORINE) EXTRACT IMAGE NDVI per Decade (AVHRR/VGT) C-NDVI = Weighted NDVI-mean REGIONS IMAGE 596 EU-NUTS2 regions UNMIXING: Make 1km²-Pixel Signal specific per Region and per Crop (or Group of Crops) C-Indicators: Weighted Regional Means with AFI-Thresholds (EU-CGMS) For 10 relevant CORINE classes For 5 thresholds (0, 20, 40, 60, 80%) Stored in ORACLE-DB (Alterra) Waiting for new CLC2000

Slide38: 

CROP YIELDS LINEAR UNMIXING MODEL: For a "homogeneous" region and a given Period (dekad/month): N = Mixed 1km²-pixels in region (pixels: n = 1…N) C = Pure classes / crops in region (classes: c = 1…C) m= (N x 1) vector: measured signal of N mixed pixels p = (C x 1) vector: pure signals of C classes F = (N x C) matrix: area fractions (fn,c=fraction of pixel n covered with class c) APPLICATION: Spectral Unmixing via Matrix Inversion Goal: Find Vector p (pure signal of each Class) Inputs - Vector m  SPOT-VGT / NOAA-AVHRR S10/S30 - Matrix F  Area Fraction Images (AFI’s) Output - Vector p  Database with pure class signals per Region When repeated for every Dekad => Pure Profiles per Region/Class p = (FT * F)-1 * FT * m

Slide39: 

AREA FRACTION IMAGE Wheat (from IACS) REGIONS 26 Circumscriptions EXTRACT-IMAGES NDVI / DMP per Decade Simple Thresholding: Only use pixels with CROP AREA > 50% Linear Unmixing CROP YIELDS BELGIUM (B-CGMS)

Slide40: 

Sanjiang Plain (± 10 000 km²) VGT-S10, NDVI, Year 2002 RESULT: Pure signatures for 15 classes Logic/Phenologic evolutions Better than C-Indicators Linear Unmixing (AFI’s from HiRes Classification) CROP YIELDS HEILONGJIANG (H-CGMS)

Slide41: 

CROP YIELDS RUM-DATABASES 1. Extraction In: Single RS-IMG, regions raster, hard classification / AFIs of C-classes Out: RUM-file (ASCII-TXT) 2. Storage of all RUM-data in ORACLE-Database (1 per ROI) STATIC AUXILIARY TABLES: Regions Classes Sensors Variables Unmix Methods RUM_HEADERS: All combinations of “labels” RUM10_VALUES: S10-data (regional means per class) RUM10_VALUES: S30-data RUMnn_LTA: Historical means

Slide42: 

WHY EXTRACTION OF RUM-DATABASES? 1. Bridge between RS-Images and Agro-Statistics (+CGMS) RUM-values are specific per Region and Crop/Class 2. Important data reduction: IMG: profiles per pixel DB: profiles per Region x Class Suppose: N = 1000 (Average Nr. of pixels per region) C = 10 (Nr. of classes) Size-Ratio DB / IMG = 0.01 3. Further analysis can be done with GIS/Database Software CROP YIELDS RUM-DATABASES

Slide43: 

RUM-DATABASES WHY? RUM-Viewers

Slide44: 

Winter Wheat in Belgium Calibration with Neural Network Y = f(X1, X2,…Xn) Xi = 5 Monthly DMP (Unmixed for Cropland!) 200 data points (25 regions x 8 years) Only Remote Sensing! Per Crop: Definition of Optimal Set of Yield Indicators (X-variables) CROP YIELD FORECASTING: CALIBRATION & VALIDATION CROP YIELDS CALIBRATION & VALIDATION

Slide45: 

Imagery of SPOT-VGT and AVHRR does contain useful information on actual state and final yield of crops, in spite of low resolution and mixed signal Qualitative Crop Monitoring  Inspection of QuickLook maps Quantitative Yield Forecasting  Image-derived indicators mostly improve the statistical calibration Example for Winter Wheat in Belgium (B-CGMS). CAL: 1995 to 1997, VAL:1998 CROP YIELDS CALIBRATION & VALIDATION

Slide46: 

CONCLUSIONS OUTPUTS & DATA DISSEMINATION MARS-STAT: EUROPE MARS-Bulletins + Website MARS-FOOD IGAD, CIS, MERCOSUR, MEDITERRA Bulletins for Somalia http://b-cgms.crag.wallonie.be Belgium/Heilongjiang Website + bulletins

Slide47: 

The Use of Satellite Data for Crop State Monitoring and Yield Forecasting CONTENTS INTRODUCTION IMG PRE-PROCESSING: Raw → 10-Daily Syntheses (S10) IMG POST-PROCESSING: S10 → Value-Added Products CROP AREAS CROP YIELDS CONCLUSIONS Herman Eerens & Dong Qinghan Vlaamse Instelling voor Technologisch Onderzoek (VITO) Mol, Belgium

Slide48: 

CONCLUSIONS

Slide49: 

PRICE Free Low Medium Expensive Very Expensive 5 DMC Current EO-Systems: Quasi-perfect Linearity: Spatial resolution  Synoptic View High spatial detail (small pixels)  Limited field of view (low synoptic power)  Low temporal frequency  Higher Cost SPOT-VGT CONCLUSIONS NEW SENSORS 1 DMC MSG GLOBAL COVER Daily Weekly Monthly Yearly 10 years MODIS-250m LANDSAT

Slide50: 

METEOSAT SECOND GENERATION (MSG) Geostationary orbit, permanent sight on Africa/Europe SEVIRI: - Spectral: 11 bands (VIS →TIR) - Spatial: 3km sub-nadir (de facto 4-20 km over Europe) - Temporal: every 15 minutes Free, value-added products from Land-SAF (Portugal) MEDIUM-RESOLUTION SENSORS TERRA-MODIS: 2-daily, global coverage at 250m in 2 bands (RED, NIR) ENVISAT-MERIS: 15-daily, global coverage at 300m in 15 bands (VIS → NIR) DISASTER MONITORING CONSORTIUM (DMC) 5 Micro-satellites, launched in 2002/3, Near-Polar orbit Developed by SSTL-Surrey (UK), Funds: UK, Algeria, Nigeria, Turkey, China Sensor: - Spectral: Green, Red, NIR - Spatial: 32 m resolution ( Landsat-TM), Swath Width = 620km ! Synchronised orbits  Daily global coverage at 32m resolution (revolution!) CONCLUSIONS NEW SENSORS GENERAL: Funding should place more focus on operational EO-systems: High repetitivity  Large swath width (HiRes: Landsat >> SPOT) ENVISAT-MERIS 300m: 15-days is too long for crop monitoring What with: Failure of Landsat-7, future of SPOT-VGT programme?

Slide51: 

The Use of Satellite Data for Crop State Monitoring and Yield Forecasting CONTENTS INTRODUCTION IMG PRE-PROCESSING: Raw → 10-Daily Syntheses (S10) IMG POST-PROCESSING: S10 → Value-Added Products CROP AREAS CROP YIELDS CONCLUSIONS THANKS FOR YOUR ATTENTION Herman Eerens & Dong Qinghan Vlaamse Instelling voor Technologisch Onderzoek (VITO) Mol, Belgium