AN RPC EVALUATION OF NASA REMOTE SENSING INPUTS AND MODEL DERIVED DATA FOR REGIONAL CROP YIELD PREDICTION MODELING Charles O’ Hara Preeti Mali Bijay Srestha Geo Resources Institute Mississippi State University May 17, 2006 NASA RPC PDR David Lewis Bob Ryan Institute for Technology Development and SSAI Stennis Space Center May 17, 2006 GENERAL OVERVIEW • RPC Evaluation of Soybean Yield Modeling • Regional Level Prediction • Integration of Remote Sensing • Advantages, Disadvantages and Tradeoffs • Best Possible Solution NASA RPC PDR CROP YIELD MODELS Crop (Soybean) Yield Prediction • Crop models have been used for predicting crop yield before harvest. • These pre-harvest crop yield estimations also help in regional and global crop prices and trade policies. Integration of Remote Sensing Data to Crop Yield Remote Sensing based methods have been used to provide inputs to a number of crop prediction models. NASA RPC PDR RPC: INTEGRATING BASELINE & FUTURE SENSORS DATA FOR CROP YIELD PREDICTION Sensors in Current Use Moderate Resolution Imaging Spectro-radiometer (MODIS) Advanced Very High Resolution Radiometer (AVHRR) Both have large Swath Width and High Temporal Resolution RPC Evaluation: Implement a baseline configuration of the Sinclair Model for selected soybean production areas in Brazil with current remote sensing data streams and compare results against results derived from model outputs using synthetic VIIRS and modeled LIS as data inputs. Include a well-devised ground data collection campaign, collaboration with USDA FAS for data sharing and exchange, participation of Dr. Tom Sinclair as the model owner, programmers to integrate the model, and researchers who will conduct tests and evaluations of results. NASA RPC PDR CROP MODEL SUITABILITY FOR REGIONAL YIELD PREDICTION • Regression based empirical methods • Montieth based models • Mechanistic or agro-meteorological based methods The agro-meteorological based crop yield prediction method provides a good scope in regional yield predictions using remote sensing. The variables in these methods are mostly obtained from meteorological stations, derived from remote sensing data sources, or computed by models; thus, they provide global or regional coverage and enhanced regional model applicability. NASA RPC PDR STUDY AREA: ARGENTINA NASA RPC PDR STUDY AREA DETAILS: ARGENTINA Study Area Details: MODIS 10 x 10 tiles are shown for areas to be considered. Field areas are shown from previous NASA/ ITD/USDA FAS work as well as the fields selected by Dr. Louis Salado and Dr. Tom Sinclair for field data campaign. NASA RPC PDR COMPLETED NASA RESEARCH IN STUDY AREA FOR USDA FAS Can NASA Research contribute to the foreign crop type assessment performed by the USDA Foreign Agriculture Service (FAS) Crop Assessment Estimates Crop Condition Data Retrieval and Evaluation (CADRE)? # 82 N # # 83 84 W E S Landsat Path 228 Other Provinces Cordoba 200 0 200 400 600 800 Miles NASA RPC PDR CROP TYPES IN PROJECT Crop # Samples _______________________________ Corn 140 Forest 40 Pasture (Cultivated) 100 Pasture (Natural) 100 Soybeans 150 Urban 40 Water 40 Wetland 40 Wheat 150 Sorghum <30 Peanuts <30 _________________________________ Total NASA RPC PDR ~800 = total samples to be collected FIELD DATA COLLECTION METHODOLOGY Example of ground truth equipment and digital sampling forms created for this study NASA RPC PDR DATA PRE-PROCESSING FLOW • RED • QC • • • Make QC Mask • Make Buffer Mask Make Look Angle Mask • NIR Clip to Bounds Apply Median Filter • Set Bad Pixels to -2 • Create NDVI • Set Background to 0 • • Save as ESRI Grid Apply Masks • Import to Imagine • Creation of daily NDVI datasets to be used for the hypertemporal composite NASA RPC PDR STEPS FOR GENERATION OF A MODIS-BASED NDVI Download MOD09 HDF Resample Export to GeoTIF Generate Daily NDVI Composite NDVI NASA RPC PDR HYPERTEMPORAL NDVI PLOTS FOR 4 MAIN CROPS Sep NASA RPC PDR Oct Nov 2004 Dec | Jan Feb 2005 Mar April CONCLUSIONS • Moving window compositing produced dataset for good classification results • Masks and filters applied significantly reduced anomalous and noisy pixels • The NDVI profiles of the hypertemporal dataset were separable for the corn, soybean, wheat, forest, other ag, and non-agriculture classes. • Best classification method from those tested was Minimum Distance classifier • The overall accuracy was improved using this classifier by separating the soybean class into two classes for single and double-cropped soybeans • An overall classification accuracy of 69% was achieved NASA RPC PDR FUTURE WORK • Investigate a classification system that combines the Growing Degree Days and Minimum Distance into a rules-based classifier (or decision tree classification system) in order to raise the overall accuracy achieved. • Develop a weighting rule for the data layers in a decision tree classification scheme • Use more sample sites in order to separate the pasture and other-agriculture classes • Identify sample sizes by crop distribution and acreage • To reduce noise, expanding the buffer mask to include a two and possibly three pixel buffer away from identified cloud or ‘bad’ pixels. • Refine methods for integrating results with crop yield prediction models. NASA RPC PDR SINCLAIR MODEL SINCLAIR MODEL • Semi-mechanistic model Named after Thomas Sinclair (University of Florida) • Used by USDA/FAS PECAD for regional soybean estimations Basic model inputs are based on the following relationships (Speath & Sinclair, 1987): • Leaf emergence as a function of temperature • Leaf area index as a function of leaf number and plant population • Interception of solar radiation as a function of leaf area • Biomass accumulation proportional to intercepted radiation • Seed yield proportional to biomass NASA RPC PDR INPUTS TO THE MODEL Fraction of Transpirable Soil Water Temperature Leaf Area Index Plant growth rate Temperature Plant Leaf Area as a function of Plant growth rate Precipitation LAI (Leaf Area Index) = PLA * Plant Population Soil Moisture Fraction of Intercepted Radiation based on LAI Planting Date Efficiency of solar radiation in Biomass assumption Soil Water Precipitation Seed Yield Solar Radiation Photoperiod NASA RPC PDR Incident Solar Radiation Daily Nitrogen Fixation Daily photosynthetic Biomass Production Daily Vegetative Biomass Calculate Vegetative growth Calculate Seed Growth rate based on Harvest Index Calculate Daily Nitrogen Budget for vegetative growth and Seed growth MODEL INPUTS LEAF AREA INDEX Sinclair Model simulation Planting Date Temperature DIPI (Daily Increase in Plastochron index) PLA (Plant Leaf Area) LAI = PLA * Plant Population Soil Water BASELINE SPATIAL SUBSTITUTE NOAA-AVHRR (NOAA-Advanced Very High Resolution Radiometer) MODIS (Moderate Resolution Imaging Spectro-radiometer) RPC EVALUATION NASA LIS – Temperature & Soil Moisture (NASA Model) Visible/Infrared Imager/Radiometer Suite (VIIRS) – Synthetic NASA RPC PDR BASELINE MODEL INPUTS REMOTE SENSING BASED LAI AVHRR (1km): NDVI ~ LAI relationship MODIS (250m): EVI ~ LAI relationship MODIS LAI ( MOD 15 LAI : 1km) NASA RPC PDR MODEL INPUTS METEOROLOGICAL DATA LOCAL GROUND INPUT DATA STATIONS SOURCES Data Temperature, Precipitation, Solar Radiation, Resolution Needs interpolation Temporal Hourly, Daily, Weekly cycle Coverage NASA RPC PDR Depends upon countries GOES Satellite Systems Precipitation METEOSAT, TRMM METEOSAT: Precipitation, Thermal AVHRR, MODIS Land Surface Temperature TRMM:Precipitation 4 km Daily 2.5-5km Daily METEOSAT: North and Europe/Africa/Indian South America Ocean TRMM: Tropics MODIS : 1km AVHRR LAC: 1km Daily Global MODEL INPUTS INPUT DATA SOURCES INTEGRATED DATA SOURCES NCDC ( National USAF-AGRMET NASA-LIS (Land Climatic Data (Agriculture Information System) RPC INPUT Center) Meteorology model) Source Ground Met Stations Data Temperature, Precipitation, Solar Radiation, Resolution Needs interpolation Temporal cycle Hourly, Daily, Weekly Coverage United States NASA RPC PDR High-performance land Integrated, Interpolated surface modeling and and Assimilated dataset data assimilation system Precipitation, Temperature, Soil Precipitation, Temperature, Soil Temperature, Soil Moisture, EvapoMoisture etc transpiration etc ½ degree ( ~ 40 km) 1 km 3 hourly,Daily Global Daily Global MODEL INPUTS OTHER INPUTS Day length: Calculated based on Latitude and Day of year Planting date: Important variable usually estimated from local knowledge and crop reports NASA RPC PDR MODEL INPUTS PLANTING DATE ESTIMATION Improved through remote sensing Zonal function Temporal NDVI Phenology curve Temporal NDVI cube Detect onset of greenness Develop refined estimation of crop planting date NASA RPC PDR RPC CHALLENGES * Critical RPC Items Baseline Configuration Challenges Mitigation Solutions Spatial Variability Use of sensors and products with comparable spatial resolution * Include synthetic VIIRS for RPC comparison Temporal Issues Use of temporal computational solutions such as temporal map algebra Dataset Adaptability issues: Temporal, Spatial, Geometric, Radiometric Use of integrated systems such as *NASA-LIS Model Manipulation challenges Involve the model developer into the process (Dr. Thomas Sinclair) Validation challenges Field campaign with local experts to collect critical field data NASA RPC PDR NPOESS VIIRS • In 2008, the National Polar-orbiting Operational Environmental Satellite System (NPOESS) Visible Infrared Imager Radiometer Suite (VIIRS) instrument will be launched into 1330, 1730, and 2130 local-time ascending-node sun-synchronous polar orbits. • VIIRS will replace three different currently operating sensors: – The Defense Meteorological Satellite Program (DMSP) Operational Line-scan System (OLS), – The NOAA Polar-orbiting Operational Environmental Satellite (POES) Advanced Very High Resolution Radiometer (AVHRR), and – The NASA Earth Observing System (EOS Terra and Aqua) Moderate Resolution Imaging Spectroradiometer (MODIS). NASA RPC PDR VIIRS SIMULATION • VIIRS will have a ground sample distance (GSD) ranging from 371 m by 387 m at nadir to 800 m by 800 m at the edge of the scan • Since the MODIS red-band and NIR-band reflectances have a GSD of 250 m at nadir, simulations of the types of NDVI images to be expected from the VIIRS sensor can be created from MODIS data • Temporal VIIRS simulations, such as near-daily NDVI time series plots and temporally-processed image videos, can be created using the TSPT. NASA RPC PDR Synthetic VIIRS for RPC Evaluation – Bob Ryan • MODIS data will be collected for the study area for the period from 2005 to 2007. • VIIRS data will be simulated for specific desired time intervals • IRS ResourceSat 1 AWiFS image data are in active use by the USDA FAS for crop monitoring and acreage estimation. • AWiFS image data provides an opportunity to create simulated products for comparison to actual MODIS products as well as to the synthetic VIIRS products to perform preliminary validation and uncertainty quantification of the synthetic products. NASA RPC PDR SENSOR SPECIFICATION SHEETS NASA RPC PDR Scale Issues, Synthetic Product Validation, and Uncertainty Analysis Selecting large fields as study sites with areas that include semicontinuous features enables crop characteristics to be measured by a plurality of image pixels by operational sensors. Synthetic image products with reduced spatial resolution will be produced that provide pixels that still remain within the boundaries of the selected study sites. A set of images with significantly higher spatial resolution and similar spectral characteristics will be employed to test the results of the data simulation and develop preliminary quantification of uncertainty. NASA RPC PDR PRELIMINARY VIIRS NDVI SIMULATION SHEELY FARM CROP FIELDS MODIS 250 m GSD NDVI NASA RPC PDR VIIRS 400 m GSD NDVI PRELIMINARY VIIRS NDVI SIMULATION SHEELY FARM COTTON FIELD, 2003 MODIS NDVI Time Series NASA RPC PDR VIIRS NDVI Time Series PRELIMINARY VIIRS NDVI SIMULATION SHEELY FARM GARLIC FIELD, 2003 MODIS NDVI Time Series NASA RPC PDR VIIRS NDVI Time Series VIIRS PIXEL AGGREGATION • VIIRS uses a pixel aggregation technique whereby three pixels are aggregated in-scan from nadir to a sensor zenith angle (SZA) of 31.71°, two pixels are aggregated in-scan at SZA’s from 31.71° to 47.87°, and no aggregation occurs beyond an SZA of 47.87°. • Due to this technique, although VIIRS has a larger GSD than MODIS at nadir, it has a smaller in-scan GSD at large SZA. NASA RPC PDR RESOLUTION VS SCAN ANGLE Source: Dr. Robert E Murphy, NPP Project Scientist, NASA GSFC NASA RPC PDR Synthetic VIIRS Data Product Validation IRS (Indian Remote Sensing) RESOURCESAT-1 RESOURCESAT-1 Orbit and Coverage Details RESOURCESAT-1 was launched into a sun-synchronous orbit at an altitude of 817 km following the current IRS 1C ground track. The RESOURCESAT-1 satellite was launched October 17, 2003 with a design life of 5 years. Orbits/cycle Semi-major axis Altitude Inclination Eccentricity Number of orbits/day Orbit Period Repetivity Distance between adjacent paths Distance between successive ground tracks Ground trace velocity Equatorial crosing time NASA RPC PDR 341 7195.11 817 km 98.69 degrees 0.001 14.2083 101.35 minutes 5-24 days 117.5 km 2,820 km 6.65 km/sec 10.30 ± 5 min A.M. (at descending node) Synthetic VIIRS Data Product Validation AWiFS Characteristics Advanced Wide Field Sensor (AWiFS) The Advanced Wide Field Sensor (AWiFS) with twin cameras has a 56 meter NADIR resolution with a 700 km combined swath and a five day revisit time. To cover such a wide swath,the AWiFS camera is split into two separate electro-optic modules (AWiFS-A and AWiFS-B) tilted by 11.94 degrees with respect to each other. AWiFS specifications IGFOV 56m (nadir) 70m (at field edge) Spectral Bands B2: 0.52-0.59 B3: 0.62-0.68 B4: 0.77-0.86 B5: 1.55-1.70 Swath 370 km each head 740 km (combined) Integration time 9.96 msec Quantization10 bits No. of gains1 NASA RPC PDR NASA AWiFS Characterize/Validation Activities Some additional input may be provided here by NASA about their efforts to characterize and validate calibrate reflectance products from AWiFS data sources. NASA RPC PDR CONCEPTUAL REPRESENTATION Compile MODIS data for area of interest and temporal range defined. Create synthetic VIIRS data to match the area and temporal range of the MODIS data. NASA RPC PDR Running NDVI for Daily Model Runs Desired Event Based Product for Critical Phenological Development RPC IMPLEMENTATION OF PARALLEL TEMPORAL MAP ALGEBRA FOR RAPID DATA PRODUCT DEVELOPMENT TMA is the temporal extension to conventional map algebra. Treats time series of imagery as three dimensional data set. XY plane represent Earth’s surface. Z dimension represents time. X Y Y Image n Image 1 Time X NASA RPC PDR Z TMA PARALLEL PROCESSING Bijay Srestha – MS Thesis • Global or regional coverage requires large volume of satellite data. • Need for intensive computing to integrate and process large datasets. • Parallel processing is the decomposition of a large problem into smaller problems that can be solved simultaneously to provide faster execution time. • Many spatial programs are inherently parallel. • Parallel processing can provide leap in performance. NASA RPC PDR TMA Parallel Processing Temporal Cube Block Distribution & Processing 1 Temporal Composite NASA RPC PDR 2 n …………………………………… MODIS or Synthetic VIIRS Pre-processing Surface reflectance day 1 NDVI day 1 Surface reflectance day 2 NDVI day 2 … … Surface reflectance day N NDVI day N Surf.Refl. Quality day 1 QMask day 1 Surf. Refl. Quality day 2 QMask day 2 QMask day N Geolocation Angles day 1 view zenith angle day 1 Geolocation Angles day 1 view zenith angle day 2 … … NASA RPC PDR Quality Mask cube … … Surf. Refl. Quality day N Geolocation Angles day N NDVI cube view zenith angle day N View zenith angle cube Input to Compositing Algorithm TMA Compositing NDVI Cube View Angle Cube Surface reflectance Quality Cube Model based constraints to create masked NDVI Masking Masked NDVI Cube TMA Operations NDVI Composite NASA RPC PDR Experimental Results High quality temporal composites may be efficiently created for custom products and desired temporal and geographic ranges of interest! NASA RPC PDR Implementation of parallel TMA abilities in the RPC will enable the rapid generation of custom temporal composites of real and simulated data sources and enable rapid use of desired products in evaluations! CONCLUSIONS • More research is needed for validating LAI-based inputs from remote sensing for agricultural modeling purposes. • A single sensor does not provide sufficient information to meet the needs for modeling regional agricultural systems, therefore integrated systems such as NASA-LIS are necessary to address spatial, temporal and adaptability issues. • NASA–LIS provides up to 1km resolution, enhancing compatibility with other inputs of comparable resolution. • Employing a set of synthetic VIIRS data products will enable the evaluation to consider the sensitivity of the model to the characteristics of the data streams from the future NASA sensor. • Agricultural yield prediction requires multi-temporal analysis and implementation of solutions such as temporal map algebra offers opportunity to implement robust solutions. NASA RPC PDR PDR Questions and Discussion Items RPC Experimental Design: Baseline and Future Data Assimilation Plan: Strength of RPC Team: Adequacy of Field Data Campaign and Local Knowledge Expertise: Identification of Pathway to ISS: NASA RPC PDR