Article Integrating Vegetation Indices Models and Phenological Classification with Composite SAR and Optical Data for Cereal Yield Estimation in Finland (Part I)

During 1996–2006 the Ministry of Agriculture and Forestry in Finland, MTT Agrifood Research Finland and the Finnish Geodetic Institute carried out a joint remote sensing satellite research project. It evaluated the applicability of composite multispectral SAR and optical satellite data for cereal yield estimations in the annual crop inventory program. Three Vegetation Indices models (VGI, Infrared polynomial, NDVI and Composite multispetral SAR and NDVI) were validated to estimate cereal yield levels using solely optical and SAR satellite data (Composite Minimum Dataset). The average R2 for cereal yield (yb) was 0.627. The averaged composite SAR modeled grain yield level was 3,750 kg/ha (RMSE = 10.3%, 387 kg/ha) for high latitude spring cereals (4,018 kg/ha for spring wheat, 4,037 kg/ha for barley and 3,151 kg/ha for oats).


Introduction
Remote sensing has been extensively applied in world crop production estimations by the European Union, the United States of America and the Food and Agriculture Organization (FAO) of the United Nations.The advantage of applying satellite based remote sensing data is global coverage and equally calibrated data, which enables temporal and spatial comparison over the years in monitoring areas.
New precision farming techniques in agriculture can be used more accurately to assess crop growth and soil conditions during growing seasons [1][2][3][4][5].New high accuracy GPS satellites used for position control, e.g., the new EU-funded GALILEO system [6], and new hyperspectral spectrometers operating over a wide range of wavelengths (λ = 400-2,400 nm) provide new possibilities and tools for cereal cultivation [1,[7][8][9].Major climatic events, such as global drought periods during the forthcoming climate change can be monitored with new generation satellites of the enhanced spectral resolution [10,11] .The Joint Research Centre of EU has applied JRC-FAPAR index (Fraction of Absorbed Photosynthetically Active Radiation) for monitoring large area vegetation stress states and biomass productivity during Pan-European drought periods [12][13][14][15][16][17].In addition, JRC-FAPAR model with Radiative Transfer models (RT) and using SeaWiFS data from the OrbView-2 (AKA SeaStar) satellite has bee applied to identify key vegetation phenology events (start, end, length) during growing season.
Recently Harrison et al. [18] used the AFRCWHEAT2 crop model for scaling-up wheat phenological development in Europe.Moulin et al. [19] applied SPOT satellite data to validate the AFRCWHEAT2 and SAIL reflectance models to estimate wheat yield and biomass NPP (Net Primary Production).
In northern latitudes the cloudiness during the growing season seriously reduces the applicability of optical spectrum satellite data .However, new generation microwave based satellite systems (e.g., COSMO-SkyMed/Italy, ENVISAT/ESA/EU, ALOS/PALSAR, Radarsat-2 /Canada, TerraSAR-X/ Germany can map images through cloud coverage and during the night-time.The Japanese ADEOS II (Miradori II) and GOSAT/IBUKI (Greenhouse gases Observing SATellite) satellites can also measure atmospheric aerosol columns, vertical greenhouse gas profiles (CO 2 , CH 4, NO 2 , N 2 O , O 3, CFC-11, CFC-12) and water vapour .Global greenhouse gas estimates can be used in large area climate change studies to assess changes in biomass NPP production and cereal yield production [19,20].Recently multispectral remote sensing studies have focused on integrating both optical and microwave SAR data [1,21].Recently McNairn et al. [1] integrated multispectral classification techniques with SAR and optical data in Canada.Cereal, grass and oil crop fields in four prairie testing sites were classified with Neural Network and Decision trees methodologies.
In Finland Kuittinen [22], Kuittinen et al. [23] and Karvonen et al. [24] have reviewed the yield estimation of spring wheat and barley (Hordeum vulgare L.) combined with Landsat and SPOT data for Nordic high latitude agricultural regions.Yara Co. (formerly Kemira GrowHow Finland) [25] has established in Finland a commercial Kemira Loris™ (LOcal Resource Information System) integrated expert system for farmers.Loris utilizes GPS based precision farming techniques with optimum fertilization and cultivation practises combined with infrared aerial photographs.
During 1996-2006 the Ministry of Agriculture and Forestry in Finland (MAFF), MTT Agrifood Research Finland (MTT), the Finnish Geodetic Institute (FGI) and the Technical Research Centre of Finland (VTT, Dept. of Space Research) performed a joint remote sensing satellite research project combined with ground truth measurements to evaluate the applicability of multispectral SAR and optical satellite data for cereal yield estimations in the annual crop inventory program.Both optical and composite SAR microwave Vegetation Indices models were validated to estimate spring cereal yield levels using solely optical and SAR satellite data (Composite Minimum Dataset).
In summary, new composite multispectral optical and SAR classification and modeling techniques provide new integrated tools for cereal yield estimations in national inventory programs [1,17,[26][27][28].In Finland the Ministry of Agriculture and Forestry performs the annual ground based crop inventory sampling to estimate cereal yield production and cultivation areas on national level.New classification and modeling techniques integrated with digitized Finnish Land Parcel Identification System (FLPIS) and Integrated Administration and Control System (IACS) can be used to support Finnish national crop inventory programs [29,30].More accurate cereal inventory estimates will also support Finnish national CAP-policy (Common Agricultural Policy) goals and objectives on EU scale.These aspects are reviewed in Part II of this publication.
The aim of the present study was to estimate actual non-potential grain yield levels for high latitude spring cereals (spring wheat, barley and oats, Avena Sativa L.) in large area field conditions in southern Finland.The cereal theoretical maximum yielding capacity is limited by environmental and vegetation stresses (e.g., drought periods, nutrient deficiencies, pathogen epidemics) during growing season in actual field growing conditions.These stress factors result to reduced non-potential baseline yield levels (y b , kg/ha) on field parcel level.
The objectives of the present study were: (i) to construct a dynamic SatPhenClass phenological classification model, which classifies both optical and SAR satellite data based on cereal actual phenological development in both vegetative and generative phases (ii) to calibrate and validate multispectral Composite Vegetation Indices (VGI) models, which integrate both phenologically preclassified optical (Models I-II) and microwave SAR data (Composite SAR and NDVI Model III), and finally (iii) VGI models were used to estimate cereal non-potential baseline yield (y b ) levels in growing zones (I-IV) in southern Finland during 1996-2006.

System Analysis and Strategy
The research applied in this study focused on utilizing multispectral optical and microwave SAR data extracted from experimental sites in southern Finland.A dynamic SatPhenClass phenological algorithm was developed to classify measured satellite data based on phenological development of different spring cereals.The pre-classified SAR and optical satellite data were used as input for VGI models.The system analysis [31,32] of this study is depicted in Figure 1.[31,38].
The analysis strategy applied in this study is depicted in phases I-XI.Phases I-II (Section 2.2) illustrate satellite and ground truth data pool consisting of solely SAR and optical satellite data (Composite SAR andOptical Minimum Datasets 1996-2006).Correspondingly phase III (Section 2.4) explains geometric and radiometric calibration of SAR and optical data.Phases IV-V (Sections 2.3.1, 3.2.1)illustrate SatPhenClass phenological classification model and LAI and ETS ground truth sampling methodology.Phases VI-VII (Sections 2.3.2,3.2.2) explain calibration of VGI models using Minimum Datasets without extensive agrometeorological ground truth data required by conventional dynamic crop models [33][34][35][36].Phases VIII-IX (Sections 2.5, 3.In addition, Phase XI(c) scheme presents an alternative cereal specific LAI-bridge coupling system [23] reviewed in Part II.LAI-bridge coupling system utilizes optical satellite data as input data for dynamic crop models.Hodges [33] and Bowen [36] have reviewed the principles of dynamic crop models explaining phenological development and crop physiological processes (root and soil systems, photosynthesis, respiration and translocation of assimilates to heads).Dynamic crop models require extensive agrometeorological and phenological ground truth data for daily integration of biomass and yield accumulation [34,35].
In LAI-bridge coupling system, the GEMI-index (Global Environment Monitoring Index, [37]) is applied to estimate cereal LAI development (LAI satellite ) from optical reflectance data during growing season.A similar model, VGI GEMI (model IV) is used in this study in Part II.The LAI-bridge coupling method applies iterative Kalman filter (KF) algorithm [38] commonly applied in System Control theory, Markov chains and Bayesian estimations.The GEMI based LAI estimates (LAI Satellite ) and LAI values estimated by crop models (LAI CropModel ) are corrected in Kalman filter iterations with LAI Satellite and LAI CropModel variances to obtain optimized LAI ( Final ) values for non-potential baseline yield (y b ) calculations.LAI-bridge coupling method was originally developed by Karvonen et al. [24] using SPOT and Landsat data as input for PotCropF [24] dynamic crop model used in Finnish cereal yield estimations.Later on, Kuittinen et al. [23] applied LAI-bridge coupling system with GEMI index by using SPOT and NOAA optical data as input for CropWatN [39] and WOFOST [40] crop models.
(d) Estimating non-potential baseline grain yield levels (y b ) for different spring cereals in actual field growing conditions in Southern Finland with calibrated VGI models and optical and SAR data.

Overview of Satellite and Ground Truth Sites
A general overview of satellite and ground truth experimental sites is displayed in Figure 2. A detailed description of the experimental sites is given by Kuittinen et al. [23].Table 1 depicts the microwave SAR and optical data sources with publication division (I-II), growing zones and Agriculture Advisory Centres in Finland.A general scheme of the measuring satellite systems and calibration parameters used in this study is displayed in Tables 12 and 13 (Appendix C).The abbreviations applied in this study are given in Table 7 (Appendix B).
The satellite measurement and ground truth experiment sites (Lapua, Kirkkonummi , Jokioinen , Mellilä and Porvoo) with corresponding soil classifications [44] are depicted in Table 1 and Table 9 (Appendix B).The Lapua experimental site, consisting of Lapua , Ilmajoki and Seinäjoki experimental sites (Figure 2), was located near the Gulf of Bothnia on sandy clay type soils.Respectively Porvoo and Kirkkonummi experimental sites were located close to the Baltic Sea with humid marine climatic conditions.Jokioinen and Mellilä sites were located on inland plateaus with mainly clay type soils.

Phenological classification algorithm (SatPhenClass) for satellite data
The calibrated optical reflectance and microwave SAR backscattering data (Tables 11-13, Appendix B.) measured from different satellite systems was classified with a phenological classification algorithm (SatPhenClass) for spring cereals.The detailed SatPhenClass classification model and VGI models applied in Parts I and II (detailed optical) are depicted in Figure 3a,b.The SatPhenClass classified satellite data were used in calibration of the VGI yield models (Figure 1). Figure 3a depicts the simplified 'Black Box' diagram and Figure 3b depicts corresponding components on system analysis level.The pseudocode for the SatPhenClass-algorithm is presented in Appendix D. The radiometric and geometric corrected surface reflectance (rf) and backscattering (σ 0 ) data were used as input data for the dynamic SatPhenClass classification model using Julian date (DOY) as a driving variable.Figure 3a,b depict SatPhenClass-model diagram for spring cereal (wheat, barley and oats) classification using optical and microwave satellite data.
Table 2. SatPhenClass Effective classifiers: Effective temperature sum (ETS) and LAI max values for high latitude cereals on growing zones I-IV [23,45].
In class a p (Figure 3b, phase VII), when the BBCH value is below 12, the cereal plant is in vegetative phase between germination, two-leaf and double ridge stages.
In class b p (Figure 3b, phase VIII), the BBCH value is between 12 and 50.When BBCH value is between 12 and 40, the cereal leaf development and stem elongation occur.The booting of the uppermost flag leaf and the initialization of heading occur, when BBCH value is between 40 and 50.In addition, the Leaf Area Index maximum is usually observed in class b p .
In class c p (Figure 3b, phase IX), the BBCH value is between 50 and 90.The cereal inflorescence emergence occurs, when BBCH values are between 50 and 60.When the BBCH value is above 60, the cereal transition to generative grain filling phase begins with the initialization of anthesis and flowering.When the BBCH value is between 70 and 80, the development of grains in the head occurs with early milk, hard dough and fully ripened phases.Finally in class d p (Figure 3b, phase X), when the BBCH value is above 90, the cereal is in senescence phase.On the average, in Finnish long day growing conditions, the phenological class a p corresponds for May, class b p for June, class c p for July and d p for August and September.
The SatPhenClass phenological classification model enabled the temporal synchronization of different optical and microwave satellite data along with spring cereal phenological development.The classified optical and microwave SAR data can in turn be used as input for cereal VGI (Vegetation Indices) or dynamic crops models, like CERES-Wheat model [33][34][35][36]48].In CERES-Wheat phenological submodel, the genotype x cultivar variation is taken into account by using genetic coefficients controlling cultivar photoperiodism, vernalization and phyllochron development.With SatPhenClass submodel, the cultivar and genotype level effects for phenological development were assumed to be constant.The genetic coefficients for high latitude Nordic spring cereals have been previously calibrated and assessed by Laurila [35] and later reviewed by Saarikko [49].In Sweden Åfors et al. [50] have reviewed the use phenological growth scales for cereals grown in high latitude Nordic growing conditions.Recently Peltonen-Sainio et al. [51] revised the Feekes and Zadoks cereal growth scaling [43,47] for Finnish long day growing conditions.

LAI and ETS Ground Truth Sampling Methodology for the SatPhenClass Algorithm
During growing season crop phenological observations (Julian Day of Year, DOY), LAI measurements (Leaf Area Index), biomass, growing density and other ground truth parameters were recorded simultaneously with satellite measurement dates (JDay SAT , Table 9, Appendix B) in the experimental sites (Table 1).Kuittinen et al. [23] presented a detailed description for the ground truth sampling methodology.Portable Li-Cor 2000 Plant Canopy Analyzer [52] was used to measure spring cereal LAI values with corresponding Julian dates (JDay LAI ) during the growing period.
The sampling methodology on experimental sites was applied after Kuittinen et al. [23] and is displayed in Figure 6 (Appendix A).Sampling plots were randomly selected from the field parcels in experimental areas (Table 1).Sampling plots were adjusted to Landsat pixel size (30 × 30 m).In each plot inside selected parcels, the LAI and biomass were measured using 10 points.The distance between the points varied between 20 and 30 m depending on the size of the parcel.The LAI value of each point was the mean of 9 LAI measurements around the original point integrated automatically by the Li-Cor 2000 Plant Canopy Analyzer.The average plant biomass on the field parcel was estimated by selecting randomly 20 plant samples.Late on, the dry-weights of the dried plant samples were measured.The growing density was determined by calculating the number of plants from the randomly selected 1 meter long sowing row with two replicates.After the harvest in the autumn, the final cereal grain yield levels (kg/ha corrected to 15% moisture content) were estimated by the farmers in the experimental areas.Yield samples were measured from the granary silos.
The Finnish Meteorological Institute provided Cumulative Effective Temperature Sum (ETS, T b = 5 °C) measurements with corresponding Julian dates (JDay ETS , Figure 4b).Kriging interpolation method was applied to estimate ETS values in experimental sites from the nearest weather stations (Figure 4c).The cumulative ETS(T b = 5 °C) and measured LAI values were used as categorical classifiers in the SatPhenClass algorithm [23,24].The Julian dates of the measured satellite data (JDay Sat ) were compared with the estimated Julian dates for spring cereals using the SatPhenClass algorithm (Figure 3b, sections II-VI: LAI development and ETS accumulation).The LAI max -values for high latitude spring cereals (Table 2) with corresponding Julian dates (JDay LAI , Figure 3b, section V) were used to estimate the transition date from generative (class b p ) to grain filling phase (class c p ).The LAI max for spring cereals varied between 3.44 and 4.27 in b p phase and between 2.14 and 3.87 in c p phase.Table 2 depicts the cereal ETS temperature sum requirements in Finnish growing conditions for different phenological classes (a p -d p ) with corresponding threshold temperatures (T b ) and typical Julian dates (JDay ETS , Figure 3b, section IV).In growing zones I-IV, the ETS requirements for a spring cereals [53,54] vary over years and locations (Figure 4a).In Nordic high latitude long day growing conditions the growing season is generally defined as the period when mean air temperature exceeds +5 °C (T b 5 °C, Table 2 [45]).The average thermal time requirement of 1050 ± 30° degreedays (dd, T b 5 °C) from sowing to yellow ripening stage is considered adequate for spring cereal ideotypes grown in growing zones I-IV [23,55,56] In the final stage the SatPhenClass algorithm (Figure 3b, sections IV-V: ETS and LAI Classifiers) compares the ETS and LAI classification results.The classification result is accepted only, if both submodels reach concord (Figure 3b, section VI: ETS and LAI Evaluation).The measured optical or microwave satellite data with corresponding Julian date (JDay Sat ) is finally set to corresponding cereal phenological class a p -d p shown in Figure 3b (graphical plot in left lower corner).The phenologically classified optical and SAR satellite data can be used as input for optical VGI infrared and NDVI models (Table 3, Figure 3b, XI: Output VGI I-II ) and for Composite NDVI and SAR model (Figure 3b, XII: Output VGI III ).Table 3. Composite Vegetation Indices Models (I-III) with dependent baseline yield (y b ) and independent variables [23,[57][58][59] (1), (2) .

Composite Vegetation Indices (VGI) models I-III
Three different Composite Vegetation Indices (VGI) models (I-III, Table 3) were calibrated with linear and non-linear polynomial regression models (Equations 1-2, Appendix C).Calibrated optical reflectance data were used for Models I-II (Models I-Infrared, II-NDVI) and composite SAR and optical data for the Model III (Composite NDVI and SAR).The NDVI optical and SAR satellite data were pre-classified with SatPhenClass model into the vegetative pre-anthesis phenological classes (a p , b p ) and into the generative post-anthesis and senescence classes (c p , d p , Figure 3b, phases IV-VI).
The VGI models (I-III) were calibrated with SAS © linear stepwise REG/Stepwise (STW) regression algorithm (Equation 1, Appendix C, [57,58]) and with non-linear polynomial RSREG regression algorithm (Equation 2, Appendix C) by using SAR and optical satellite data (Table 12, Appendix C).The non-linear RSREG algorithm takes into account (i) the general linear component, (ii) the polynomial quadratic response of non-linear variation (iii) the cross product effect of dependent variables and (iv) the total model component (Table 5).The statistical significance levels are given in Table 8 (Appendix B).
The polynomial infrared model (I) is a polynomial linear regression model, which estimates baseline yield production (y b , kg/ha) with infra red (rf 3 ) and near infrared (rf 4 ) channels.Respectively in VGI model II the NDVI Index [57][58][59] and in Model III the composite SAR and optical NDVI components were applied.
The composite SAR and optical reflectance datasets were used for spring cereal yield modeling under actual non-potential field conditions.Water stress limited significantly cereal yield accumulation on clay type soils in Kuuma and Porvoo experimental areas during growing season (Table 1, Figure 2).
The SAR calibration details for ERS, Radarsat and Envisat data have been previously published by Karjalainen and et al. [27,28] and by Matikainen et al. [26].The set of Radarsat-1 images consisted of scenes acquired with various beam modes of Radarsat-1 satellite.The Radarsat incidence look angles varied between 39-44° in the experimental sites (Table 1).Envisat ASAR images were acquired in dual-polarization mode, the VV and VH polarization channels were used in calibration of VGI models (Tables 11 and 12, Appendix B).
In addition, two older auxiliary optical [23] and microwave (HUTSCAT scatterometer, Table 10, Appendix B) [61,63] datasets consisting from Mellilä and Porvoo sites (Nylands Svenska Agric.Advisory Centre, Figure 2) were used in calibration of SatPhenClass phenology and VGI models (I-III, Table 3).The Porvoo experimental site (Kullogård experimental farm, Figure 5) was used for both optical Landsat and SPOT and microwave HUTSCAT measurements for spring cereals.The helicopter mounted HUTSCAT (Helsinki Univ. of Technology, Lab.Of Space Technology [63]) is a monostatic scatterometer system measuring with C-(f = 5.4 GHz) and X-bands (f = 9.8 GHz) with horizontal (HH), vertical (VV) and cross-polarizations levels (VH, HV).In this study only C-band data from HUTSCAT was used for the calibration of VGI models.Both the transmitter and receiver are mounted in the same flight rack under the helicopter [63].The technical configuration of HUTSCAT resembles SAR/ASAR spectrometers currently onboard ERS, Radarsat and ENVISAT satellites (Table 10, Appendix B).The technical configuration of the monostatic HUTSCAT scatterometer has been previously published by Hallikainen et al. [61] and Koskinen et al. [62].The HUTSCAT calibration details have been published Hyyppä et al. [63].
The Finnish Geodetic Institute applied ortho-rectification and radiometric correction for the optical Landsat and SPOT satellite data (Table 1, [23]).The optical calibration procedure for different Vegetation Indices models was originally presented by Price [57] and later revised and utilized for cereals by Flowers et al. [64], Prasad et al. [65], Maas [66][67][68] and Maas and Dunlap [69].Table 11 (Appendix B) presents calibration parameter values applied in optical calibration equations 3-5 (Appendix C), which convert binary digital count (dc) values from satellite optical sensors into surface reflectance values (rf).The Landsat and SPOT raw data coded as 8 bit binary values (digital counts, dc) were converted to integer values between 0-255 (surface reflectance , r f ) with higher values indicating higher reflectance radiation to satellite sensors from ground cover and atmosphere.The satellite digital count-value (dc) to spectral radiance (R) transformation was performed after Equation 3. The spectral radiance -values were transformed into surface reflectance (r f ) by means of solar equivalent radiance (S) (Equation 4).In addition, the sun zenith-angle during the observation time was taken into account (Equation 5).
The calibrated SAR backscattering and ortho-rectified and radiometric corrected optical reflectance data were applied in the calibration of VGI models.Total of 18 SAR and cloud free optical satellite images (Landsat5 /TM and SPOT /HRV2) were obtained during the 1996-2006 campaign.Images were temporarily well dispersed along the growing season between May, June, July and August (Tables 11 and 12, Appendix C).Additional details of the optical calibration procedure and corresponding calibration parameters are presented in Electronic Supplementary Information Appendix (ESI).

Validation of Composite Vegetation Indices (VGI) Models
After the calibration procedure for VGI Composite models (I-III)), the validation procedure was applied to test the VGI Composite models for cereal non-potential baseline grain yield (y b , kg/ha) predictions.Yield differences were calculated as absolute values (kg/ha) and per cent (%) differences vs. observed MAFF annual yield inventory [29,70] and MTT Official Variety Trial grain yield estimates in corresponding sites and years [55,56,71].MAFF and MTT validation datasets with over 3º latitude variation were used to assess cereal baseline yield variation (y b ) in growing zones I-IV (Figure 4).In addition, the Polynomial Infrared and NDVI models (Models I, II) were analyzed with soil*species and soil*cultivar covariance categories.The MTT Official Variety Trial yield data for one Finnish malting barley cultivar (cv.Inari) and for two Finnish spring wheat cultivars (cv.Manu, Satu) in cultivation zones I-IV were used in analyzing soil*cultivar covariances (Table 1, Figure 4a).

Validation Data
The calibrated VGI yield models (I-III) were validated with two independent cereal yield datasets consisting of (i) the Ministry of Agriculture and Forestry Official Inventory Statistics 1994-2006 data for cereal yields [29,30] and (ii) the MTT Agrifood Research Finland Official Variety trial results 1994-2006 (Figure 4, [56,71]).The MAFF validation dataset consisted yield inventory statistics from Etelä Pohjanmaa (Averaged III-IV Zone) Agricultural Advisory Centre.The MTT Official Variety Trial dataset consisted of yield trials from the Ylistaro (Zone III) and Ruukki (Zone IV) MTT Experimental Stations Averaged to Zone III-IV.Zones III-IV were pooled together because the satellite and ground truth measurement sites in Lapua, Seinäjoki and Ilmajoki were located in the middle of 1,100 ETS (Tb = 5 °C) isoline (Figure 4b).

Cereal Canopy Soil Backscattering Covariance Variation
The SAR (Radarsat, ERS), ASAR (Envisat) and HUTSCAT backscattering variance (σ 0 ) between different spring cereals in Lapua, Seinäjoki and Porvoo experimental sites is presented in Table 4.In addition, SAR backscattering horizontal and vertical polarization levels (HH, VV) and the cross-polarization (VH/HV) of the monostatic HUTSCAT scatterometer are presented.The Radarsat horizontal (HH) backscattering (σ 0 ) variation between malt (MBrl) and fodder barley (FBrl) is presented in Radarsat sensor column.Table 4 depicts cereal canopy*soil backscattering covariance during generative phenological phases c p and d p (anthesis, grain full maturity and canopy senescence, Table 2) grown on clay, coarse and fine sand type soils.
Both the soil type and soil*cereal canopy covariances expressed extensive variation with SAR backscattering signal (σ 0 ) on clay type soils, which were dominant in experimental areas in zones I-IV (Table 1).According to Henderson & Lewis [60] the vertically oriented components (VV) with cereals, especially the stems, interact effectively with vertically polarized signals resulting increased backscattering attenuation.Oat species with more planophile canopy and panicle inflorescence structures have different backscattering polarization properties from corresponding properties of wheat and barley, which have more erectophile head and canopy structures [23,28,72].
On sandy clay soils the HUTSCAT backscattering signal (σ 0 ) varied between -6 dB and -29 dB.On fine and coarse sands the Envisat and ERS SAR signal amplitude varied between -17 dB (Envisat in VV, VH dual polarization mode) and -8 dB (ERS) respectively.In the Porvoo experimental area the topsoil (5-10 cm) and the subsoil (<10 cm) consisted of fine and coarse sandy alluvium deposits from the Porvoo river (Figure 5) and gyttja clays containing minor fractions of organic compounds (mould, peat and mud) and silt.In Lapua and Seinäjoki areas the major soil type was sandy clay with minor fractions of organic mould and humus fractions (Table 1).
The SAR backscattering signals indicated extensive topsoil*cereal canopy layer covariance interaction especially after the anthesis (phenological stage c p ) and during grain filling, yellow ripening and canopy senescence (d p ).When analyzing the variation between different SAR sensors the HUTSCAT backscattering signal (f = 5.4 GHz, Table 10,  The helicopter mounted HUTSCAT signal amplitude varied more significantly when compared to other backscattering signals (Envisat, ERS, Radarsat) measured on clay and sandy soils.In addition, the HUTSCAT C-band measurement frequency was slightly higher (f=5.4GHz) than with Envisat, ERS and Radarsat (f = 5.3 GHz).Table 4. SAR backscattering variation (σ 0 , f=5.3-5.4GHz) on sand and clay type soils with different polarization levels during anthesis (c p ), full grain maturity and canopy senescence (d p ). ( 1  (±1.02) (1)MBrl-Malt barley, FBrl-feed barley (2) Soil types in Table 1 (3) HUTSCAT σ 0 VH/HV cross-polarization When analyzing the cereal variation for the wheat, the SAR signal varied between -25 dB (HUTSCAT *c p *VH* fine and coarse sandy clay) and -6 dB (HUTSCAT *d p *VH* sandy clay).The wheat VH cross-polarization amplitude was higher when compared with VV vertical levels both in anthesis (c p ) and full maturity (d p ) on clay and sandy soils.The wheat HH horizontal signal amplitude was higher in stage c p compared to d p .
The oat SAR amplitude varied between -29 dB (HUTSCAT *c p *VH* sandy clay) and -6 dB (HUTSCAT *d p *VH* sandy clay).The oats VH cross-polarization amplitude was higher when compared with VV vertical levels both in anthesis (c p ) and full maturity (d p ) on clay and sandy soils.Both the oats horizontal and vertical (HH, VV) signal amplitudes were higher in anthesis stage (c p ) compared to maturity stage (d p ).
With barley the SAR amplitude varied between -22 dB (HUTSCAT *c p *VH* sandy clay) and -5 dB (HUTSCAT *c p *VV* sandy clay).Especially the barley experimental plots with clay and peat topsoil layers in Porvoo (Table 1, Figure 2) were affected by excessive water from melting snow in the beginning of the growing season (a p ) causing the topsoil to reach the full soil water capacity.Later on during mid-summer, drought periods especially in generative phases (c p , d p ), caused the soil moisture content to recede close to wilting point.These factors affected the translocation of assimilates to the head with a reduction effect on final grain filling.The barley VH cross-polarization amplitude was higher when compared with VV vertical levels both in anthesis (c p ) and full maturity (d p ) on clay and sandy soils.Both the barley horizontal (HH) and vertical (VV) σ 0 signal amplitudes were higher in anthesis stage (c p ) than in full maturity stage

Testing the SatPhenClass Phenological Model Accuracy
The SatPhenClass classification results for spring cereals are depicted in Table 5.The SatPhenClass classification accuracy for spring cereals varied between 61% and 89% with corresponding classification error ranging between phenological classes (a p 21-28%, b p 11-22%, c p 14-28% and d p 25-39%) as calculated from the observed mean phenological Julian dates (DOY).The observed phenological Julian dates for spring cereals were averaged from ground truth experimental sites in cultivation zones I-IV (Figures 2  and 4).In addition, the mean observed Julian dates for spring cereals from MTT experimental stations in cultivation zones I-III were used for phenological validation (Zone I-Pernaja and Inkoo, Zone II-Jokioinen, Zone III and IV-Ylistaro and Ruukki).
The early emergence in vegetative phase (a p , BBCH 0-12) in two leaf stage before double ridge induction and the senescence phase after full maturity and harvest (d p ), BBCH > 90) were difficult to estimate.This increased the classification error in a p and d p classes [35].In addition, the SatPhenClass classification result was accepted only, if both submodels (ETS and LAI comparisons) were in line (Figure 3b).47 (1)

Testing the Composite VGI model (I-III) calibration performance
The spring cereal calibration results derived from phenologically pre-classified (SatPhenClass) SAR and reflectance data, are depicted in Table 5.It contains linear (REG/Stepwise) and non-linear polynomial (RSREG) VGI model (I-III, Table 3) baseline yield estimates (y b, kg/ha) with corresponding R 2 and RMSE, C v (%) and Anova F-test significance level estimates.The calibrated VGI model (I-III) equations are presented in Table 12 (Appendix C).
There was a large variation in VGI Composite model cereal yield responses between years and locations with averaged R 2 0.627 and RMSE 387 kg/ha in growing zones I-IV in southern Finland (Table 1).The average modeled yield response varied between 3.5 t/ha and 4.2 t/ha with spring wheat, between 4.1 t/ha and 4.4 t/ha with barley and between 3.4 t/ha and 3.7 t/ha with oats.coefficient of variation (C v ,%) varied between 1.6% and 28.7% with spring cereals, respectively the VGI model subcomponents (linear, quadratic and cross products) were significant on 0.1% error level.
The calibration results indicate that the R 2 of the VGI modeled (I-III, Table 3) final grain yield accuracy varied significantly between different spring cereals.The overall R 2 average for yield (kg/ha) prediction was for spring sown cereals 0.63 (mean response yield 3750 kg/ha, RMSE 387 kg/ha).According to calibration results obtained from the optical VGI models (I-II) and Composite VGI model (III), the R 2 tends to stabilize on the 0.60-0.70level.With relatively few observations (n < 500) the R 2 tends to increase above 0.80, whereas with larger data sets (n > 1,000 observations) the R 2 varies between 0.50 and 0.70.The overall R 2 for wheat, barley and oats varied between 0.615 and 0.794 for Infrared model (I), between 0.611 and 0.737 for NDVI (II), and between 0.417 (oats using ERS data) and for 0.730 (wheat using Envisat and Radarsat data) for Composite SAR model (III).The total cultivation area used for the calibration of VGI models was 1253 hectares.The R 2 varied for wheat ranged between 0.731 (SAR-model (IIII), RMSE 300 kg/ha) and 0.794 (Infrared model (I), RMSE 42 kg/ha), for barley between 0.448 (III) and 0.615 (I) and for oats between 0.417 (III) and 0.760 (I). The

Soil Species Covariance Validation with Composite Models
The validation results assessing both soil*species (Composite SAR and NDVI Model III) and soil cultivar covariance (Infrared NDVI Models I, II) categories for cereal baseline yield (y b ) are depicted in Table 6 (Model Categories I-II).
In model category I (Table 6), the SAR Composite Model (III) validation results were compared and validated with averaged inventory estimates (kg/ha) from the MAFF Etelä-Pohjanmaa Agricultural Advisory Centre.The SAR and NDVI modeling results depict detailed soil*canopy interactions between cereal canopies and soil top cover (5-10 cm) with fine and coarse sand soil textures.The SAR composite model validation results take into account both the pre-anthesis a p and b p phenological phases (Figure 3b, phases I-II) using the NDVI component and post-anthesis and senescence phases (c p and d p, phase III) using the SAR component with HH and VV levels.
In model category II (Table 6), the soil cultivar covariance validation results are depicted.Two spring wheat (cv.Manu, Satu) and one malting barley (cv.Inari) cultivars grown on heavy clay type soils and currently cultivated in Finland were analyzed by using optical VGI Infrared and NDVI models [44,56,71].The response mean yields (kg/ha) were compared with the MTT Official Variety Trial mean yield results correspondingly grown on clay type soils in cultivation zone I (Pernaja and Inkoo Exp.Stations), II (Jokioinen Exp.Station) and III-IV (Ylistaro and Ruukki Exp.Station, Figure 4a).
The Model I and II validation results (Model Category II, Table 6) using MTT validation data, indicated that both observed wheat and barley yield levels from the MTT official variety trials (kg/ha) exceeded corresponding modeled spring wheat and barley yield levels.The soil*species and soil*cultivar covariance validation averaged results with Infrared and NDVI models indicated an underestimated yield difference (DP MTT , %) by -10% in soil*species category and between -31% and -14% in soil*cultivar category.The modeled results for spring wheat and barley baseline yield levels (y b ) in both categories were lower than the corresponding yield levels in MTT Official Variety trials.

Implications from SAR Soil*Canopy Interactions
Both the SAR backscattering (σ 0 ) signal and SAR and NDVI baseline yield (y b ) levels varied significantly both in cereal soil*species and species*canopy covariance categories.In the species*soil covariance category SAR backscattering signal varied significantly especially on clay type soils with minor fractions of sand, silt and organic mould in the Porvoo, Mellilä, Kirkkonummi, Jokioinen and Lapua experimental areas (Table 1).
With spring wheat the VH cross-polarization amplitude was higher when compared with VV vertical levels both in anthesis (c p ) and full maturity (d p ) on clay and sandy soils.The wheat horizontal signal (HH) amplitude was higher in c p stage compared to d p .
The microwave Radarsat VH amplitude in barley plots was higher when compared with VV vertical levels both in anthesis (c p ) and full maturity (d p ) on clay and sandy soils.Both the barley horizontal and vertical (HH, VV) σ 0 signal amplitudes were higher in anthesis stage when compared to full maturity stage.
With oat cereals, the VH cross-polarization amplitude was higher when compared with VV vertical levels both in anthesis (c p ) and full maturity (d p ) on clay and sandy soils.Both the horizontal and vertical (HH, VV) signal amplitudes were higher in anthesis stage than in full maturity stage containing grain filling, yellow ripening and canopy senescence stages.
In the species*canopy covariance category level there was a significant variation both in backscattering amplitude and polarization properties between spring wheat, barley and oats.Especially the canopy structure of oat cereals differs morphologically from other spring cereals.Oat cereals with more planophile canopy and panicle inflorescence structures differ in polarization properties from those of wheat and barley with more erectophile head and canopy structures [23,60,72].

Implications from the SAR Composite Modeling Results for Baseline Yield Levels (Y b )
The SAR calibration results with Composite NDVI and SAR models indicated that averaged baseline yield (y b ) response varied between 3.5 t/ha and 4.2 t/ha with spring wheat, between 4.1 t/ha and 4.4 t/ha with barley and between 3.4 t/ha and 3.7 t/ha with oats.Correspondingly the SAR validation results compared with MAFF data indicated that with spring wheat the overestimation was between +104% (Envisat) and 107% (Radarsat) from the observed MAFF inventory reference.Respectively with oats the underestimation was between 85% (Envisat) and 88% (ERS2).With barley the underestimation was 97% with Envisat and overestimation 112% with ERS.With malt barley the yield was underestimated (96%) with Radarsat data and overestimated with fodder barley (102%).
The use of Composite Multispectral model (III) combined with the SatPhenClass phenological classification algorithm for spring cereals provides a promising integrating technique for combining both microwave SAR and optical reflectance data.The Composite Multispectral model takes into account both the pre-anthesis phenological phases using the NDVI component and post-anthesis and senescence phases using the SAR component with backscattering polarization levels.In addition, the Composite multispectral model can be used in assessing the soil*canopy covariances between cereal canopies and soil top layers on different soil types.
The Composite model (III) calibration results indicate that the R 2 tends to stabilize on the 60%-70% level similar to optical VGI models (I-II).The Composite model results indicated that the use of the Envisat ASAR additional cross-polarization component (VH) did not increase the 2 level compared to ERS (VV) and Radarsat (HH) with only one polarization level in the 5.3 GHz measurement spectrum.The Radarsat signal with the HH horizontal polarization component yielded highest R 2 values with wheat, barley and oats.ERS signal with only one VV vertical polarization component produced consistently lower R 2 values.In Finland Karjalainen et al. [28] reported SAR modeling results with the mean R 2 of 0.55.The average crop height was used to estimate the amount of biomass using dual-polarization (VV/VH) Envisat SAR data in the Lapua and Seinäjoki experimental sites [28].
Recently McNairn et al. [1] reported the composite VV-VH SAR and optical data to be the most suitable for wheat, maize and soybean classification with over 85% overall accuracy (κ Kappa range 0.47-0.89) in Canadian growing conditions.McNairn et al. [1] applied three primary classification methodologies Neural Networks [73], Gaussian Maximum-Likelihood Classifier [21,74] and Decision trees [75] for crop classification using composite SAR and optical data.Respectively in this study the SatPhenClass classification accuracy for spring cereals varied between 61% and 89% in phenological classes (a p -d p ). Especially the early vegetative phase before double-ridge induction (a p ) [35,50,51] and post-harvest senescence (d p ) phases were major source for error variation decreasing the overall classification accuracy.Highest classification accuracies (>80%) were obtained during the anthesis (b p ) near the LAI maximum proximity.
The composite model validation results indicated an overestimation with wheat and feed barley cereals.On the contrary, an underestimation was noticed with oats and malting barley cereals by using different SAR/ASAR and NDVI sources.In summary, the results suggest the potential inclusion of horizontal polarization component (HH) in the SAR composite models.The oat canopy and head structures differ from those of wheat and barley potentially changing the backscattering polarization properties [23,28,60].

Conclusions
The cereal validation results obtained in this study suggest that spring cereal baseline yield estimates derived from Vegetation Indices Models tend to retain between the observed MAFF yield inventory estimates and MTT Agrifood Official Variety trial data results.The annual MAFF non-potential yield inventory estimates are obtained from suboptimal growing conditions on actual farm level, where as MTT Agrifood Official Variety trial results are obtained from more optimal growing conditions.

Figure 1 .
Figure 1.System Analysis and design diagram by satellite systems and VGI models [31,38].
3) explain the validation of VGI models.Phases X-XI(a,b) (Sections 3.2, 3.3) explain the estimation of baseline grain yield levels (y b ) for different spring cereals resulting outputs for SAR & NDVI models (Phase XI(a), Part I) and for detailed optical VGI models (Phase XI(b), Part II).

( b )
Calibration of Composite Vegetation Indices (VGI) models (models I-III in Part I, models IV-V in Part II) with pre-classified C-band SAR and optical data.Both linear and non-linear polynomial response functions in VGI models were applied for cereal baseline yield (y b ) estimations.

Figure 4 .
Figure 4. (a) Growing zones (I-V) and MTT Experimental Station locations in Finland, (b) ETS (T b 5 °C) cumulative isolines and (c) Meteorological Weather Stations in Finland (Original Data (C) MTT Agrifood Research Finland, Finnish Meteorological Institute and Finfood Finland).

Figure 5 .
Figure 5. Landsat/TM satellite image from the Porvoo field experimental area.The image location was used for optical Landsat, SPOT and microwave HUTSCAT measurements.Porvoo River on the right corner of the image.

Table
Composite Model III (NDVI * SAR) calibration results for ERS, Radarsat and Envisat are presented in Table 5 (Model III SAR, NDVI ).The R 2 for the Composite model III varied for wheat between 0.723 (Envisat SAR-component, RMSE 302 kg/ha) and 0.731 (Radarsat SAR-component, RMSE 300 kg/ha), for barley between 0.694 (ERS component, RMSE 482 kg/ha) and 0.702 (Radarsat SAR component, RMSE 322 kg/ha) and for oats between 0.417 (ERS SAR component, RMSE 584 kg/ha) and 0.624 (Radarsat, RMSE 483 kg/ha).The corresponding optical calibration results for soil*species and soil (clay) * cultivar covariances with Infrared Polynomial Model I are reviewed in Part II.

Table 10 .
SAR and optical satellite systems used in remote sensing campaigns.