Key Advances and Remaining Knowledge Gaps in Remote Sensing ...

Key Advances and RemainingKnowledge Gaps in Remote Sensingfor Precision AgricultureD. J. MullaProfessor and Larson Chair forSoil and Water ResourcesDirector Precision Agriculture CenterDept. Soil, Water and ClimateUniversity of Minnesota

Precision Agriculture• A managementpractice applied atthe right time andthe right place inthe right amount• Field sub-regionmanagement– Nutrients– Drainage orIrrigation– Pests and Weeds– Tillage and Seeding

Precision ManagementOptimal ResourceManagementImplementationWISDOMDecisionProcessKNOWLEDGEMap BasedorReal TimeApproachAnalysis andDiagnosticsDataCollectionDATAInformationManagementINFORMATION

The Electromagnetic Spectrum

Remote Sensing• Bare soil reflectance– Soil organic carbon and water content– Iron oxides or carbonates• Crop reflectance– Leaf area index– Crop growth stage– Crop color and leaf N status– Weeds and disease• Thermal emission of energy– Surface temperature and crop water stress

Remote Sensing Applications• Mapping crop variability– Crop growth, biomass, and yield– Crop pphenology and evapotranspirationp– Crop stresses due to nutrients, disease,weeds, and insects• Mapping soil variability• Targeted sampling• Management zone delineation• Ability to monitor changes over time

Remote Sensing Sugar Beetsg g(Seelan et al., 2003)

History of Remote Sensing

Satellite Remote SensingSatelliteSpectral BandsReturnSuitability(Year)(Spatial Resolution)Frequencyfor PA(d)Landsat 1 (1972) G, R, two IR (56x79 m) 18 LAVHRR (1978) R, NIR, two TIR (1090 m) 1 LLandsat 5 TM (1984) B, G, R, two NIR, MIR, TIR (30 m) 16 MSPOT 1 (1986) G, R, NIR (20 m) 2-6 MIRS 1A (1988) BGRNIR(72 B, G, R, m) 22 MERS-1 (1991) Ku band altimeter, IR (20 m) 35 LJERS-1 (1992) L band radar (18 m) 44 L

Satellite Remote SensingSatelliteSpectral BandsReturnSuitability(Year) (Spatial Resolution) Frequencyfor PAIKONOS (1999) Panchromatic, B, G, R, NIR (1-4 m) 3 H(d)Terra EOS ASTERG, R, NIR & 6 MIR, 5 TIR bands (15-90 m) 16 M(2000)EO-1Hyperion (2000) 400-2500 nm, 10 nm bandwidth (30 m) 16 HQuickBird (2001) Panchromatic, B, G, R, NIR (0.61-2.4 m) 1-4 HEOS MODIS (2002) 36 bands in VIS-IR (250-1000 m) 1-2 LRapidEye (2008) B, G, R, Red edge, NIR (6.5 m) 5.5 HGeoEye-1 (2008) Panchromatic, B, G, R, NIR1, NIR2 (1.6 m) 2-8 HWorldView-2 (2009) PBGYRRededge P, B, G, Y, R, Red edge, NIR (0.5 m) 11 1.1 H

Increasing Spatial Resolution ofImagery

Pixel Size

Increased Homogeneity of Pixels

Limitations to SatelliteRemote Sensing• Coarse spatial resolution and infrequentrepeat coverage for older satellite platforms• Difficulty obtaining images when needed• Interference from clouds• Changes in irradiance on multiple passes• Slow turn-around time due to imageprocessing for calibration, corrections, andgeo-rectification• Cost ranges from $23 - $45/km 2

Aerial RemoteSensing• Airplanes• UnmannedAerial Vehicles

Proximal Remote Sensing• Sensors can be mounted on tractors,spreaders, sprayers or irrigation booms– GreenSeeker– Crop Circle– WeedSeeker– Infrared thermometers• Allows real time site specificmanagement of fertilizer, pesticides orirrigation

Data Interpretation• Classification (Bauer, 2000)– Supervised, unsupervised, multi-temporal temporal data sets,decision tree classifiers, neural network classification,guided clustering, fuzzy sets, contextual classifiers,and use of auxiliary information• Pattern Detection (Civco, 2002)– Image differencing, cross correlation, chemometricanalysis, principal component analysis, neuralnetworks, object oriented classification (Walker, 2003),fuzzy set theory (Viscarra Rossel, 2006)• Assess accuracy (Congalton, 1991; Foody, 2002)Vi l t A i S li t t i– Visual assessment, Area comparisons, Sampling strategies,Kappa statistic, Confusion matrix

Traditional Classification

Fuzzy Classification

Spectral Decomposition andChemometrics• Spectral Mixing ing Analysis (Huete, 1991)• Principal Component Analysis• Partial Least Squares Regression (Lindgren,1994)• Stepwise Multiple l Linear Regression• Multiple Regression Analysis• Multivariate i t Adaptive Regression Splines

Spectral Mixing Analysiswhere ρ is reflectance, m are end-members, k are bands, and i,j are pixels

Principal Component Analysis

Artificial Neural NetworksOutput Layerweighting factorsa summation functiona transfer functionInputLayerNeuron(hidden)Miao, Y., D. J. Mulla andP. C. Robert. Prec. Agric.7(2):117-136, 2006scaling and limiting processesan output tfunctionan error functiona back-propagated value/learning function

Pixel Accuracy AssessmentNon-Spatial Accuracy TestBorder Pixel Accuracy Important

Confusion MatrixOmission errors involve underestimation of a classOmission errors involve underestimation of a classCommission errors involve overestimation of a class

Accuracy AssessmentStatisticsKappa Statisticfor ClassificationAccuracy

Confusion MatrixDCNIRB Leaf NUser's CommissionHigh Moderate Low None Total Accuracy ErrorHigh 19 2 1 0 22 0.86 0.14Moderate 2 3 0 2 7 0.43 0.57Low 2 1 2 4 9 0.22 0.78None 0 1 9 32 42 076 0.76 024 0.24Total 23 7 12 38 80Producer's Accuracy 0.83 0.43 0.17 0.84 0.70Omission Error 0.17 0.57 0.83 0.16Nigon, Rosen, Mulla et al., 2012

Estimate Biophysical orBiochemical VariablesCourtesy of Drs. V.C. Patilyand K.A. Al-Gaadi,King Saud University

Spatial Pattern

Spatial Pattern is Complicated whenHyperspectral Data are involved!

Integration of Imagery in GIS

Statistical Advances• Traditional statistics– Mean and dispersion– Histograms• Spatial and non-traditional statistics– Spatial filtering (low and high pass kernalfilters, Kalman filters, wavelets, etc)– Homogeneity, Contrast, Dissimilarity– Geostatistical analysis

Using Geostatistics toCorrect Errors in RS Images• Used when pixels are missing– Cloud cover or satellite malfunctionRossi et al., Rem. Sens. Environ. 49:32-40, 1994Zhu et al., Rem. Sens. Environ.124:49-60, 2012

Spatial Modeling & InterpolationCokriging KrigingMulla D J Geostatistics remote sensing and precision farming 1997 pp 100 119Mulla, D. J. Geostatistics, remote sensing and precision farming. 1997. pp 100-119.In: A. Stein and J. Bouma (eds.), Precision Agriculture: Spatial and Temporal Variabilityof Environmental Quality. Ciba Foundation Symposium 210. Wiley, Chichester.

400Kriged Phosphorus (ppm)300(m)DIS STANCE2001000 0 100 200 300 400 500 600DISTANCE (m)

400Cokriged Phosphorus (ppm)300(m)DISTANCE2001000 0 100 200 300 400 500 600DISTANCE (m)

Targeted Soil Sampling StrategyBased on Bare Soil NIR Image1501005000 100 200 300 400 500 600 700 800

Applications of Remote Sensingin Precision Ag• Crop growth stages, LAI, biomass, andyield• Crop chlorophyll and nitrogen stress• Crop ET and water stress• Weed, insect and disease infestations• Soil salinity, organic matter, moisture, etc• Delineating management zones• Precision conservation

Crop SensorsYear Innovation Citation1992 SPAD meter (650, 940 nm) used to detect NSchepers et al., 1992dfii deficiency in corn1995 Nitrogen Sufficiency Indices Blackmer and Schepers,19951996 Optical sensor (671, 780 nm) used for on-the-goStone et al. (1996)detection of variability in plant nitrogen stress2002 Yara N sensor (450-900 nm) Link et al. (2002), TopConIndustries2002 GreenSeeker (650, 770 nm) Raun et al. (2002), NTechIndustries2004 Crop Circle (590, 880 nm or 670, 730, 780 nm) Holland et al (2004), HollandScientific2002 CASI hyperspectral sensor based indexHaboudane et al. (2002;measurements of chlorophyll2004)

Spectral Signatures

Visible Absorption Spectrum

The Red-NIR Reflectance Space

Satellite RS-based N Managementof Rice in Northeast China2010-6-3 2010-6-25 25 2010-7-14 14 2010-7-30生 Dfii Deficient生 NormalExcessiveVillage levelField levelCourtesy ofDr. Y. Miao,Beijing Agric. Univ.

Crop Sensor-based N Management Strategy (China)(Li et al., 2009. Soil Science Society of America Journal 73:1566-1574)1)N Rate (kg ha -1500400300200100Soil_basedSensor_basedFPN Rate02006 2007Yeard (kg ha -1 )Grain Yiel700060005000400030002000100002006 2007YearWinter Wheat Yield

Spectral Indices• Canopy reflectance is affected by water,chlorophyll, h ll canopy density and age, soil, etc• Leaf water potential and leaf temperature– 1981 Crop Water Stress Index (Jackson)• Leaf Area Index– 1974 NDVI (Rouse, 670 & 800 nm)• Soil Adjusted Vegetative Index (670 & 800 nm)– 1994 MSAVI (Qi)• Chlorophyll Absorption Ratio Index– 2000 MCARI (Daughtry, 550, 670 & 700 nm)

Crop Water StressHigh water Treatments (low temperatures)Blue and green spotsThermal Crop WaterStress Indices.– Crop: Cotton– Location: Arizona– Sensor: Thermalscanners (helicopter)Source: http://www.uswcl.ars.ag.govLow water Treatments (high temperatures)Yellow and orange spots.

RGBThermalCourtesy Drs. V. Alchanatis &Y. Cohen, Volcani Center, AROand Pivot Irrigation & Pumping

Types of Reflectance Spectra• Panchromatic reflectance– An average over all wavelengths• Broad band or multispectral reflectance– Reflectance at a few specific discretewavelengths– B, G, R NIR portions of spectrum• Hyperspectral reflectance– Reflectance at specific narrow banddiscrete wavelengths across a largecontinuous spectral range

Multi-spectral broad-band vegetationindices available for use in PrecisionAgricultureIndex Definition ReferenceGDVI NIR-G Tucker, 1979NDVI (NIR-R)/(NIR+R) Rouse et al., 1973GNDVI (NIR-G)/(NIR+G) Gitelson et al., 1996SAVI 1.5 * [(NIR-R)/(NIR+R+0.5)] Huete, 1988GSAVI 1.5 * [(NIR-G)/(NIR+G+0.5)] Sripada et al., 2005OSAVI (NIR-R)/(NIR+R+0.16) Rondeaux et al., 1996

Hyperspectral Data CubeNigon, Rosen, Mulla et al., 2012

Best Reflectance Wavelengths?Thenkabail et al. (2000)• The greatest information about plant characteristics withmultiple narrow bands includes the longer redwavelengths (650-700 nm), shorter green wavelengths(500-550550 nm), red-edge edge (720 nm), and NIR (900-940940 nmand 982 nm) spectral bands– The information in these bands is only available in narrowincrements of 10-20 nm, and is easily obscured in broadmultispectral bands that are available with older satellites• The best combination of two narrow bands in NDVI-likeindices is centered in the red (682 nm) and NIR (920 nm)wavelengths, but varies depending on the type of crop, aswell as the plant characteristic of interest

Hyperspectral narrow-bandvegetation indices available for usein precision agricultureIndex Definition ReferenceSR1 NIR/Red = R 801 /R 670 Daughtry et al., 2000SR7 R 860 /(R 550 * R 708 ) Datt, 1998NDVI (R 800 -R 680 )/(R 800 +R 680 ) Lichtenthaler et al.,1996Green NDVI (R 801 -R 550 )/(R 800 +R 550 ) Daughtry et al., 2000(GNDVI)NDI1 (R 780 -R 710 )/(R 780 -R 680 ) Datt, 1999NDI2 (R 850 -R 710 )/(R 850 -R 680 ) Datt, 1999

Hyperspectral narrow-bandvegetation indices available for usein precision agricultureIndex Definition ReferenceMCARI [(R 700 -R 670 )-0.2(R 700 -R 550 )](R 700 /R 670 ) Daughtry et al., 2000TCARI 3*[(R 700 -R 670 )-0.2*(R 700 -R 550 )(R 700 /R 670 )] Haboudane et al., 2002OSAVI (1+0.16)(R 800 -R 670 )/(R 800 +R 670 +0.16) Rondeaux et al., 1996MSAVI 0.5[2R 800 +1-SQRT((2R 800 +1) 2 -8(R 800 -R 670 ))] Qi et al., 1994MCARI21.5[2.5( R800 R670)1.3(R800 R550)]Haboudane et al., 2004(2R8001)2 (6R800 5R670) 0.5

Derivative Spectra• Hyperspectral reflectance data are collected innearly continuous narrow bands across a widerange of wavelengths• The derivative of hyperspectral reflectance datacan be very useful at indicating portions of thespectrum where the slope of the reflectancecurve changes rapidly• Derivative spectra reduce interference from soil

Derivative SpectraNigon, Rosen, Mulla et al., 2012

Lambda-Lambda Plots• Lambda-lambda plots involve calculating, forexample, the r 2 coefficient for leaf N content at allhyperspectral reflectance bands• A graph of the r 2 coefficient for all possiblecombinations of band 1 on the x-axis and band 2 onthe y-axis results in a lambda-lambda plot• The lambda-lambda plot is useful for identifyingwhich combinations of two bands contain redundantinformation about N stress• Spectral bands or narrow band indices should beselected with low r 2 coefficients to eliminateredundancy and maximize information about cropcharacteristics (such as N stress).

LambdalambdaplotNigon,Rosen,Mulla et al.,2012

Temporal Analysis of Data

In-Season N Management(From J. Schepers, 2005)

Properties of N deficient Plants• Green reflectanceincreases• Red reflectanceincreases & NIRreflectancedecreases• Differences inreflectance greatestbetween 550 – 600nm, followed byred-edge (680 –730 nm)

Hyperspectral Imaging forN Sufficiency in MaizeCM = 56.74 + 6.48 * R 713 + 2.02 * R 953 + 8.42 * R 666 - 9.87 * R 695- 0.75 * R 972 - 19.57 * R 554 + 16.68 * R 499 - 1.66 * R 934Miao et al., Precision Agriculture 10:45-62, 2009

Promising Indices for Crop N StressNormalized Difference Index 2:

Results/discussionLeaf NN stress in potatoNDI2Nigon, Rosen, Mulla et al., 2012

Leaf N (CV 14%) NDVI (CV 5%)Nigon,Rosen,Mulla,2012DCNI (CV 15%) NDVI1 (CV 26%)

LiDAR• Light Detection and Rangingscans earth’s surface usinga laser with visible radiation• Provides a high resolutionDigital Elevation Model(DEM) which can beprocessed in GIS with terrainanalysis– Slope, curvature, flowaccumulation, etc.– Stream power, topographicwetness, etc.

Side InletPrecision Conservationin Critical Source AreasGalzki, J., A. S. Birr and D. J. Mulla. 2011.J. Soil Water Conserv. 66:423-430.430.Gully• Two criteria:• Accumulationof surface flow• Hydrologicconnection tosurface waters• GIS basedLiDAR dataused toidentify CSAs

Example: Using Specific Catchment Area to Identify Critical Source AreasBeauford Watershed (Blue Earth County)

Conclusions• Precision agriculture dates back to the middleof the 1980’s• Remote sensing applications in precisionagriculture began with sensors for soilorganic matter, and have quickly diversified toinclude satellite, aerial, unmanned aerialvehicles, tractors, spreaders, sprayers andirrigation booms• Spatial resolution of aerial and satelliteremote sensing imagery has improved from100’s ofmtosubmeter sub-meter accuracy

Conclusions• Wavelengths in use range from the ultravioletto microwave portions of the spectrum– Red edge reflectance is increasingly available– LiDAR, fluorescence and thermal spectroscopyhave advancedd• Spectral bandwidth has decreased with theadvent of hyperspectral remote sensing• A variety of useful spectral indices now existfor various precision agriculture applications– The best spectral indices for N stress are robustover crop species, variety and growth stage (withhigh r 2 and CV)

Conclusions• Return frequency of satellite remote sensingimagery has improved dramatically– Timeliness, cloud interference and cost of imageryare still issues• More research on remote sensing withunmanned aerial vehicles is needed toimprove the timeliness of data collection• Errors in geo-rectification and geometriccorrection from aerial platforms are becomingmore important relative to image pixel size

Remaining Knowledge Gaps• More emphasis is needed on chemometricmethods of analysis to assess cropbiophysical and biochemical characteristics• Spectral indices that t simultaneously l allow forassessment of multiple crop stresses (e.g;water, nitrogen and disease) are lacking• Higher spatial resolution of aerial remotesensing is needed for early detection ofdisease

Remaining Knowledge Gaps• Sensors are needed for direct estimation ofnutrient t deficiencies i i without t the use ofreference strips• There are many algorithms and spectralindices for management of crop N stress– The utility of these seems to vary by year, croptype, variety and growth stage– Simple, generalizable solutions are needed• Remote sensing needs to be better integratedwith automated, online soil and cropmanagement decision support systems

Remaining Knowledge Gaps• At present there is considerable interest incollecting remote sensing data at multipletimes in order to conduct near real time soil,crop and pest management• We have progressed from farming by soil, tofarming by grid cell to farming bymanagement zone and now the challenge isto manage individual plants in real-time

Questions?Acknowledgement: Some images were obtained from“Fundamentals of Satellite Remote Sensing”by Chuvieco and Huete (2010)