Innovationer för hållbar växtodling 3d. Underlag för ett webbaserat beslutsstödssystem för precisionsodling Lena Engström, Kristin Piikki, Mats Söderström och Bo Stenberg. Sveriges lantbruksuniversitet (SLU), Inst. för Mark och miljö, Precisionsodling och pedometri, Skara, Sverige.
3d. Underlag för ett webbaserat beslutsstödssystem för precisionsodling Syfte och mål med delaktiviteten: utveckla och anpassa algoritmer som översätter tillgängliga gröd- och markdata till platsanpassade åtgärdsrekommendationer.
3d. Underlag för ett webbaserat beslutsstödssystem för precisionsodling Med hjälp av olika verktyg som handburna sensorer, drönare och satelliter: I höstvete bestämma skörd vid kompletteringsgödsling. I höstraps bestämma kväveinnehåll på hösten för optimal beräkning av vårgivan. I gräsvall bestämma näringsinnehåll för optimal skördetidsprognos.
Data från drönare: Röda polygonen visar jordbruksblocket 5m Fotografering med RGB-kamera (RX-100, Sony) Fotografering med multispectral kamera (RedEdge, Micasense) Skapande av mosaik (SOLVI) Skapande av mosaik (Micasense ATLAS) Georeferering Mosaikerna georefererades med hjälp av Jordbruksverkets blockdatabas. Ofta 5-10 m fel innan georeferering. (ArcGIS, ESRI). Dataextraktion Ett medelvärde för alla celler inom 5 m radie från inmätta punkten beräknades Detta gjordes band för band (www.r-project.org). K. Piikki
Data från drönare i försök: Mosaikbildens pixlar 8 x 8 cm, medianvärdet i varje ruta tas fram för de fem våglängdsbanden
Data från satelliter: 10x10, 20x20, 30x30 m raster/pixel = 1 värde 10 m 10 m
Winter oilseed rape Decision support for calculating the optimal N rate in spring to winter oilseed rape by estimation of N- uptake in autumn by means of remote sensing. Four fields, crop sampling (1 m 2 ) at five positions in each field in late autumn, n= 25. RGB-image 13/10 2016, Entorp, Skara, 22 144 kg N/ha 1. 22 kg N/ha 3. 50 kg N/ha 5. 144 kg N/ha 2. 23 kg N/ha 4. 108 kg N/ha
Winter oilseed rape Lanna, skifte 13, 12 84 kg N/ha 5. 84 kg N/ha 4. 12 kg N/ha 3. 40 kg N/ha 2. 72 kg N/ha 1. 65 kg N/ha
Winter oilseed rape Bjertorp skifte 31, 48 145 kg N/ha 5. 48 kg N/ha 4. 49 kg N/ha 3. 55 kg N/ha 2. 65 kg N/ha 1. 145 kg N/ha
Results 2016: Modelling the autumn N-uptake of winter oilseed Best prediction of N-uptake (11-90 kg N/ha, 4 sites, n= 20 ) with RGB-pictures (drone) and handheld GreenSeeker (NDVI), RMSECV = 12 kg N/ha. Good prediction models also with all wavebands (RMSEC = 14 kg N/ha) or single indices, RMSECV = 15-16 kg N/ha (drone). No good prediction with satellite-pictures since sampling area (1 m 2 ) did not properly represent the area of 10-30 m pixels in the satellites. Regression models (PLS and SLR) Response: N-uptake (kg N/ha). Predictors: All wave bands (B, G, R, Rededge, NIR ) Indices (NDVI, MSAVI, SAVI) RGB Regression models Cross-validated r 2 RMSECV rpd UAV- Micasense, indices NDVI all 0,38 33 1,2 not 2 high 0,64 15 1,6 val. gårdar 0,58 16 1,5 MSAVI2 all 0,37 33 1,2 not 2 high 0,61 16 1,5 val. fields 0,52 17 1,4 SAVI all 0,38 33 1,2 not 2 high 0,64 15 1,6 val. fields 0,58 16 1,6 UAV - Micasense, wavebands (B, G, R, Red edge, NIR) not Entorp all 0,24 37 1,0 ej 2 höga 0,69 14 1,7 val. fields 0,30 21 1,1 UAV Sony (R, G, B) all 0,16 38 1,1 not 2 high 0,77 12 2,1 val. fields 0,71 13 1,9 Handheld GreenSeeker NDVI all 0,57 28 1,5 Not 2 high 0,78 12 2,1 val. fields 0,68 14 1,8
Results: UAV - Sony (R,G, B) Cross validated model (4 sites), R 2 = 0,71, RMSECV = 13) 120 RGB: PLS calibration on respons variable N-uptake (kg N per hectar) 110 100 Lanna Entorp 90 80 70 Entorp Bjertorp21 Lanna Predicted Y (kg N/ha, Factor-3) 60 50 40 30 20 10 0 Lanna Bjertorp21 Bjertorp21 Entorp Bjertorp21 Entorp Lanna Bjertorp31 Bjertorp31 Bjertorp21 Reference Y (kg N/ha, Factor-3) 0 10 20 30 40 50 60 70 80 90 100 110 120 Bjertorp31 Bjertorp31 Lanna
Plans for 2017: 50-70 crop samples at 4-6 sites 2017 10-20 samples with high biomass, 100-200 kg N/ha. Improve georeferencing by using 2-3 markers around each sampling plot, instead of only one (2016). Adapt sampling strategi for satellite pictures. Sample in larger homogenious areas!
Results 2016: Modelling yield in winter winter wheat Using 6 fields with 20 samples in each (n= 120): No good prediction of yield (or N-uptake and biomass) was achieved with the models tested (multispectral images from drones and satellites). No good prediction due to sampling area (0,5 x 0,5 m) did not properly represent the area from the satellite and UAV pictures. Using 4 field trials from 2015 (n= 225) and a handheld sensor (multispectral): Good prediction of yield was achieved with all wavebands, NDVI+Rededge and MSAVI+Rededge. Regression models (PLS and SLR) Response: Yield (kg/ha) Predictors: All wave bands (400-1000) Indices (NDVI, MSAVI, SAVI..) n= 225 Korsvalidering Validering gårdsvis r2 rmsecv rpd valid. r2 rmsecv rpd valid. Alla våglängder 0,82 895 2,3 0,73 1326 1,6 utan 0N 0,45 797 1,4 Index NDVI 0,66 1219 1,7 MSAVI 1,45 1549 1,4 SAVI 0,34 1694 1,2 Rededge (NDRE) 0,73 1078 1,9 Clor.index 0,72 1098 1,9 Rededge+Cl.index 0,75 1043 2,0 NDVI+Rededge 0,76 1032 2,0 0,68 1187 1,8 MSAVI+Rededge 0,75 1050 2,0 0,59 1374 1,5
NDVI+Rededge, cross validation with sites (4) Predicted vs. Reference 12500 12000 11500 11000 10500 10000 9500 9000 Elements: Slope: Offset: Correlation: R2(Pearson): R-Square: RMSECV: SECV: Bias: 225 0,7442809 2,5637e+03 0,826647 0,6833452 0,6976857 1,1872e+03 1,1899e+03 9,0038147 33 2Ba 56 2Ba43 2BbvIII 3950 2BbIII 2BbvI 4752 53 2BbII 57 2Ab 332BbvII 15 2Ba 2Ba16 2235 37 19 17 2BbII 19 2Ab 2BbvIII 36 4 2BbvII 2BbII 2Aa 5 2BbIII 7 2BbI 48 2Aa 54 1Bb 21 2BbvI 51 1AaIII 46 1Aav26 2Ab 27 2BbI 25 2BbIII 45 1Ba 49 1Ab 55 1AaI 16 2BbII 34271Ab 2BbI 40 1AaII 4 2Aa 33 48 2Aa 19 2BbvIII 39 2BbIII 47 2BbII 25 2BbIII 1 2BbvI 52Ba 2BbIII 25 2BbIII 520 2BbIII 139 1Ab 2Ab 17 2BbvII 26 2Ab 15 2Ba 37 2BbvIII 18 1AaII 9 2Ab7 32BbI 48 29 2Aa 501AaIII 28 2BbvI 36 1AaI 57 2Aa27 2BbvII 2BbI 15 4 2Aa43 2BbvIII 7 2BbI 13 1Ab 11 53 2Ab 56 52 2Ba 5711 1Ba 2Ba 2BbvII 1Bb16 1Aav 28 2BbII29 1AaIII 1AaI 2BbI 35 2BbvII 36 2Aa 46501Aav 2BbvI 2 1AaI 23 40 1AaII 1 2BbvI 5522 1Aav 17 30 1Ba 6 1AaII 26 34352Ab 1Ab 2BbvII 54 2BbII 3 1Aav19 392BbvIII 2BbIII 5247 53 37 2BbII 2Ab 56 2Ba 43 2BbvIII 2BbvIII 24 1Bb 8 1AaIII 1Bb 18 1Ba 45 1Ba 306 2 21Ba 1AaII 28 2 1AaI 49 831AaIII 1Ab 1Aav 512BbvI 1AaI 40 1AaII 21 2BbvI 22 46 2BbII 1Aav 1AaIII 24 1Bb 24 1Bb 1 2BbvI 23 13 111Ab 1Bb20 1AaII 645 1Aav 1AaII 1Ba 55511AaI 1AaIII 49 54 1Ab1Bb 18 1Ba 23 1Aav 19172BbvIII 27 2BbI 22 2BbII 30 1Ba 43 50 2BbvIII 48 57 2Aa 2BbvII 39 2BbIII 21 2BbvI 29 1AaIII 34 561Ab 2Ba 29 1AaIII 3047 1Ba2BbII 4 2Aa34 1Ab 5 2BbIII 2625 2Ab2BbIII 33 2Ba 24 1Bb 28 1AaI 36 2Aa 52 15 352BbI 2Ba 2BbvII 37 8 1AaIII7 2BbvIII 53 2Ab 6 1AaII 20 1AaII 9 2Ab16112BbII 1Bb 13 1Ab 20 1AaII 4021AaII 3 1Aav 45 1Ba 23 1Aav 46 1Aav54551Bb 18 1Ba 49 1Ab 51 1AaIII 8500 Predicted Y (Skörd kg/ha, Factor-1) 8000 7500 7000 6500 6000 5500 5000 4500 4000 3500 44 0AaII 44 0AaII 42 0AaI 14 0AaI 32 0AaII 12 0AaII 44 0AaII 41 380AaIII 10 0AaIII 12 0AaII 42 0AaI 42 0AaI 41 38 0AaIII 38 0AaIII 41 0AaIII 14 0AaI 31 0AaI 32 0AaII 12 0AaII 31 0AaI 10 0AaIII 10 0AaIII 32 0AaII 14 0AaI 38 0AaIII 10 0AaIII 12 14 0AaII 0AaI 31 0AaI 32 0AaII 31 0AaI 3000 Reference Y (Skörd kg/ha, Factor-1) 4000 5000 6000 7000 8000 9000 10000 11000 12000 13000
All wave bands, cross validation with sites (4) Predicted Y (Skörd kg/ha, Factor-3) 13500 13000 12500 12000 11500 11000 10500 10000 9500 9000 8500 8000 7500 7000 6500 6000 5500 5000 4500 4000 3500 Elements: Slope: Offset: Correlation: R2(Pearson): R-Square: RMSECV: SECV: Bias: 10 0AaIII 14 0AaI 41 0AaIII 12 0AaII 42 0AaI 14 0AaI 14 0AaI 38 0AaIII 41 0AaIII 12 0AaII 10 0AaIII 44 0AaII 10 420AaIII 41 0AaIII 38 0AaIII 31 0AaI 32 0AaII 225 0,930308 1,2582e+03 0,8527921 0,7272543 0,6247552 1,3256e+03 1,6483e+03 561,99744 12 14 0AaII 0AaI 10 0AaIII 31 0AaI 32 0AaII 38 0AaIII 33 2Ba 31 0AaI 32 0AaII 44 0AaII 12 0AaII42 0AaI 44 0AaII 32 0AaII 38 0AaIII Predicted vs. Reference 43 2BbvIII 15 2Ba16 19 17 2BbvI 2Ab 4 2Aa 5 2BbIII 7 39 2BbIII 192BbII 2BbvIII 33 2Ba 50 2BbvI53 2Ab 47522BbII 2BbI 37 1 2BbvI 5 2BbIII 36 2Aa 2527 2BbIII 2BbI 15 92Ba 2Ab 33 7 16 2BbI 2BbII 443 2Aa 2BbvIII 26 2Ab 30 362Ba 2Aa 22 35 35 37 2BbII 2BbvIII 36 272Aa 2BbI 17 2BbvII 19 1Ba29 48 2Aa 5037 28 2BbvIII 57 2BbvII 5253 1AaIII 2Ab1AaI 5635 2Ba 2BbvII 392BbvIII 2BbIII 47 2BbII 21 2BbvI 56 2Ba 21 2BbvI 22 2BbII 34 1Ab 51 1AaIII 48 132Aa 1Ab 33 2Ba 25 2BbIII 22 472BbII 40 26 2Ab 6 1AaII 1 1AaII 2BbvI 11 1Bb 11 1Bb 26 2Ab 24 1Bb 57 2BbvII 54 1Bb 3 1Aav 3 1Aav 46 1Aav 5 2BbIII 25 2BbIII 1913 62BbvIII 1Ab 40 1AaII 8 2 18 1Ba 28 1AaI 15 2Ba 23 424 2Aa 1Aav 1Bb 34 9 2Ab 1Ab 27 45 1Ba 131AaIII 820 11 1Ab 23 45 1Aav 1Ba 55 46 1AaI 511Aav 1AaIII 49 54 1Ab1Bb 1AaIII 1AaII 1Bb49 1Ab 621AaII 17 2BbvII 1856 1Ba 20 2Ba57 29 21 1AaIII 8 1AaIII 302BbvII 1Ba34 1Ab 39 2BbIII 7 2BbI 43 2BbvIII 48 2Aa 23 1Aav 29 2 3 1Aav 2050 1AaII 2BbvI 401AaIII 2852 2BbI 16 2BbII 18 1Ba 531AaI 30 1Ba 24 55 2Ab 1Bb 1AaI 45 1Ba 49 1Ab 54 1Bb 46 51 1Aav 1AaIII 48 2Aa 19 2BbvIII 39 2BbIII 4 2Aa5 2BbIII 9 2Ab 15 2Ba 16 2BbII 2 1AaI 53 2Ab 56 52 2Ba 2BbI 47 2BbII 17 1 2BbvI 7 2BbvII 2BbI 8 3 1AaIII 1Aav 13 6 1Ab 1AaII 11 1Bb 4650 43 2BbvIII 3745 1Aav 2BbvI 57 2BbvII 2BbvIII 40 1AaII 55 1AaI 1Ba 51 1AaIII 54 1Bb 18 49 1Ba 1Ab 35 2BbvII 36 2Aa 22 2BbII 25 2BbIII 24 21 1Bb 2BbvI 26 27 2Ab 2BbI 20 1AaII 23 1Aav 28 1AaI 29 1AaIII 30 1Ba 34 1Ab 31 0AaI 3000 Reference Y (Skörd kg/ha, Factor-3) 4000 5000 6000 7000 8000 9000 10000 11000 12000 13000 55 1AaI
Results 2017: Modelling yield in winter winter wheat Using 2 field trials, Bjertorp (n= 40) and Skofteby (n= 64), and a multispectral sensor equipped drone Bjertorp: good prediction of yield was achieved with all wavebands, RGB, the indices ChII and NDRE. Skofteby: good prediction of yield was achieved with all wavebands Regression models (PLS and SLR) Response: Yield (kg/ha) Predictors: 5 wave bands (B,G,R, Rededge, NIR) Indices (NDVI, ChII, NDRE, MSAVI 2) Cross validation of N-treatments Bjertorp 2017 r2 rmsecv rpd Wave bands (5) 0,91 668 3,3 RGB 0,84 897 2,4 NIR 0,41 1681 1,3 Rededge No correlation! Indices: NDVI 0,51 1532 1,4 ChII 0,80 972 2,2 NDRE 0,76 1064 2,0 MSAVI2 0,63 1333 1,6 Cross validation of N-treatments Skofteby 2017 r2 rmsecv rpd Wave bands (5 ) 0,92 601 3,6 Without 0N 0,80 553 3,9 RGB No correlation! NIR 0,7 1172 1,9 Indices NDVI 0,69 1339 1,6 ChII 0,26 2058 1,1 MSAVI2 No correlation! Prediction of Bjertorp: Model: Skofteby Wave bands (5) 0,79 1302 1,7
Bjertorp 2017 All wave bands and cross validation with N-treatments (3) Predicted vs. Reference 13500 13000 12500 12000 11500 11000 10500 Elements: Slope: Offset: Correlation: R2(Pearson): R-Square: RMSECV: SECV: Bias: 57 0,9481089 448,17334 0,9537517 0,9096423 0,9521046 668,8598 666,6731-103,5271 10000 9500 9000 8500 8000 7500 Predicted Y (Skörd, Factor-5) 7000 6500 6000 5500 5000 Reference Y (Skörd, Factor-5) 5000 6000 7000 8000 9000 10000 11000 12000 13000 14000
Skofteby 2017 All wave bands and cross validation with N-treatments (16) Predicted vs. Reference 16000 15000 14000 13000 12000 11000 Elements: Slope: Offset: Correlation: R2(Pearson): R-Square: RMSECV: SECV: Bias: 64 0,9448155 654,89075 0,9615323 0,9245444 0,9331001 601,4043 606,15411-2,320618 10000 9000 8000 7000 6000 5000 Predicted Y (Skörd, Factor-5) 4000 3000 2000 1000 Reference Y (Skörd, Factor-5) 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 12000 13000 14000 15000 16000
Results 2016: Modelling protein, energy and NDF in timothy grass leys No good regression models could be made when using UAV-pictures from 3 fields with 5 samples in each at 3 harvest dates, n= 45 or a field experiment (n= 70). Satellite pictures not analyzed yet. Visual differences between treatments in field trials were well described by e.g. NDVI 8/7 2016 (UAV) and IR/R och Si1 15/6 2017 (handheld N-sensor)! NDVI, timotej-försök 2016 9 Index: IR/G, timotejförsök 2017 NDVI Fröskörd kg/ha 8 7 0,82 0,80 0,78 0,76 0,74 0,72 0,70 70 100 130 160 190 130 160 160 190 220 1200 1000 800 600 400 200 0 Index-värde 6 5 4 3 2 1 0 70 100 130 160 190 130 160 160 190 220 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 giva vår 2 givor vår 2 givor 3 givor
Results 2017: Modelling protein, energy and NDF in timothy grass leys Using a field trial in a timothy grassley with increasing N-rates (n= 40) and crop sampling at three harvest dates (15/6, 27/6, 10/7, n=120). Good prediction of yield was achieved with all wavebands, RGB, the indices GNDVI and ChII (two dates n= 80) The results from the third harvest date have not yet been analysed. Regression models (PLS and SLR) Response: Yield (kg/ha) Predictors: 5 wave bands (B,G,R, Rededge, NIR) Indices (GNDVI, ChII) Two harvest - dates, n=80 Crossvalidation (N-treatments) Protein: r2 RMSECV rpd Wavebands (5 ) 0,89 7,0 3,1 RGB 0,88 7,3 3,0 NIR 0,74 11,1 1,9 Indices: GNDVI 0,79 9,85 2,2 ChII 0,79 9,85 2,2
All wavwbands and crossvalidation of N-treatments (10) Predicted vs. Reference 150 145 140 135 130 125 120 115 110 105 100 95 90 Elements: Slope: Offset: Correlation: R2(Pearson): R-Square: RMSECV: SECV: Bias: 80 0,9012116 9,8190861 0,9452664 0,8935285 0,9020361 7,0486441 7,0918074 0,135352 Predicted Y (RåproteinNIR, Factor-5) 85 80 75 70 65 60 55 Reference Y (RåproteinNIR, Factor-5) 40 50 60 70 80 90 100 110 120 130 140 150 160
GNDVI and cross validation with N-treatments (10) 135 Predicted vs. Reference 130 125 120 115 110 105 100 Elements: Slope: Offset: Correlation: R2(Pearson): R-Square: RMSECV: SECV: Bias: 80 0,8178606 17,702501 0,8903896 0,7927936 0,8087648 9,8481894 9,9091473-0,1517141 95 90 85 80 Predicted Y (RåproteinNIR, Factor-1) 75 70 65 60 55 50 45 Reference Y (RåproteinNIR, Factor-1) 50 55 60 65 70 75 80 85 90 95 100 105 110 115 120 125 130 135 140 145 150
Projektets finansiärer Förutom Interreg Öresund-Kattegatt-Skagerack har projektet även följande finansiärer: promilleafgiftsfonden för landbrug