Crop Type & Yield Mapping:
Preparatory study for Sentinel 2
Sentinel 2 for Science Meeting - 20-22 May 2014
N.M. Knox; L.T. Tsoeleng; C. Adjorlolo; T. Newby
Free State Agricultural production • 2 million ha under agricultural production
• 40% of South Africa's total maize crop,
• 50% of wheat,
• 80% of sorghum,
• 33% of potatoes,
• 18% of red meat,
• 30% of groundnuts
• 15% of wool.
Crop Calendar
• Free State Province
Crop Type Modelling
Sentinel 2 for Science Meeting - 20-22 May 2014
Crop Type Modelling
• Focused on 5 primary crop types
• Maize, Sunflower, Soy beans, Pasture, Fallow
• Simple models based on NDVI and EVI time series
• Based on National Crop Estimates Committee annual aerial samples
• Calibration – 132 samples (balanced)
• Validation - 153 samples (heavily weighted on maize)
• 16 Cloud free BOA Spot4-Take5 scenes 31 Jan – 15 Jun
• Curve fitting and least squares fit
In-situ Data
• National Crop Estimates Committee Data
• Annual PICES aerial survey carried out. – Applicable to commercial not subsistence farming
– Approx 300 random sample points
• Only within field boundary layer
– Crop type identified
– Subset of the maize crop sites selected for yield estimation
– Yield estimation used to generate provincial yield estimate
• Annually updated field boundary layer
Maize
Sunflower
Soy Bean
Pasture
Fallow
NDVI
EVI
Sunflower 68% Maize
Maize 29% Pasture
Fallow 75 %Maize
Pasture 41 % Maize
30% SFlower
Soy bean 70% Maize
Limitations & Way Forward
• Late start in season prevents crucial sowing/green-up phase of crop
• Limited sample points of certain crops
• Result of random sampling and maize prevalence
• Heterogeneous in-field variability of crops
• Capability of expanding study to make use of greater spectral information
• Integration of climate data and crop calendar (restrict date inclusion past harvest)
Maize Yield Modelling
Sentinel 2 for Science Meeting - 20-22 May 2014
Maize Yield
• Small data set
• Maize yield derived from precision agriculture – during harvesting
• 3 Fields obtained through GrainSA Co-operative from a farmer in Bothaville region – hardcopy digistized and classified
• Yield ranged from 0-13 tonnes/ha
• Data set:
• Calibration data set – 314 stratified random sample
• Validation set – 155 random samples
• Applied to NDVI and EVI 16 date Spot4-Take5 layer stack
• Random forest modelling
SANSA’s role in earth observation R2+/-RMSE = 0.51+/-1.213
Measured yield (t/ha)
Pre
dic
ted y
ield
(t/
ha)
Variable Importance
Way Forward to Sentinel 2
• Present initial findings to farmers in region - July
• Request additional farm yield data in digital format
• Integrate data from the Automatic Weather Stations
• Evaluate effect of time series (5 vs 10 day intervals)
• Cloud masking
• Area mapping/field boundary discrimination
• Integrate spectral rather than simple indices for models
Thank you!
Questions
Sentinel 2 for Science Meeting - 20-22 May 2014