Evaluation of benefits and constraints for the use of Sentinel-2 for agricultural monitoring

Similar documents
Transcription:

23/05/2014 Evaluation of benefits and constraints for the use of Sentinel-2 for agricultural monitoring Isabelle Piccard, Herman Eerens, Kris Nackaerts, Sven Gilliams, Lieven Bydekerke Sentinel-2 for Science Workshop, 20-22 May 2014, ESA-ESRIN, Frascati (Italy)

Evaluation of the use of Sentinel-2 for agricultural monitoring» Studied aspects:» Spatial resolution» Temporal resolution» Evaluation based on time series of DMC-1/2 and Deimos-1 images (22m resolution) over Belgium (Flanders) 23/05/2014 2

Spatial aspects» Belgium: small agricultural parcels (very) high resolution imagery required for crop monitoring Study area in E-Belgium 23/05/2014 3

Spatial aspects» MODIS provides insufficient detail <10% 10-20% 20-30% 30-40% 40-50% 50-60% 60-70% >70% Crop area fraction (%), derived from IACS data Winter wheat Example: MODIS vs. DMC derived anomaly maps for winter wheat, 2012, municip. Hoegaarden Sugar beets fapar return period MODIS 21-31 May 2012 MODIS 21-31 July 2012 Potatoes DMC 27 May 2012 DMC 25 July 2012 Maize 23/05/2014 4

Temporal aspects» Optical imagery: cloud cover» 10-day revisit time seems insufficient» MULTI-SENSOR approach: combination of S2, Landsat, DMC/Deimos, intercalibration!» Number of cloud free DMC/Deimos observations, 2009-2013 (intensive acquisition campaigns in 2011 & 2012; other years: archive data) 2012 2009 2010 2011 2013 23/05/2014 5

Continuous time series required! Gap filling: how?» Combination of different optical HR sensors: intercalibration» Combination of optical HR and MR sensors: resampling & intercalibration or using data fusion techniques (eg. STARFM and derived methods)» Combination of optical and SAR sensors» Combination of optical HR data with a crop growth model + = MODIS: - high frequency - low resolution Landsat: - low frequency - high resolution 23/05/2014 6

Phenology monitoring» Continuous time series temporal profiles» Phenology detection (calibration with field observations)» Development stage field practices / planning & logistics» Input for crop growth modelling» Control purpose (e.g. harvest date early potatoes ) Potato parcels, 2012 (IACS data) Declared earlies Declared earlies Not OK OK 23/05/2014 7

Crop monitoring» Common practice for low / medium resolution imagery:» Selection of EO index, preferably biophysical parameters» Comparison of actual values with long term average (LTA): absolute or relative differences, z-score, return period, VCI, VHI, VPI,» Similar approach for high resolution imagery 23/05/2014 8

High resolution Biophysical parameters» Algorithm for DMC/Deimos developed by INRA Source: INRA Source: INRA 23/05/2014 9

High resolution Biophysical parameters» Validation:» DHP measurements» N-Belgium» Maize & potatoes, 2012 23/05/2014 10

DMC fapar Building a high resolution long term average» DMC long term average (LTA) derived from MODIS (spectrally similar)» Comparison DMC-fAPAR vs. MODIS-fAPAR at 250m resolution (same day, equal sampling intervals) equation» unmixing of MODIS-S10 fapar [2003-2013] per crop & per municipality MODIS LTA statistics MODIS LTA image at 25m resolution equation DMC LTA image R²=0.92 fapar DMC = 0.0312 + 0.763 * fapar MODIS MODIS fapar DMC fapar, long term average 1st dekad of June (winter wheat) 23/05/2014 11

Building a high resolution long term average» Validation: comparison of MODIS based LTA per municipality with actual DMC LTA (2009-2013) for winter wheat MODIS derived LTA < actual DMC LTA MODIS: mixed pixels, influence summer crops MODIS derived LTA > actual DMC LTA 23/05/2014 12

Crop monitoring & anomaly detection» Monitoring at parcel level: comparison of actual and average index values» Intra-field variability: can be improved with 10m S2 data» Inter-field variability 14 May 25 May 27 May 28 May 24 July 25 July 10 Aug 18 Aug 29 Aug 4 Sept 7 Sept 8 Sept 16 Sept 30 Sept 15 Oct Example: potato monitoring with DMC/Deimos at 25m resolution based on fapar return period (May-Oct 2012) Field 140 (high yield: 69 ton/ha) 23/05/2014 13

Crop monitoring & anomaly detection» Validation: comparison with MODIS, pure pixels Min. 20% winter wheat 23/05/2014 14

Yield estimates» Development of yield models based on time series of HR data (example: potatoes) Crop harvester with integrated Probotiq Yield Master Pro sensor Spatial variability of yield for single field 23/05/2014 15

Other examples» Approach also tested in other regions:» Ethiopia (ISAC & Spot4Take5)» Morocco (Spot4Take5)» Input: DMC, Landsat, SPOT-4, RapidEye NDVI images & land cover maps» LTA computed based on MODIS, only resampled & translated to HR» Anomaly maps computed & validated with field data» Check out our poster! 23/05/2014 16

Conclusions» S2 benefits:» Improved spatial (and spectral) resolution» S2 constraints:» 5-10 day revisit time may be too long, especially in cloud-prone areas!» S2 opportunities for crop monitoring at parcel level!» Calculation of biophys. par s & high resolution LTA (has limitations )» Continuous time series Multi-sensor approach required need to invest in gap filling & intercalibration!» Not mentioned but equally important for (HR) time series analysis:» Geometric correction make sure basic products are OK» Atmospheric correction make sure basic products are OK» Accurate cloud & shadow detection: not easy to automate 23/05/2014 17

Thank you! 23/05/2014 18