University of Twente Student Theses


Yield monitoring with Sentinel-2: A first assessment for The Netherlands

Chamnan, Kingkan (2021) Yield monitoring with Sentinel-2: A first assessment for The Netherlands.

[img] PDF
Link to full-text:
(only accessible for UT students and staff)
Abstract:Yield reductions have become more common for the Netherlands in recent years due to a changing climate and more extreme weather events. Various studies applied satellite image time series for timely prediction of yield. Although for larger regions, mostly coarse-resolution data were used, that do not allow observing smaller individual fields without contamination from field surroundings. The primary objective of this study is to assess if Sentinel-2 timeseries allow for an accurate yield prediction at a regional level. Following an assessment of maize area and yield variability, silage maize was selected for this study, given that it is the crop with the largest harvest area in the Netherlands and substantial yield variability in the past five years. Using the full Sentinel-2 level 1C record available in Google Earth Engine, clouds and shadows were masked, and monthly-maximum vegetation index (VI) composites were created to further reduce remaining atmospheric effects. Given that yield data were available at the province level (12 provinces in the Netherlands), the monthly VIs were spatially aggregated for all maize fields within each province. Annual location and spatial extent of maize fields were obtained from a Dutch reference dataset, and a 10m internal buffer was used on each field to avoid edge effects. The aggregated VIs were then regressed against statistical yield. From the two VIs tested, i.e. NDVI and EVI, EVI correlated more strongly with yield, probably due to the saturation of NDVI for high biomass conditions. The strongest correlation was obtained for the cumulative EVI from July to August (EVIJA), which is during the maturity stage of maize. The linear regression model revealed that EVIJA retrieved from Sentinel-2 could explain 76% of the yield variation (R²=0.76, RMSE=2.26 (x1000kg-ha)). Cross-validation revealed that for the major maize producing provinces like Noord-Brabant, Gelderland, and Overijssel, yield prediction was most accurate with an RMSE of 0.91, 0.98, and 1.25 (x1000kg-ha), respectively. Taking only these three provinces, more variability in yield could be explained (R²=0.95) as compared to the other nine provinces with smaller maize areas (R²=0.72). When grouping provinces differently, more accurate predictions could be made for those provinces where maize is predominantly cultivated on sandy soil (Drenthe, Overijssel, Noord-Brabant, Gelderland, Groningen, Friesland, and Limburg). Nonetheless, municipal level anomaly maps of EVIJA reveal substantial within-province variability of maize performance, suggesting the possibility to further improve yield prediction by better accounting for this spatial variability. Given the continuing acquisition of Sentinel-2 imagery for years to come, this study demonstrated its potential in accurate regional yield estimation.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page