Time series analysis for spring barley phenology monitoring using Sentinel 2 – A case study of Southern-central Sweden

Abstract: Monitoring crop phenology at parcel scale aligns with the concept of precision agriculture (PA) and can provide invaluable information to agronomic management systems. Satellite time series data are commonly deployed for detecting crop growth stages, while the recent advancements in remote sensing (RS) technologies such as the launch of Sentinel 2 (S2) are providing unprecedented opportunities for crop monitoring. In this thesis, the focus is on spring barley parcels with available in-situ crop growth stage measurements (recorded in Zadoks scale) located in south-central Sweden over the period of 2017 – 2019. More specifically, the aim was to detect three specific crop growth stages of spring barley that are crucial for applying external inputs (e.g. fungicides applications) named according to the Zadoks Scale: (i) first node detectable (31DC) (ii) flag leaf ligule just visible (39DC) and (iii) first spikelet of inflorescence just visible (51DC). This thesis describes a simple empirical approach based on Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index 2 (EVI2) S2 time series. TIMESAT 4.0 was deployed to reconstruct the S2 NDVI and EVI2 trajectories using the Double Logistic (DL) smoothing function. Moreover, the available in-situ crop growth stage measurements were utilized to optimize the dynamic thresholds (% of the amplitude of the season) for detecting the different crop growth stages of interest (31DC, 39DC, 51DC). Two types of thresholds were conceptualized and optimized: the (i) local threshold and the (ii) global threshold. The optimal local threshold for each crop growth stage of interest refers to the threshold that created an agreement between the vegetation index (NDVI and EVI2) results and the in-situ measurements for each spring barley field individually. The global threshold refers to the optimal threshold that resulted in the smallest Root Mean Square Error (RMSE) between in-situ measurements and S2 derived results when applied on all the studied spring barley parcels. The optimal local thresholds showed high variability especially for the crop growth stage 31DC, where the standard deviation (SD) of the local threshold values was 16.1% (NDVI) and 22.1% (EVI2). The variability of the optimal local thresholds showed a decreasing trend with latter crop growth stages, where for the stage 51DC the SD was 5.9% (NDVI) and 3.1% (EVI2). According to the results of the global threshold optimization, the optimal thresholds for the stages 31DC, 39DC and 51DC based on NDVI were 74%, 92%, and 99% respectively, where for EVI2 the optimal global thresholds were 70% (31DC), 91% (39DC) and 99% (51DC). When applying the optimized global thresholds, the RMSE of the retrieved dates for the different crop growth stages against the in-situ measurements was smaller than 7.6 days (for both NDVI and EVI2). EVI2 consistently outperformed NDVI regarding all the crop growth stages of interest (31DC, 39DC, 51DC) where it resulted in lower RMSE (5.1 days, 4.8 days, 4.2 days) and higher coefficient of determination (R-square; 0.43, 0.45, 0.69) compared to the NDVI induced RMSE (6.9 days, 7.6 days, 7.2 days) and R-square (0.33, 0.27, 0.39). The results showed the feasibility of using S2 data in crop phenology studies and demonstrated its potential uses and inaccuracies regarding the detection of three specific crop growth stages of spring barley that are of interest in agronomic decision making.