Estimation of winter wheat yield using time series of airborne hyperspectral data

0570654 - ÚVGZ 2024 eng A3 - Přednáška/prezentace nepublikovaná
Homolová, Lucie - Pikl, Miroslav - Švik, Marian - Janoutová, Růžena - Slezák, Lukáš - Veselá, Barbora - Klem, Karel
Estimation of winter wheat yield using time series of airborne hyperspectral data.
[12th EARSeL Workshop on Imaging Spectroscopy in Potsdam. Potsdam, 22.06.2022-24.06.2022]
Způsob prezentace: Prezentace
Pořadatel akce: University of Greifswald
URL akce: https://is.earsel.org/workshop/12-IS-Potsdam2022/ 
Institucionální podpora: RVO:86652079
Klíčová slova: imaging spectroscopy * yield * machine learning * winter wheat
Obor OECD: Agriculture
https://is.earsel.org/workshop/12-IS-Potsdam2022/

Wheat (Triticum spp.) is one the most important cereals produced worldwide, as well as in Europe. Accurate and timely assessment of crop yields and its spatial variability within a field can help to optimise fertilisation application and water management. The main objective was to evaluate airborne hyperspectral data from three consecutive growth seasons for prediction of winter wheat yields. We aimed to evaluate different machine learning methods and contribution of acquisition days.
For this study we selected two sites that are part of the Czech Republic long-term crop rotation experiment: Ivanovice (49°18’40’’N, 17°05’45’’E, 225 m.a.s.l) and Lukavec (49°33’23’’N, 14°58’39”E, 620 m.a.s.l). At both sites, the same fertilisation experiment design was applied on winter wheat plots (12 combinations of organic and mineral fertilisation in four replicas, 48 subplots in total). The average yield per subplot varied between 2.25 and 10.01 t/ha. The average yield was 5.4 t/ha in 2019, 7.5 t/ha in 2020 and 7.3 t/ha in 2021. Yields at Ivanovice were generally higher (4.55 – 10.01 t/ha), which is likely due to a lower elevation than at Lukavec (2.25 – 9.10 t/ha).
Airborne hyperspectral data were acquired several times during three consecutive vegetation seasons: 2019 (5 acquisitions for Ivanovice and 3 acquisitions for Lukavec ), 2020 (6 acquisitions at both sites) and 2021 (4 acquisitions for Ivanovice and 3 acquisitions for Lukavec) using CASI and SASI spectroradiometers (Itres, Canada) on board Flying Laboratory of Imaging Systems (https://olc.czechglobe.cz/en/flis-2/). For this study we evaluated only the visible and near infrared CASI data with 48 bands between 383 and 1053 nm with the spectral step of 14.25 nm and the spatial resolution was 0.5 m. The images were corrected for radiometric, geometric and atmospheric effects. Average spectral signatures were extracted for each wheat subplot (5 x 5 m) for each image acquisition.
Yields were estimated from the spectral data using machine learning methods available in the ARTMO toolbox. Data were divided into two parts, 70% of the data were used for model training and 30% for validation.
A pooled model, when all acquisitions dates were combined together provided promising results. We tested five machine learning methods, namely canonical correlation forests (CCF), gaussian processes regression (GPR), support vector regression (SVR), least square linear regression (LSLR) and partial least square regression (PLSR). All methods provided similar results reaching R2 > 0.8 and RMSE < 0.67 t/ha. In the validation process the GPR outperformed other tested methods (R2 = 0.82, RMSE = 0,57 t/ha). A GPR model that combined only the acquisitions from the end of April and beginning of May suggested that wheat yield could be accurately estimated with R2 = 0.89 and RMSE = 0.55. Those promising results from experimental plots, however, should be further verified across the years and other localities from the long-term experiment. Consequently, the best model should be tested at a larger scale in real production conditions of a farm.

Trvalý link: https://hdl.handle.net/11104/0341974