Process-based crop models calibration and structure affect data assimilation for estimating sugarcane yield

• Ano de defesa: 2022
• Autor(a) principal: Fattori Junior, Izael Martins
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Saccharum officinarum L.; Calibração; Calibration; DSSAT; Índice de área foliar; Leaf area index; Remote sensing; Sensoriamento remoto; WOFOST
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/11/11152/tde-15092022-153908/

Sugarcane is an important feedstock of sugar and ethanol. Thus, strategies to follow the quantity and the availability of sugarcane are essential. Different methods have been developed for yield estimation and the use of process-based crop models (PBM) with data assimilation (DA) stands out. Due to the capability of using two different sources of information, and their respective uncertainty for crop yield estimation. However, inconsistencies between PBM and the assimilated data were reported in the literature, which led to systematic errors and low performance of the simulations. Such limitation was connected to the absence or poor calibration of PBMs and to the simplification of the PBMs structure to represent crop development traits. Thus, this study aimed to evaluate the impact of using one sugarcane cultivar-based calibration on other cultivars with DA methods and compare this source of uncertainties with others. Moreover, evaluate how the difference between two PBM, regarding their structure of specific sugarcane traits, affects the performance of DA methods. For that, firstly, the DSSAT/SAMUCA (DS) was used to simulate 22 field experiments and quantify the impact of using one cultivar specific calibration (cv. RB867515) compared to four non-calibrated cultivars (cv. NCo376, SP832847, R570, RB72454), on stalk fresh yield (SFY) predictions. This was performed for three different DA methods, Ensemble Kalman filter (EnKF), Ensemble smoother (ES), and Weighted mean (WM) to assimilate leaf area index (LAI) retrieved from field observation and compared to the PBM simulation without DA (Open-Loop, OP). Moreover, we analyzed the influence of the timing and amount of LAI data, to compare with the impact of calibration. Second, two different PBM, in terms of structure, one more detailed in terms of structure (DS) and other more general (WOFOST, WO), were compared to the performance of simulate SFY with the use of EnKF and LAI. The LAI was retrieved from Landsat 7 ETM+ and 8 OLI, from fields of a sugarmill database. Thus, the simulations with the EnKF of these fields were compared to the OP simulations. The results showed that the use of a genotype-specific calibration had substantially higher accuracy compared to non-calibrated, for the three DA methods. The simulation of non-calibrated cultivar experiments had a higher accuracy increase, for EnkF and ES, however, WM had opposite results. In this regard, the accuracy of the simulations with DA had a high correlation OP simulations accuracy, which was higher than the correlation with the number of LAI observations assimilated. Furthermore, our results indicated that the calibration performance and the structure of the PBMs influenced the OP simulations, with DS showing higher performance, compared to WO. However, with DA the performance was limited by the inconsistency between Landsat LAI and the LAI simulated by the PBMs, despite the improvements. Thus, assimilated Landsat LAI had the potential to improve yield estimation, but the better descriptions of DS did not inhibit the error inconsistency. Therefore, this study emphasized that the use of DA required previously calibrated PBMs regarding cultivar traits to ensure a higher performance. In addition, more detailed PBMs in terms of process description can benefit from their detailed description to improve the performance of OP simulations and consequently with DA.