University of Twente Student Theses


Mapping nitrogen status in rice crops using unmanned aerial vehicle (UAV) data, multivariate methods and machine learning algorithms

Ma, Fanshu (2020) Mapping nitrogen status in rice crops using unmanned aerial vehicle (UAV) data, multivariate methods and machine learning algorithms.

[img] PDF
Abstract:Rice as one of the most important food crops feeds more people than any other crops in the world. Nitrogen (N) is an essential element during the rice-growing stages which affect rice yield and production. Although raised N application is used to increase the yield, in order to meet the demand for food, excess of its application would cause a series of environmental problems and even would decrease yield. Therefore, estimating nitrogen in rice is important to precision N application, environmental pollution reduction, and global carbon and N cycle. This study aims to use time-series multispectral Unmanned aerial vehicle (UAV) data and field observations of nitrogen together with multivariate methods and machine learning algorithms for estimating and mapping of nitrogen in different rice-growing seasons. The study area is in IRRI (International Rice Research Institution) experimental fields in Los Baños, the Philippines. Rice N was measured in 2016 early wet season (EWS) and 2017 dry season (DS) destructively (referred to as tissue analysis) from different parts of crops (stem, grain and whole plant parts) at the end of the growing season. SPAD and leaf colour chart readings (LCC) (as nitrogen proxies) were obtained nine times during the growing seasons whereas, other relevant measurements, such as leaf area index (LAI) was measured four times during the growing seasons. Further, SPAD and LAI values were used to calculate the canopy chlorophyll content (CCC). The relationships between SPAD, LCC and tissue analysis of the whole plant parts (referred to as plant nitrogen accumulation) (PNA) were firstly explored to understand the relationship between nitrogen measurements obtained destructively and those nitrogen proxies obtained non-destructively. In order to choose the best vegetation index (VI) for N estimation, the correlation coefficients between VIs and field measurements (PNA, SPAD and CCC) were examined. PNA was then used for further analysis, and the VI, which had the highest correlation coefficient with PNA was used in simple linear and stepwise regression models for PNA estimation. The partial-least square regression (PLSR), support vector machine (SVR) and random forest (RF) were then compared for PNA estimation using R2, RMSE and NRMSE between measured and estimated PNA. Finally, the most accurate algorithm was used for mapping rice PNA. Results are as follows, 1) Strong correlations were observed among the PNA, SPAD and LCC in the rice panicle initiation and heading stages; 2) The GNDVI derived from the multispectral UAV images was the best performing VI for PNA estimation in both seasons; 3) comparison between simple linear and stepwise regressions revealed that using simple linear regression models (SR) and GNDVI from rice panicle initiation and heading stages are sufficient for PNA estimation; 4) among the machine learning algorithms, the RF was the most accurate machine learning algorithm for PNA estimation in 2016EWS (R2=0.9, RMSE=8.37, NRMSE=10.9%) and 2017DS (R2=0.93, RMSE=9.93, NRMSE=8.1%); 5) PNA estimation maps were generated for the whole study site using the RF model in both seasons. Further investigation for more accurate N status based on rice hills level and different input for machine learning algorithms could be examined in future studies.
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