Abstract
This research is an outcome of the R&D activities of Ecodevelopment S.A. (steadily supported by the Hellenic Agricultural Organization—Demeter) towards offering precision farming services to rice growers. Within this framework, a new methodology for topdressing nitrogen prediction was developed based on machine learning. Nitrogen is a key element in rice culture and its rational management can increase productivity, reduce costs, and prevent environmental impacts. A multi-source, multi-temporal, and multi-scale dataset was collected, including optical and radar imagery, soil data, and yield maps by monitoring a 110 ha pilot rice farm in Thessaloniki Plain, Greece, for four consecutive years. RapidEye imagery underwent image segmentation to delineate management zones (ancillary, visual interpretation of unmanned aerial system scenes was employed, too); Sentinel-1 (SAR) imagery was modelled with Computer Vision to detect inundated fields and (through this) indicate the exact growth stage of the crop; and Sentinel-2 image data were used to map leaf nitrogen concentration (LNC) exactly before topdressing applications. Several machine learning algorithms were configured to predict yield for various nitrogen levels, with the XGBoost model resulting in the highest accuracy. Finally, yield curves were used to select the nitrogen dose maximizing yield, which was thus recommended to the grower. Inundation mapping proved to be critical in the prediction process. Currently, Ecodevelopment S.A. is expanding the application of the new method in different study areas, with a view to further empower its generality and operationality.
Highlights
Nitrogen (N) fertilization is by far the most vital input in rice growing for ensuring high yields
Using the Root Mean Square (RMS) error between the predicted and observed yield values in the relevant years (2017–2019) for all the tested models, it was found that the XGBoost model resulted in the lowest error compared to the other three models (0.73 tn ha−1 ) provided, though, that DATE and leaf nitrogen concentration (LNC)*DATE parameters were not included in the algorithm; otherwise, RMS error was found quite high in all cases
The XGBoost model was selected for two reasons: (a) it was found more appropriate for the specific dataset, and (b) it performed slightly better than the other three models in the case of DATE inclusion (Figure 7)
Summary
Nitrogen (N) fertilization is by far the most vital input in rice growing for ensuring high yields. The XGBoost is a decision tree-based algorithm, which gives state-of-the-art results in many machine learning problems, as it uses the gradient boosting framework and focuses on areas within the dataset of high error for improving the overall initial predictor [11]. This technique is dynamic as it leaves out the data that fit in the weak models and focuses on the development of new models that can better deal with the remaining data [35]. A key technique that must be included in the model for addressing this problem is the dropout technique and the adjustment of the early stopping rounds parameter
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call