Development of an improved prediction model of oil palm yield using advanced geospatial and machine learning techniques

The accurate prediction of oil palm yield is essential to ensuring the sustainable production of oil palm for the purposes of food security and economic return. Due to the fact that crop yield varies depending on a variety of factors such as weather, nutrients, and management practices, developing a tool to predict oil palm yield in a timely and effective manner is essential in order to prevent climate risk and ensure food security, especially in the face of climate change and the acceleration of extreme climatic events. Therefore, this study aims to develop yield prediction models based on the remote sensing data and machine learning approaches. First, a comprehensive mapping of oil palm at the state level was conducted, followed by yield prediction using time-series satellite imageries from Landsat-8. A combination of vegetation indices and a relief algorithm with linear projection was employed to select the best combination of spectral indices for oil palm discrimination. It was classified using random forest and modified AdaBoost algorithms. A time-series approach known as walk-forward validation was used to train the model for 2016–2019 data, and one-step prediction was implemented for 2020 using random forest and Adaboost based on Landsat NDVI time series imageries. Then, machine learning algorithms were used to predict oil palm yield at the block level at a research plantation using MODIS-NDVI and Landsat-NDVI. In addition to one variable, several variables, such as land surface temperature (LST), and CHIRPS, weather stations, and field-surveys from multiple sources, were selected to develop a generic yield prediction model using machine and deep learning algorithms. Lastly, the best yield prediction model was deployed to develop a web application with an interactively designed graphical user interface as a production for plantation management. In terms of state level, the proposed method showed that the RF model (RMSE = 0.384; MSE = 0.148; MAE = 0.147) outperformed the AdaBoost model (RMSE = 0.410; MSE = 0.168; MAE = 0.176). In terms of blocks, the Landsat-7 NDVI revealed that neural networks with a deeper network topology (R2= 0.85; RMSE = 1.42 tonnes per hectare; MAE = 0.57 tonnes per hectare; MAPE = 0.06 tonnes per hectare) achieved the highest accuracy among the models. Using multisource data, it provides the highest accuracy with the proposed deep neural network architecture. With backward elimination, deep neural networks consistently achieved the highest prediction accuracy compared with the other models, with a 14% increase in R2, a 11% increase in RMSE, a 32% decrease in MAE, and a 1% decrease in MAPE. Following the improved accuracy of the machine learning and deep learning algorithms used, the proposed deep neural network model was deployed for developing a web application with a timely prediction capability that can act as decision tools to maintain the efficiency and productivity of plantations.

Text FK 2023 26 - Declaration Form.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (1MB) Text FK 2023 26 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (12MB) Text FK 2023 26.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (12MB) Official URL or Download Paper: http://ethesis.upm.edu.my/id/eprint/18942

Actions (login required)

View Item