|Title:||Machine learning approaches for rice crop yield predictions using time-series satellite data in Taiwan||Authors:||Nguyen-Thanh Son
|Issue Date:||Oct-2020||Publisher:||Taylor & Francis||Journal Volume:||41||Journal Issue:||20||Start page/Pages:||7868-7888||Source:||International Journal of Remote Sensing||Abstract:||
Rice is the most important food crop in Taiwan, directly feeding more than 23 million people in the country. Information on rice production is thus crucial for crop management and food policymaking. This study aims to develop a machine learning (ML) approach for predicting rice crop yields in Taiwan using time-series Moderate Resolution Imaging Spectroradiometer (MODIS) data. We processed the data for the period from 2000 to 2018, following three main steps: (1) data pre-processing to generate smooth time-series Normalized Difference Vegetation Index (NDVI) data, (2) establishment of models for yield predictions using the heading date (HD) NDVI value, and the accumulated NDVI value of the dates from heading to maturity (DHM). The data from 2000 to 2017 were used for building predictive models using the random forests (RF) and support vector machines (SVM), leaving the 2018 data for model assessment, and (3) evaluation of model performance. The results compared with the government's yield statistics indicated good predictions, with the root mean square error (RMSE) and mean absolute error (MAE) values between 7.1% and 11.8%, and Willmott's index of agreement (d) values between 0.81 and 0.84 for the first crop, and 5.6% and 11.3% anddvalues between 0.91 and 0.95 for the second crop, respectively. A slight underestimation of yield predictions was observed for both crops, with the relative error (RE) values of -6.5% to -8.2% and -3.8% to -6% for the first and second crops, respectively. The results of regression analysis also confirmed a close agreement between these two datasets, with the correlation coefficient (r) higher than 0.84 (p-value <0.05), in both cases. Although some factors, including mixed-pixel issues, boundary effects, and cloud cover potentially affected the modelling results, our study demonstrated the effectiveness of ML methods for regional rice yield predictions from MODIS NDVI data in Taiwan.
|Appears in Collections:||SCI期刊|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.