Comparison between Support Vector Machine and Random Forest Models in Estimating Tea Yields at Mwera Tea Farm using Sentinel and Landsat 8 Imagery.

Abstract

Tea is the second-largest export cash crop of Uganda making tea a very important crop since it generates a lot of income for the government and the local people through its revenue generated from the export and selling of tea products by the local people. The production of tea is however greatly affected by high temperatures and fluctuation in rainfall amounts leading to low yields. Traditionally in-situ methods of crop yield estimation have been used to estimate crop yields however these methods are time-consuming, laborious and costly. Crop growth models like the EPIC, Holland’s WOFOST, the United States DSSAT and Australia’s ASPIM model have been widely used to forecast crop yields however the accuracy of the estimated crop yields are low compared however the accuracy of the produced from machine learning models. This study aimed at comparing the estimates of tea yields by a support vector machine and random forest models to determine whether the machine learning models can produce predictions with accuracy comparable to the observed tea yields at Mwera Tea Estate based on Sentinel 2A derived NDVI, Landsat 8 LST, and rainfall data. Tea yields estimated by the support vector machine model were compared to those estimated by the random forest model for four different years. Four years were considered from 2017 to 2020 because of the year the Sentinel 2A imagery was launched and by the time of project execution, some images were not available due to sensor problems. The statistical analysis performed showed that both support vector machine and random forest models can be used to estimate tea yields with the support vector machine model producing a percentage error of tea yield of 2.67% compared to 3.23% produced by the random forest model. This made the support vector machine model be the best model in comparison between the two when estimating tea yields with data of a small sample size.