The Age Dynamics of Para Rubber Plantations using Landsat TimeSeries (1991 - 2018) using Machine Learning Algorithms

Abstract

The investigated the potential of using Landsat time series data and secondary land use and land cover (LULC) data to identify the ages of para rubber plantation in the lowland of Thalang district, Phuket province, southern Thailand. The LULC data, including high spatial resolution historical image from Google Earth ProTM, were used to identify rubber plantation and the event (year) of rubber planting (To). The inter-annual vegetation profiles of 2,168 rubber plantations were extracted from the distribution of sample NDVI values, which depend on the particular plot's size, for each summer period of 129 Landsat NDVI images (October 1991 to April 2018). The predictor variables were generated from difference and ratio of NDVI distribution values (minimum, Q1, median, Q3, and maximum) at different seasons (two years before and six years after To) for Recursive Partitioning (RP) supervised classification algorithm. Modeling data (outcome and predictors) from 336 plantations were divided into the training and testing datasets. The predicted RP model was learning on training data (30- time repeated) and we used testing data for cross-validation assessment to optimize an appropriated hyperparameter of the RP model. Then, the RP model with a complexity parameter as 0.01 was applied on both modeling data and predicting data (1,832 plots that unknowns To). The predicted To for each plantation was selected based on the maximum nominated in 100-time repeated prediction. Finally, the result validation of To prediction was carried out using 131 records of rubber farmers' registration from Rubber Authority of Thailand (RAOT). There are 15 plots (11.5%) that have a correct prediction, and 54 plots (41.2%) have one-year error prediction because many farmers start planting one or two years after approval from RAOT. The average error prediction is 3.62 years. We found that there is a possibility of using a 30-meter spatial resolution Landsat NDVI time series to identify rubber plantation ages with high accuracy, especially in the larger plots. The high precision of para rubber stands ages database will enable accurate yield prediction that, subsequently, resulting in better decision-making, planning, and development in the agricultural sector.