Impact of Training Set Configurations for Differentiating Plantation Forest Genera with Sentinel-2 Imagery and Machine Learning

Abstract

Forest plantations in South Africa impose genus-specific demands on limited soil moisture. Hence, plantation composition and distribution mapping is critical for water conservation planning. Genus maps are used to quantify the impact of post-harvest genus-exchange activities in the forestry sector. Collecting genus data using in situ methods is costly and time-consuming, especially when performed at regional or national scales. Although remotely sensed data and machine learning show potential for mapping genera at regional scales, the efficacy of such methods is highly dependent on the size and quality of the training data used to build the models. However, it is not known what sampling scheme (e.g., sample size, proportion per genus, and spatial distribution) is most effective to map forest genera over large and complex areas. Using Sentinel-2 imagery as inputs, this study evaluated the effects of different sampling strategies (e.g., even, uneven, and area-proportionate) for training the random forests machine learning classifier to differentiate between Acacia, Eucalyptus, and Pinus trees in South Africa. Sample size (s) was related to the number of input features (n) to better understand the potential impact of sample sparseness. The results show that an even sample with maximum size (100%, s~91n) produced the highest overall accuracy (76.3%). Although larger training set sizes (s > n) resulted in higher OAs, a saturation point was reached at s~64n.