Use of Machine Learning Algorithms in the Classification of Forest Species

Abstract

Optimization in the process of managing forest resources seeks alternatives that make data collection possible. One of them alternatives is spectroradiometry, which consists of measuring the spectral response, having as product the response of the target in relation to the incident radiation along the electromagnetic spectrum, and that, using machine learning, with pre-selected models, makes it possible to identify. Given the above, the study aimed to use machine learning algorithms to classify species by vegetation indices from reflectance data. The study was developed at the Federal University from Santa Maria, working with the species Ficus benjamina, Inga marginata, Handroanthus chrysotrichus, Psidium cattleianum, Salix humboldtiana, Corymbia citriodora and Myrcianthes pungens, and spectral readings of the leaves were taken using the FieldSpec®3 spectroradiometer connected to RTS-3ZC3 integrating  sphere. The reflectance values with wavelength ranged in amplitude from 350 ƞm to 2,500 ƞm and spectral resolution of 1 ƞm. Vegetation indices were calculated using the software R Studio, being: NDVI, SAVI, RVI, GNDVI, NDWI, NDWI2, GEMI, DVI, TVI, RVI, MSAVI, WDVI. The algorithms used to develop machine learning were: Random Forest (RF), k-Nearest Neighbors (K-NN), Naive Bayes (NB) and Support Vector Machine (SVM). RF proves to be the most appropriate for data validation, with 85% global accuracy, followed by SVM, with 71%, K-NN with 64% and NB with 35%. The indices with the best performance to point the species were NDWI and SAVI.