Mapping of poplar tree growing fields with machine learning algorithms using multi-temporal sentinel-2A imagery

Abstract

The poplar trees used in peeling, packaging, furniture, fiber chip, cellulose industry and construction sector are one of the most significant wood supply sources of the countries. Monitoring the development stage of cultivated poplar trees, determination of their boundaries and mapping their fields in cheaper and more accurate ways plays an important role for the poplar tree growing sector. The main goal of this study is to map the cultivated hybrid Poplar (P. deltoides) fields in Akyazi district of Sakarya, Turkey using multi-temporal Sentinel-2A satellite imagery and different spectral band combinations. For this purpose, pixel-based supervised classification procedure was selected and three machine learning algorithms, namely, random forest (RF), support vector machines (SVMs) and Adaboost (AdaB) were applied to produce thematic map of the study area. In order to meet desired the goals of the study, three Sentinel-2A imagery from April, July and September, 2019 were used as a multi-temporal dataset consisted of three band combinations. In addition, classification results of multi-temporal datasets compared with single-dated datasets belonged to April for evaluate the effect of using multi-temporal imagery on classification accuracy. Overall accuracy and McNemar’s test were used for accuracy assessment. According to classification results, overall accuracies of multi-temporal datasets became superior from single-dated datasets. Furthermore, McNemar’s test results also affirmed there is significantly difference datasets formed by 20 m resolution bands and pan-sharpened bands of Sentinel-2A imagery between April datasets and multi-temporal datasets. One of used algorithms, AdaB, showed weaker classification performance with respect to RF an SVMs. Furthermore, according to results of F-score, poplar label class reached up to %99 at Dataset-6. © 2021 Elsevier B.V., All rights reserved.