Agricultural crop type mapping using object-based image analysis with advanced ensemble learning algorithms

Abstract

Parallel to the rapid technological advances, up-to-date remote sensing platforms and sensors have made it possible to observe the Earth's surface features at a higher spatial and spectral resolution. The WorldView-2 (WV-2) imagery has been effectively used for the detailed mapping of agricultural crop-type types in many studies. The selected area for this study covers approximately 17 km2 of various agricultural land and forest areas. Hazelnut and corn products, which are spectrally similar crop types, are the most dominant and economically valuable agricultural products for this particular region. Therefore, accurate determination of these agricultural products and mapping of their spatial locations play important role for yield estimation. The primary objective of this paper is to map cultivated areas by classifying a WV-2 imagery using conventional classifiers and advanced ensemble learning algorithms. In addition, several spectral indices were used as an ancillary dataset to identify and differentiate cultivated areas and forest species from each other. Within this context, object-based image analysis (OBIA) with multi-resolution segmentation was performed to produce image objects. Then, a total of 22 image object subsets were determined for the segmented image objects. Four classification algorithms, namely Random Forest (RF), Canonical Correlation Forest (CCF), Decision Tree (DT), and k-nearest neighbor (k-NN) classifiers were applied to produce the thematic map of the study area. Result of the study showed that the CCF classifier reached the highest overall accuracy of 94% with dataset having 22 object subsets. The improvement in classification performance reached to 7% in terms of overall classification accuracy. Moreover, the results noticeably indicated that ensemble methods (i.e., RF and CCF) outperformed the DT and k-NN classifiers in terms of applied accuracy measures. The results were confirmed by the McNemar's statistical test. Moreover, feature importance results of the RF algorithm showed that the most important vegetation indices were Chlorophyll RedEdge, green leaf, and NDVI-2 indices, respectively. © 2021 Elsevier B.V., All rights reserved.