Tree Species Extraction and Land Use/Cover Classification From High-Resolution Digital Orthophoto Maps

Understanding tree species distribution and land use/cover classes plays a key role for developing environmental monitoring and decision support systems. This study investigates a method based on integration of multiple classifiers to improve the classification accuracy for extraction of tree species and land use/cover classes from large scale data. First, a diverse set of classifiers from different families of statistical learning was selected as base classifiers namely: support vector machine, artificial neural network, and random forest. Both spectral and spatial features were, then, extracted and fed into individual classifiers to classify data into four classes (tea gardens, other trees, impervious surfaces, and bare land). Finally, the results obtained from each classifier were combined to obtain final output by maximum voting. The proposed method was evaluated by using an area-based accuracy assessment on a dataset consisting of ten high-resolution digital orthophoto maps. Experimental results showed that integrating the outputs of individual classifiers improved (4%-7%) overall classification accuracy.