Deep and machine learnings of remotely sensed imagery and its multi-band visual features for detecting oil palm plantation

Characterization of oil palm plantation is a crucial step toward many geographical based management strategies, ranging from determining regional planting and appropriate species to irrigation and logistics planning. Accurate and most updated plantation identification enables well informed and effective measures for such schemes. This paper proposes a computerized method for detecting oil-palm plantation from remotely sensed imagery. Unlike other existing approaches, where imaging features were retrieved from spectral data and then trained with a machine learning box for region of interest extraction, this paper employed 2-stage detection. Firstly, a deep learning network was employed to determine a presence of oil-palm plantation in a generic Google satellite image. With irrelevant samples being disregarded and thus the problem space being so contained, the images with detected oil-palm had their plantation delineated at higher accuracy by using a support vector machine, based on Gabor texture descriptor. The proposed coupled detection-delineation was benchmarked against different feature descriptors and state-of-the-art supervised and unsupervised machine learning techniques. The validation was made by comparing the extraction results with those ground surveyed by an authority. It was shown in the experiments that it could detect and delineate the plantations with an accuracy of 92.29% and precision, recall and Kappa of 91.16%, 84.97%, and 0.81, respectively.