Remote Sensing Approach to Oil Palm Plantations Detection Using Xception

Palm oil is an essential export commodity for Malaysia's economy. Thus, it is important for the government to promote sustainable ways of managing oil palm plantations. One of the crucial aspects of sustainable plantation is the ability to detect and map the plantation areas accurately. Therefore, this paper proposed a deep convolutional neural network (CNN) framework for oil palm plantations detection using remote sensing images. The main objective of this work is to build an Xception-based detection network, which will be optimized for robust detection, regardless of the tree age. The dataset is downloaded from the Kaggle platform that consists of both the remote sensing images and their ground-truth labels. Each image is classified as positive if there are some plantation areas inside the image and vice versa. A transfer learning of XceptionNet architecture will be retrained using saved parameters from Keras library. Adam optimizer is used in the training process and yields a detection accuracy of 98.96%. This work can be extended to include a segmentation network so that the plantation areas can be inferred directly from the remote sensing images.