Assessment of image segmentation and deep learning for mapping paddy fields using Worldview-3 in Magelang, Central Java Provinces, Indonesia

Abstract

Paddy fields are complex land-use entities with various surface covers depending on the timing of the planting stages. Therefore, the best practice to map paddy fields using remote sensing has benefited from the availability of multi-temporal data which were used to characterize the phenology related to the paddy fields. However, this practice may require more RS data to be obtained and processed. Other mapping methods by capitalizing the spatial configuration, such as image segmentation in Object-Based Image Analysis (OBIA) and object recognition in Deep Learning using Convolutional- Neural Network (CNN) architecture has been used in the mapping application. This study aims to assess the accuracy from using mean-shift image segmentation and Random Forests and Extreme Gradient Boosting as the classifiers, with the accuracy from simple CNN architecture, by using Worldview-3 (WV3) full-spectrum image (16 bands). The image segmentation and deep learning analysis were conducted by using 16-bands from the WV3 image and classified by using RF and XGB, and CNN. The results showed that RF was able to identify the paddy fields with an accuracy of 88.09 % (User’s accuracy (UA)) and 81.61 % (Producer’s accuracy(PA)), while XGB produced an accuracy of 85.71 % (User’s accuracy (UA)) and 82.44 % (Producer’s accuracy (PA)), respectively. While CNN produced the accuracies of 49.5 % (PA), 96.3 % (UA) and 82.9 % (OA). The lower producer’s accuracy indicated the higher omission error where more paddy fields were classified as non-paddy fields. CNN produced promising accuracy results for identifying paddy field tiles with 82.9 % accuracy without using data augmentation, although it will be needed to increase the accuracy and more complex CNN architecture such as U-net is needed to determine the boundary of the mapped objects.