Instance segmentation of center pivot irrigation systems using multi-temporal SENTINEL-1 SAR images

• Published in: Remote sensing applications Vol. 23; p. 100537
• Main Authors: de Albuquerque, Anesmar Olino, de Carvalho, Osmar Luiz Ferreira, e Silva, Cristiano Rosa, de Bem, Pablo Pozzobon, Trancoso Gomes, Roberto Arnaldo, Borges, Díbio Leandro, Guimarães, Renato Fontes, Pimentel, Concepta Margaret McManus, de Carvalho Júnior, Osmar Abílio
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2021
• Subjects: Center pivot; Deep learning; Mask R–CNN; SAR imagery; Time series; Deep learning; Center pivot; SAR imagery; Mask R–CNN; Time series
• ISSN: 2352-9385; 2352-9385
• DOI: 10.1016/j.rsase.2021.100537

The mapping of Center Pivot Irrigation Systems (CPIS) is essential for agricultural and water resource management. In this context, methods based on Deep Learning (DL) have reached state-of-the-art in the classification of remote sensing images. However, the mapping of CPIS with DL is still restricted to optical images with limitations in tropical environments due to the extensive cloud cover for long periods. The present research proposes the detection of CPIS using instance segmentation from multi-temporal SAR images that are cloud-free. The research developed a CPIS database for the Cerrado biome based on visual interpretation, totaling 3675 instances in the Common Objects in Context (COCO) annotation format. The training used the Mask-RCNN with the ResNeXt-101-32x8d backbone considering different data arrangements: (a) variation in the number of Sentinel-1 temporal images with an interval of 12 days (from 1 to 11 images), and (b) comparison of VV, VH, and VV + VH polarizations. For mapping large areas, we applied mosaicking with a sliding window technique. The results show an accuracy improvement with the increase in the number of temporal images, reaching a difference greater than 15% AP when comparing a single temporal image with the optimal number of temporal images in the VV (eight), VH (ten) and VV + VH (nine) polarizations. The combined use of the two polarizations (VV + VH) had slightly better results (75% AP, 91% AP50, and 86% AP75) than the others. However, VV polarization may have an advantage, obtaining close results from less image and computational cost. The instance segmentation with the sliding window provides the automatic identification of different objects belonging to the same class in large areas, allowing the total CPIS count and the size calculation.