Scalable Deep Learning Framework for Non-Urban Landcover Mapping in Tropical Regions Using IFSAR Data Yanuar A.N. (a*), Muhammad Nizar Y.P. (b)
a) Product and Technology Lead, Solution Engineering, Esri Indonesia,
Jalan Gatot Subroto 6, Kota Jakarta Selatan 12710, Indonesia
*ynugroho[at]esriindonesia.co.id
b) Data Scientist, Professional Services, Esri Indonesia
Abstract
Indonesia has 1,705,029 km square of non-urban territory, which faces significant landcover mapping challenges due to persistent cloud coverage limiting optical remote sensing applications. We propose a scalable GeoAI framework leveraging cloud-penetrating Interferometric Synthetic Aperture Radar (IFSAR) for comprehensive tropical landcover assessment. Our methodology integrates high-resolution X-band airborne SAR Ortho-rectified Radar Imagery (X-ORI) and complementary RGB aerial photographs. To overcome single-polarization grayscale SAR limitations in automated classification, we develop a flexible two-stage deep learning pipeline with alternative enhancement approaches. The first stage employs either: (1) CycleGAN for unsupervised image-to-image translation, converting grayscale SAR into synthetic RGB imagery, or (2) advanced pansharpening techniques, including Brovey Transform, Gram-Schmidt, and wavelet-based fusion to enhance spatial and spectral resolution. Both approaches improve semantic interpretability for subsequent processing. The second stage utilizes U-Net architecture for semantic segmentation, generating detailed landcover maps across diverse tropical landscapes. A comparative analysis demonstrates the effectiveness of both enhancement methods in various scenarios. Results yield three key outcomes: (1) high-resolution, cloud-free enhanced imagery for tropical environments, (2) development of flexible, reusable GeoAI models with multiple enhancement options, and (3) an efficient, scalable solution for national-level landcover monitoring. This framework provides a transformative approach for land management, deforestation monitoring, and sustainable development planning in tropical, cloud-dense regions.
Keywords: IFSAR, Landcover Mapping, CycleGAN, Pansharpening, Deep Learning
