Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhanced LULC Segmentation via Lightweight Model Refinements on ALOS-2 SAR Data

Ali Caglayan, Nevrez Imamoglu, Toru Kouyama

TL;DR

The study tackles dense LULC segmentation from single-polarization ALOS-2 SAR over Japan, addressing boundary fidelity, thin-structure recovery, and rare-class robustness. It leverages a SAR-tailored foundation model via SAR-W-MixMAE pretraining and adds three lightweight refinements—high-resolution feature injection, a progressive refine-up head, and an alpha-scale tempered focal-dice loss—to improve dense predictions without increasing pipeline complexity. Empirically, the approach yields a mean IoU of about $0.50$ on the Japan LULC test set and enhances water-detection IoU to around $0.936$, with stronger gains for under-represented classes. Overall, the combination of self-supervised SAR pretraining and minimal architectural refinements demonstrates substantial improvements in boundary accuracy and rare-class performance for nationwide SAR-based LULC mapping and water delineation.

Abstract

This work focuses on national-scale land-use/land-cover (LULC) semantic segmentation using ALOS-2 single-polarization (HH) SAR data over Japan, together with a companion binary water detection task. Building on SAR-W-MixMAE self-supervised pretraining [1], we address common SAR dense-prediction failure modes, boundary over-smoothing, missed thin/slender structures, and rare-class degradation under long-tailed labels, without increasing pipeline complexity. We introduce three lightweight refinements: (i) injecting high-resolution features into multi-scale decoding, (ii) a progressive refine-up head that alternates convolutional refinement and stepwise upsampling, and (iii) an $α$-scale factor that tempers class reweighting within a focal+dice objective. The resulting model yields consistent improvements on the Japan-wide ALOS-2 LULC benchmark, particularly for under-represented classes, and improves water detection across standard evaluation metrics.

Enhanced LULC Segmentation via Lightweight Model Refinements on ALOS-2 SAR Data

TL;DR

The study tackles dense LULC segmentation from single-polarization ALOS-2 SAR over Japan, addressing boundary fidelity, thin-structure recovery, and rare-class robustness. It leverages a SAR-tailored foundation model via SAR-W-MixMAE pretraining and adds three lightweight refinements—high-resolution feature injection, a progressive refine-up head, and an alpha-scale tempered focal-dice loss—to improve dense predictions without increasing pipeline complexity. Empirically, the approach yields a mean IoU of about on the Japan LULC test set and enhances water-detection IoU to around , with stronger gains for under-represented classes. Overall, the combination of self-supervised SAR pretraining and minimal architectural refinements demonstrates substantial improvements in boundary accuracy and rare-class performance for nationwide SAR-based LULC mapping and water delineation.

Abstract

This work focuses on national-scale land-use/land-cover (LULC) semantic segmentation using ALOS-2 single-polarization (HH) SAR data over Japan, together with a companion binary water detection task. Building on SAR-W-MixMAE self-supervised pretraining [1], we address common SAR dense-prediction failure modes, boundary over-smoothing, missed thin/slender structures, and rare-class degradation under long-tailed labels, without increasing pipeline complexity. We introduce three lightweight refinements: (i) injecting high-resolution features into multi-scale decoding, (ii) a progressive refine-up head that alternates convolutional refinement and stepwise upsampling, and (iii) an -scale factor that tempers class reweighting within a focal+dice objective. The resulting model yields consistent improvements on the Japan-wide ALOS-2 LULC benchmark, particularly for under-represented classes, and improves water detection across standard evaluation metrics.
Paper Structure (9 sections, 4 equations, 4 figures, 3 tables)

This paper contains 9 sections, 4 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. LULC segmentation loss
  4. Water detection loss
  5. Experiments
  6. Dataset
  7. Evaluation and Training Details
  8. Results and Analyses
  9. Conclusion

Figures (4)

  • Figure 1: Overview of the proposed pipeline. Left: self-supervised pretraining on unlabeled ALOS-2 HH using SAR-W-MixMAE sar_w_mixmae, followed by supervised finetuning for downstream dense prediction. Right: finetuning architecture with a Swin encoder and a UPerNet-style xiao2018unified decoder (PPM + FPN), incorporating (i) highest-resolution post-embedding feature injection and (ii) progressive upsampling to improve overall performance while mitigating boundary smoothing.
  • Figure 2: Validation mIoU during training; scratch run shown up to 200 epochs.
  • Figure 3: Qualitative LULC segmentation comparison among RSSJ'25 imamoglu2025rssj, our no-pretraining model, and our pretrained model with proposed refinements.
  • Figure 4: Large-area water detection example showing SAR input, model prediction, and reference water mask.