Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Detecting Cadastral Boundary from Satellite Images Using U-Net model

Neda Rahimpour Anaraki, Maryam Tahmasbi, Saeed Reza Kheradpisheh

TL;DR

This work tackles automated cadastral boundary detection from satellite and UAV imagery using a transfer-learned U-Net with a ResNet34 backbone for three-class semantic segmentation (boundary, field, background). It carefully curates a small, patch-based dataset and employs Laplacian pre-processing, a 2-pixel boundary buffer, and a Jaccard–Focal loss to train the model, achieving high precision (approximately 0.88), solid recall (0.75), and a strong F-score (0.81) on two test farmland images from Iran. The boundaries are converted to vector shapefiles via skeleton-based post-processing, enabling GIS-ready outputs, and the approach demonstrates improved boundary quality and rural/urban discrimination relative to a Mask R-CNN baseline. The study provides a practical, transfer-learning pipeline for cadastral mapping in contexts with limited labeled data, while acknowledging vectorization limitations and suggesting avenues for future enhancement.

Abstract

Finding the cadastral boundaries of farmlands is a crucial concern for land administration. Therefore, using deep learning methods to expedite and simplify the extraction of cadastral boundaries from satellite and unmanned aerial vehicle (UAV) images is critical. In this paper, we employ transfer learning to train a U-Net model with a ResNet34 backbone to detect cadastral boundaries through three-class semantic segmentation: "boundary", "field", and "background". We evaluate the performance on two satellite images from farmlands in Iran using "precision", "recall", and "F-score", achieving high values of 88%, 75%, and 81%, respectively, which indicate promising results.

Detecting Cadastral Boundary from Satellite Images Using U-Net model

TL;DR

This work tackles automated cadastral boundary detection from satellite and UAV imagery using a transfer-learned U-Net with a ResNet34 backbone for three-class semantic segmentation (boundary, field, background). It carefully curates a small, patch-based dataset and employs Laplacian pre-processing, a 2-pixel boundary buffer, and a Jaccard–Focal loss to train the model, achieving high precision (approximately 0.88), solid recall (0.75), and a strong F-score (0.81) on two test farmland images from Iran. The boundaries are converted to vector shapefiles via skeleton-based post-processing, enabling GIS-ready outputs, and the approach demonstrates improved boundary quality and rural/urban discrimination relative to a Mask R-CNN baseline. The study provides a practical, transfer-learning pipeline for cadastral mapping in contexts with limited labeled data, while acknowledging vectorization limitations and suggesting avenues for future enhancement.

Abstract

Finding the cadastral boundaries of farmlands is a crucial concern for land administration. Therefore, using deep learning methods to expedite and simplify the extraction of cadastral boundaries from satellite and unmanned aerial vehicle (UAV) images is critical. In this paper, we employ transfer learning to train a U-Net model with a ResNet34 backbone to detect cadastral boundaries through three-class semantic segmentation: "boundary", "field", and "background". We evaluate the performance on two satellite images from farmlands in Iran using "precision", "recall", and "F-score", achieving high values of 88%, 75%, and 81%, respectively, which indicate promising results.

Paper Structure

This paper contains 7 sections, 3 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Data
  3. Pre-Process
  4. Method
  5. Post-Process
  6. Evaluation
  7. Conclusion

Figures (7)

  • Figure 1: Image13 and its corresponding initial mask.
  • Figure 2: From left to right, Original patch, it’s corresponding initial mask, secondary mask and final mask with buffer 1, 2 and 5 pixels.
  • Figure 3: From left to right, Original patch, high-pass filter, Laplacian filter, Sharped filter and Sharped then Laplacian filter applied on it.
  • Figure 4: U-Net model architecture.
  • Figure 5: NegarKhatun image and it’s corresponding initial output.
  • ...and 2 more figures