MULTI-TEMPORAL CLASSIFICATION AND CHANGE DETECTION USING UAV IMAGES

S. Makuti, Francesco Nex, Michael Ying Yang

Abstract

In this paper different methodologies for the classification and change detection of UAV image blocks are explored. UAV is not only the cheapest platform for image acquisition but it is also the easiest platform to operate in repeated data collections over a changing area like a building construction site. Two change detection techniques have been evaluated in this study: the pre-classification and the post-classification algorithms. These methods are based on three main steps: feature extraction, classification and change detection. A set of state of the art features have been used in the tests: colour features (HSV), textural features (GLCM) and 3D geometric features. For classification purposes Conditional Random Field (CRF) has been used: the unary potential was determined using the Random Forest algorithm while the pairwise potential was defined by the fully connected CRF. In the performed tests, different feature configurations and settings have been considered to assess the performance of these methods in such challenging task. Experimental results showed that the post-classification approach outperforms the pre-classification change detection method. This was analysed using the overall accuracy, where by post classification have an accuracy of up to 62.6 % and the pre classification change detection have an accuracy of 46.5 %. These results represent a first useful indication for future works and developments.

Introduction

Monitoring rapid urban development and construction sites presents significant logistical hurdles. While satellite imagery is commonly used for change detection, it often lacks the spatial resolution and temporal flexibility required to track daily progress on a building site, often hampered by cloud cover or fixed orbital schedules. Unmanned Aerial Vehicles (UAVs) provide a low-cost, high-resolution alternative. However, the manual generation of change maps from large volumes of UAV imagery is inefficient. This research addresses the need for an automated pipeline to accurately identify land-cover changes across multiple time-stamps (epochs).

Methodology

The study evaluates two primary workflows for change detection:

  1. Pre-classification Change Detection: This method extracts features from two consecutive epochs, calculates the differences between these features, and then uses a classifier to identify the type of change.
  2. Post-classification Change Detection: Each epoch is classified independently into land-cover types (road, building, vegetation, bare soil). These classified maps are then overlaid to identify changes.

The technical implementation utilizes a mix of 2D spectral data (HSV color space), textural data (Gray Level Co-occurrence Matrix - GLCM), and 3D geometric data (linearity, planarity, and normalized Digital Surface Models) derived from UAV photogrammetry. The classification engine employs a Conditional Random Field (CRF) model, where a Random Forest (RF) classifier provides the unary potential (initial pixel labels) and a Fully Connected CRF defines the pairwise potential to spatially smooth the results and reduce noise.

Experimental Setup

The study utilized a dataset captured over a construction site in Lausanne, Switzerland, with a 5 cm Ground Sampling Distance (GSD) across 8 different epochs. Images were processed into orthophotos and Digital Surface Models (DSMs) using Pix4D.

Orthophoto and Digital Surface Model (DSM) used as primary data sources for the multi-temporal study.

Results and Discussion

The experimental results indicate that the post-classification approach is significantly more robust, achieving an overall accuracy (OA) of up to 62.6%, compared to 46.5% for the pre-classification method. The post-classification method's superiority stems from its ability to detect all 16 possible "from-to" change combinations, whereas pre-classification struggled with subtle variability between classes with similar spectral signatures, such as bare soil and concrete.

A sample classification map generated by the workflow, categorizing the site into vegetation (green), road (brown), building (yellow), and bare soil (orange).

Key findings include:

  • Impact of 3D Data: Integrating 3D geometric features from the DSM increased classification accuracy by approximately 11%. Height remains the most reliable indicator for structures in construction environments.
  • CRF Refinement: The Fully Connected CRF acted as an effective spatial filter. For a single epoch, the OA improved from 50.4% (RF alone) to 57.9% after applying the CRF, eventually reaching 66.3% with the addition of 3D features.
  • Class Confusion: Vegetation was the most accurately detected class (IoU up to 83.9%). Conversely, bare soil and buildings were frequently confused due to their similar radiometric responses.

Conclusion

This research demonstrates that a post-classification workflow incorporating 3D geometric data and CRF-based spatial smoothing provides a viable path for automated construction site monitoring. While current accuracies are modest, they provide a foundation for future work, which may involve deep learning techniques like Convolutional Neural Networks (CNNs) to better handle the radiometric similarities between construction materials.

Summary updated

Abstract

In this paper different methodologies for the classification and change detection of UAV image blocks are explored. UAV is not only the cheapest platform for image acquisition but it is also the easiest platform to operate in repeated data collections over a changing area like a building construction site. Two change detection techniques have been evaluated in this study: the pre-classification and the post-classification algorithms. These methods are based on three main steps: feature extraction, classification and change detection. A set of state of the art features have been used in the tests: colour features (HSV), textural features (GLCM) and 3D geometric features. For classification purposes Conditional Random Field (CRF) has been used: the unary potential was determined using the Random Forest algorithm while the pairwise potential was defined by the fully connected CRF. In the performed tests, different feature configurations and settings have been considered to assess the performance of these methods in such challenging task. Experimental results showed that the post-classification approach outperforms the pre-classification change detection method. This was analysed using the overall accuracy, where by post classification have an accuracy of up to 62.6 % and the pre classification change detection have an accuracy of 46.5 %. These results represent a first useful indication for future works and developments.
BibTeX
@inproceedings{makuti2018multi,
  author = {S. Makuti and Francesco Nex and Michael Ying Yang},
  title = {{MULTI-TEMPORAL CLASSIFICATION AND CHANGE DETECTION USING UAV IMAGES}},
  booktitle = {˜The},
  year = {2018},
  doi = {10.5194/isprs-archives-xlii-2-651-2018},
}
Figure