Classification of Potential Water Bodies Using Landsat 8 OLI and a Combination of Two Boosted Random Forest Classifiers

ByoungChul Ko, Hyeong Hun Kim, Jae-Yeal Nam

TL;DR

This paper introduces a robust water body classification method using Landsat 8 OLI data, combining Top-of-Atmosphere (TOA) reflectance and water indices (WIs) with two Boosted Random Forest (BRF) classifiers. The approach achieves high accuracy and outperforms existing methods, leveraging the improved signal-to-noise ratio of the Landsat 8 OLI sensor.

What Problem Does This Paper Address?

Accurate and efficient mapping of water bodies (rivers, lakes) is crucial for water quality assessment and management. Traditional methods often rely on heuristic thresholds, which can be sensitive to terrain complexity and atmospheric conditions, leading to misclassifications. While classifier-based methods offer improvements, supervised learning requires extensive labeled data, and unsupervised methods often need post-processing. This study aims to develop a more robust and accurate water body classification method that overcomes these limitations, specifically addressing the challenges posed by varying water types, topologies, and background environments.

Methods

The proposed method utilizes Landsat 8 OLI sensor data, focusing on Top-of-Atmosphere (TOA) reflectance and Water Indices (WIs) as features. The core of the approach involves two distinct Boosted Random Forest (BRF) classifiers:

  • BRF1: Trained using TOA reflectance values from six spectral bands (bands 2-7).
  • BRF2: Trained using three derived Water Indices (NDWI and two versions of MNDWI).

These two BRFs are trained separately to leverage the different characteristics of TOA reflectance and WIs, avoiding potential performance degradation from combining them into a single feature vector. The final classification probability for each pixel is obtained by linearly combining the probabilities from the two BRFs, with a learned weight (w₁) that was experimentally determined to be 0.5. This weighted combination aims to enhance robustness across diverse water body types and environments. The method also includes a pre-processing step to remove shadow pixels using a threshold on the green band reflectance. Post-classification, morphological closing is applied to merge water regions, and small regions are filtered as noise.

Datasets

The study utilizes Landsat 8 OLI sensor data. For training and evaluation, data from three specific sites in Korea were used, chosen for their diverse water types and topologies:

  • Area 1: Yedang artificial reservoir (agricultural areas).
  • Area 2: Soyangho lake (mountainous areas with shadows).
  • Area 3: Middle Nakdonggang River (agricultural and mountainous areas).

The paper also compares its performance against methods using Landsat 5 TM and Landsat 7 ETM+ data, captured on similar dates.

Key Insights and Ideas

  • Landsat 8 OLI Advantage: The improved Signal-to-Noise Ratio (SNR) of the Landsat 8 OLI sensor, due to its 12-bit quantization, significantly contributes to higher classification accuracy compared to older Landsat sensors (TM and ETM+).
  • Feature Synergy: Using both TOA reflectance and Water Indices as separate features for distinct BRF classifiers, rather than a combined feature, leads to better classification performance. This highlights the complementary nature of these feature sets.
  • Boosted Random Forest Robustness: The combination of two BRFs effectively handles variations in water body topology, river properties, and background environments, demonstrating a robust classification capability.
  • Weighted Combination is Key: The linear combination of probabilities from the two BRFs, with an optimized weight (w₁=0.5), is crucial for achieving the best overall performance, balancing the contributions of reflectance and index-based features.

Key Results

  • Sensor Comparison: The Landsat 8 OLI sensor achieved an average Overall Accuracy (OA) of 99.90% and an average Kappa of 0.9942, outperforming TM (99.05% OA, 0.9469 Kappa) and ETM+ (99.73% OA, 0.9738 Kappa).
  • Algorithm Comparison: The proposed method achieved an average OA of 99.90% and average Kappa of 0.9942. This was superior to:
  • Otsu threshold method (average OA: 99.31%, Kappa: 0.965672)
    • SVM classifier (average OA: 99.64%, Kappa: 0.98198)
    • RF classifier (average OA: 99.69%, Kappa: 0.983167)
  • Processing Speed: The proposed method achieved a processing speed of 12.91 seconds per image, which is significantly faster than SVM (82.4 s/image) and comparable to RF (16.4 s/image), though slower than Otsu threshold (5.2 s/image).

Limitations and Future Work

  • Narrow Rivers and Cluttered Backgrounds: The paper acknowledges that the proposed method can still produce miss or false classifications when river widths are very narrow or when the background contains structures similar to rivers or lakes.
  • Processing Speed: While faster than SVM, the processing speed is still a consideration for real-time applications.
  • Future Work:
    • Focus on reducing missing and false classifications, particularly for narrow rivers and cluttered backgrounds.
  • Improve processing speed to approach that of simpler methods like Otsu threshold, without sacrificing accuracy.
  • Expand research to measure water quality using the classified water bodies and TOA reflectance values.

Summary updated

BibTeX
@inproceedings{ko2015classification,
  author = {ByoungChul Ko and Hyeong Hun Kim and Jae-Yeal Nam},
  title = {{Classification of Potential Water Bodies Using Landsat 8 OLI and a Combination of Two Boosted Random Forest Classifiers}},
  booktitle = {Sensors},
  year = {2015},
  url = {https://openreview.net/forum?id=Dhv9MEXnFR_},
}

Loading related content...

Figure