Research Approach

Experiments

The main objective of this research is to overcome the following challenges in flood inundation extent mapping:

• Incorporating the high local spectral variation of flood water. Generally, mixing of flood water with different land features causes variation in colour (spectral properties) of flood water in different locations.
• Distinguishing flood water and other land features belonging to different classes but characterised with high similar spectral (color) properties. Eg. flood water and permanent water bodies or muddy flood water and moist soil looks similar in the remote sensing images of moderate of high resolution.
• Difficulty in manual ground-truth collection process during flood events.

Practically, it is therefore, hard to generalise a classification model with available training data that could be able to delineate flooded areas from any remote sensing data and the robustness of the classification models depend on highly accurate human engineered approaches or semi-automated approaches. With the aim of preparing a robust classification model, deep learning based approach has been selected. Initially, the experiment tries to incorporate information from multiple sources that are freely available and easy to access as real time data.

At the beginning, a number of experiments have been undertaken by applying state-of-the art machine learning methods to evaluate their performance in extracting flood inundation from remote sensing images.

• Earlier experiments applies Extended Support Vector Machine classifier along with linear spectral unmixing method on multispectral satellite imagery to obtain the classification output. The efficiency of classifier was enhanced by introducing the Bayesian probability estimation and performance was assessed by using precision and recall metrics on the pre and post Bayesian flood probability estimation.

For more details please refer http://ieeexplore.ieee.org/document/7797054/

• Experiments have also incorporated non-linear Support Vector Machines and Markov Random Field Model classification model evaluate the state-of -the art machine learning methods on detection of flooded areas from remote sensing images.
• Two most widely used image pre-processing methods were implemented to examine the improvement in the accuracy level of classification results. The methods include:
  • Conservative Smoothening and Contrast-limited adaptive histogram equalization (CLAHE) method.
• A comparative analysis was performed with the classification results from raw unprocessed satellite image.
• Los estimation by Logarithmic Loss function has applied to estimate the error level of flood water and permanent water classification results from Linear and Non-Linear SVM.

• Iterated Conditional Modes (ICM) have been applied for estimating the maximum a posteriori of MRF image segmentation.

• With the aim of replacing the manual selection process to train the classification model, deep learning method has been introduced in this experiment. This is because human engineered methods limits the accuracy level of the classification results and every classification model has been built as image specific. That is the classification model works for multispectral satellite images may not be appropriate for aerial image classification. The Deep learning methods has been therefore considered for further experiment to generate a more generalised classification model that is trained to extract flooded inundation extent areas from real time multispectral remote sensing images, either aerial or space-borne.