Optimizing Water Body Detection in Southeast Hubei Using PCA on Landsat ETM+ Imagery

4.1 Different traditional methods for water body extraction

The extraction of water bodies from Landsat ETM+ imagery has been a subject of extensive research. Common methodologies employed in this realm include the single-band threshold method, multi-band enhancement threshold method, and post-classification extraction method. This study has undertaken a comprehensive evaluation of these traditional methods as documented in the literature to assess their effectiveness in extracting water bodies within the specified study area.

The following are some of the indices and methods referenced in previous studies and applied in this research for water body extraction:

(1) TCW

The TCW [12] index is defined as follows:

$\begin{align}

  & TCW=0.1529*TM1+0.1973*TM2+0.3279*TM3 \\ 

 & \text{       }+0.3406*TM4-0.7112*TM5-0.4572*TM7 \\ 

\end{align}$                         (1)

(2) NDWI1

NDWI1 [25] is defined as follows:

$NDWI1=\frac{TM2-TM4}{TM2+TM4}$                         (2)

(3) NDWI2 (MNDWI)

NDWI2 (MNDWI) [9] is defined as follows:

$NDWI1=\frac{TM2-TM5}{TM2+TM5}$                         (3)

(4) NDWI3

NDWI3 [26] is defined as follows:

$NDWI1=\frac{TM4-TM5}{TM4+TM5}$                         (4)

(5) Threshold1

The single-band threshold 1 [27, 28] is defined as follows, and the value of K depends on the histogram distribution and experimental validation.

$if\begin{matrix}

   TM5<K\begin{matrix}

   {}  \\

\end{matrix}  \\

\end{matrix}THEN\begin{matrix}

   {}  \\

\end{matrix}water$                   (5)

(6) Threshold2 (TM2345)

The multi-band threshold 2 [29] was defined as follows. The grayscale map generated by TM2 + TM3-(TM4 + TM5) is marked as TM2345. Through analysis of its histogram and experimental validation, the value of k is determined.

$\begin{align}

  & TM2+TM3-(TM4+TM5) \\ 

 & >K\begin{matrix}

   {}  \\

\end{matrix}THEN\begin{matrix}

   {}  \\

\end{matrix}water \\ 

\end{align}$                         (6)

In addition to the methods mentioned above, PCA and IHS transform are applied to TM453 band combination to observe the characteristics of water body from the perspective of transform space.

(7) MLC

The MLC, often referred to as Bayes classification, is a technique utilized for image classification in remote sensing. This method is grounded in Bayes' criteria, which employs statistical probabilities to categorize image pixels based on their spectral signatures. In this study, samples representing various ground objects were collected and analyzed using the MLC approach applied to the TM453, and classification results of water body were attained.

Table 4. Statistics of minimum, maximum and mean pixel values of different features

4.2 Water body extraction results of traditional methods

Table 4 shows the minimum, maximum and mean pixel values of different ground features in TCW, NDWI1, NDWI2, NDWI3, Threshhold1(TM5), and Threshhold2(TM2345).

Based on the analysis of data in Table 4 and the histogram of each gray-level image, the thresholds are determined for each gray-level image mentioned in the previous section.

Table 5 shows the classification possibility matrix of TCW, NDWI1, NDWI2, NDWI3, Threshhold1(TM5), Threshhold2(TM2345), and TM453_MLC.

Table 5. The possibility matrix of water body extracted using various methods

The comparative analysis of the aforementioned water body extraction methods reveals that while some achieve high accuracy, they also exhibit varying degrees of confusion with other land features, including buildings, vegetation, and shadows.

A close examination of specific water pixels in the Landsat ETM+ imagery highlights that smaller water bodies, such as ponds and rivers, often manifest as mixed pixels. Their spectral characteristics can be easily misinterpreted, leading to confusion with other land cover types, such as buildings, vegetation, and mountain shadows. For instance, Figure 5 displays the classification results using NDWI1, as proposed by McFeeter, which achieved a water body extraction accuracy of 90.13%. However, this method showed considerable mis-classification between ponds and buildings. The TM2345 method, on the other hand, attained the highest water body extraction accuracy at 91.37%. Despite this high level of accuracy, a significant proportion of non-water bodies were incorrectly classified as water. Figure 6 presents the results obtained from TM2345, where small water bodies such as rivers are clearly discernible, but there is notable confusion with buildings and mountain shadows.

image017.png

Figure 5. Water body extraction result using NDWI1

pi_zhu_2023-12-30_154955.png

Figure 6. Water body extraction result of TM2345

4.3 Water body extraction using PCA transformation

PCA is a statistical technique employed in multivariate analysis to explore correlations among multiple variables. It aims to uncover the underlying structure within a set of variables by deriving a small number of principal components. These components are designed to retain as much information from the original variables as possible while being mutually independent. In the context of remote sensing and image analysis, PCA transformation concentrates the information of an image into its initial few principal components. This process allows for the discarding of components containing minimal information, thereby transforming extensive hyperspectral data into a manageable number of informative components. In this experiment, the first three principal components, which encapsulate the most significant information of the image, were retained for analysis.

The results of applying PCA transformation to the TM453 composite image are illustrated in Figure 7. Additionally, Figure 8 presents the spectral characteristics post-PCA transformation.

pi_zhu_2023-12-30_155033.png

Figure 7. Transformed result of TM453 based on PCA (pca453) (3 components)

pi_zhu_2023-12-30_155043.png

Figure 8. Spectra curve of pca453

As indicated in Figure 8, the first component resulting from the PCA transformation of the TM453 composite (hereafter referred to as pca453) exhibits a pronounced differentiation between various features, particularly in distinguishing water from non-water areas. This significant disparity makes the first principal component of pca453 a valuable tool for water body extraction.

The visual analysis of the grayscale image presented in Figure 9 clearly reveals a stark contrast between the water and non-water areas. This distinction is particularly evident in the grayscale rendition, where the water parts are markedly different in tone compared to the surrounding land areas. Further supporting this visual observation, Table 6 provides a statistical breakdown of the pixel values corresponding to different ground objects. The data clearly indicate that the pixel values for water and non-water bodies are significantly different. This disparity highlights the potential for effectively distinguishing between water and non-water areas, provided that an appropriate threshold is determined.

pi_zhu_2023-12-30_155157.png

Figure 9. Grey image generated by the first component of pca453 (pca453-first)

Table 6. Statistics of minimum, maximum and mean pixel values of different features of pca453-first

Building upon the analysis results of the data given in Table 6 and the histogram of the pca453-first grayscale image, an appropriate threshold was determined for the extraction of water bodies. The final extraction results, achieved by applying this determined threshold to the pca453-first component, are displayed in Figure 10, and the classification possibility matrix is given in Table 7.

From Figure 10 and the probability matrix in Table 7, it is evident that the use of pca453-first significantly reduces the mis-classification between water and non-water bodies. Notably, the issue of mountain shadows being erroneously identified as water in the TM2345 method is effectively resolved with pca453-first. This approach demonstrates the highest accuracy in water body extraction among all the methods evaluated in this study, clearly distinguishing itself as the most effective technique in minimizing confusion and accurately identifying water bodies.

pi_zhu_2023-12-30_155241.png

Figure 10. Water body extraction result of pca453_first

Table 7. Possibility matrix of water body extracted of pca453-first