Performance of Random Forest Classifier for Flood Mapping Using Sentinel-1 SAR Images

Abstract

The city of Khartoum, the capital of Sudan, was heavily damaged by the flood of the Nile in 2020. Classification using satellite images can define the damaged area and help emergency response. As Synthetic Aperture Radar (SAR) uses microwave that can penetrate cloud, it is suitable to use in the flood study. In this study, Random Forest classifier, one of the supervised classification algorithms, was applied to the flood event in Khartoum with various sizes of the training dataset and number of images using Sentinel-1 SAR. To create a training dataset, we used unsupervised classification and visual inspection. Firstly, Random Forest was performed by reducing the size of each class of the training dataset, but no notable difference was found. Next, we performed Random Forest with various number of images. Accuracy became better as the number of images in creased, but converged to a maximum value when the dataset covers the duration from flood to the completion of drainage.

1. Introduction

Flood is one of the most frequently occurring natural disasters and life-threatening disasters all over the world every year (Kuenzer et al., 2013). Centre for Research on the Epidemiology of Disasters (CRED) and United Nations Office for Disaster Risk Reduction (UNDRR) announced that flood has been accounted for about 44% of all disaster events in the last 20 years (CRED and UNDRR, 2020). From 2000 to 2019, flood affected 1.6 billion people worldwide, which is the highest figure among all other disaster types. Unpredictable flood harms crops, destroys infrastructure, and impairs accessing to natural resources (Kuenzer et al., 2013). To recover damaged areas quickly and prepare responses for flood hazard areas, it is important to accurately determine the size and location of flooded areas.

However, it is limited and dangerous to get data from in situ measurement for flooded areas (Sanyal and Lu, 2004). Remote sensing data can provide information over a large area and a long period (Uddin et al., 2019). Furthermore, it can be used for areas that are inaccessible and difficult to acquire field data (Lim and Lee, 2018). Researches of flooded areas using remote sensing have been conducted in various ways. The usage or purpose of satellite data depends on the wavelength of the electromagnetic waves. Optical satellite imagery which uses about 480–780 nm wavelength band is typically used for analysis of post floods such as generation of flood extent map because of its poor penetration for cloud during the flood (Lin et al., 2016). For multispectral satellite imagery, there are some studies of flood mapping using spectral indices, for instance, normalized difference water index (NDWI) (Jain et al., 2005; Soltanian et al., 2019).

Radar imagery acquires data by collecting returned microwave signal that is backscattered from the Earth’s surface. As Synthetic Aperture Radar (SAR) is an active microwave imaging system, it is independent of sun azimuth and elevation in data acquisition. Microwave has longer wavelength than optical bands, typically, of 3 cm (X-band), 6 cm (C-band), and 24 cm (L-band). Therefore, it can penetrate most clouds, aerosols, and water vapour (Bamler, 2000). Hence, radar imagery shows more powerful performance on flood research even though in bad weather conditions during flood period than an optical satellite imagery (Jüssi, 2015).

Pixel-based image classification is one of the techniques that can be applied to flood mapping. Image classification is grouped into two types: unsupervised classification and supervised classification. Essentially, they have the same methodology: it is clustering pixels in an image to a set of classes based on land surface properties (Hasmadi, 2009). The difference between them is whether prior information of the land cover is given or not. Unsupervised classification performs clustering without prior information, and the resulting classes are unlabelled. Supervised classification, on the other hand, requires prior information to classify pixels. It trains features of pixels in the given training datasets and applies to the whole image. The resulting classification image is acquired with labelled classes (Hasmadi, 2009).

Random Forest, one of the supervised classifier based on machine learning, is an ensemble of decision trees suggested by Breiman (2001). A decision tree model is an analysis method that classifies or predicts the region of interest into some subgroups by depicting decision rules as tree structure (Song and Chae, 2008). By the way, there is a common issue called overfitting for decision tree model. Overfitting makes the model optimized only for the training dataset by memorizing features including even the noise. In other words, the model will be poor to predict general properties for new dataset (Dietterich, 1995). Random Forest, however, can prevent overfitting more effectively than a single decision tree model by constructing a model which is adapted bootstrap aggregating (bagging) algorithm. The bagging algorithm makes several decision tree models by sampling some of training dataset and aggregates each result of trees by averaging or voting (Breiman, 2001; Lee et al., 2021). Random Forest classifier is relatively a new method when compared with other supervised classifiers such as k-Nearest Neighbor classifier, Maximum Likelihood classifier, etc. (Feng et al., 2015).

Noi and Kappas (2017) has conducted flood mapping using several classifiers including Random Forest classifier with Sentinel-2 multispectral imagery. They compared the accuracy of classifiers with different size of training datasets, and have shown that the accuracy tends to improve as training dataset was increased (Noi and Kappas, 2017). Lee and Jeong (2020) also performed Random Forest using Sentinel-2 imagery, which showed that multi-temporal image classification showed better accuracy than a single image classification. These studies evaluated the performance of Random Forest classifier through optical satellite imagery. However, there are few studies related to verification of Random Forest classifier performance based on SAR imagery for flood mapping (Feng et al., 2015).

In this study, flood mapping was performed using Random Forest classifier with multi-temporal SAR imagery. We aim to figure out the effect of the size of training dataset and the number of images on classification through Random Forest algorithm. First, Random Forest was performed by varying the size of the training datasets from homogeneous region, and its performance was measured in terms of classification accuracy. Second, Random Forest was performed while constantly increasing the number of time-series images with regular time interval.

2. Study Area and Materials

1) Study Area

The city of Khartoum is the capital of Sudan in Africa (latitude: 15° 37′ 16.4″N, longitude: 32° 30′ 13.27″E) as shown in Fig. 1. Khartoum is located at a point of confluence of the Blue and the White Nile. The rainy season is from March to October every year. Especially, precipitation is concentrated between July and September. Khartoum experiences periodic flood every year due to these climatic and topographical characteristics. UNDRR estimated that between 2016 and 2019, more than 250,000 people in all 18 states of Sudan had been directly affected by flood. The Nile flood in 2020 is considered to be the worst ever recorded flood. According to the Sudan Meteorological Administration, on July 29, 2020, the daily precipitation of Sudan was 44 mm, about three times the daily average for the rainy season. As a result, it is estimated that more than 875,000 people were affected by the flood of Sudan in 2020, and approximately 3.3 billion USD in property was damaged across buildings and infrastructures (OCHA, 2020). According to the Food and Agriculture Organization of the United Nations (FAO), the flood caused about 2.2 million hectares of agricultural land to be flooded in Sudan (FAO, 2020).

Fig. 1. Location of Khartoum is marked in the box on the top left. The arrows indicate the flow direction of the river. Background was derived from Google Earth image taken at 2022.

2) Materials

SAR images can detect the surface of the Earth effectively with less influence from solar altitude and weather conditions due to the use of microwaves (Bamler, 2000). This can be a huge advantage for flood analysis in a cloudy rainy season over optical satellite images (Kussul et al., 2011). Sentinel-1B SAR images operating in European Space Agency (ESA) were used in this study. It was launched on April 25, 2016, equipped with a C-band, and has repetition cycle of 12 days. A total of 35 Sentinel-1B Single Look Complex (SLC) images were acquired from January 15, 2020 to March 10, 2021 at the Khartoum. We processed the data with Sentinel Application Platform (SNAP) software by ESA. Table 1 shows the description of Sentinel-1B SAR used in this study, and Fig. 2 shows dates of dataset and their baseline. The range of perpendicular baseline is –73 m to 143 m. We conducted classification based on SAR amplitude multi-temporal images. With small range of perpendicular baseline, it would conduct more accurate classification with less geometric distortion.

Table 1. Description of the dataset used in this study

Fig. 2. Acquisition dates and perpendicular baseline of Sentinel-1B dataset. The colored period is the flood duration.

3. Methods

1) Image Processing and Training Dataset

Fig. 3 shows the whole workflow of this study. First of all, before performing supervised classification, we need to know the duration of the flood and features of the study area. Hence, the first step of this study is to visually analyze the flood with RGB composite image of before, during, and after the flood. We also obtained a land cover map by performing unsupervised classification using all SAR dataset. We then acquired a training dataset to be used as an input for this study to evaluate the performance of Random Forest classifier in terms of the size of class and the number of dataset.

Fig. 3. The workflow of this study

Sentinel-1B SAR images acquired in SLC images were converted into Ground Range Detected (GRD) images. GRD image contains the detected amplitude and is composed of square pixels with 10 m sampling interval with reduced speckle noise (Filipponi, 2019). Water at calm weather have smooth surface, hence, microwaves are reflected in the opposite direction of the antenna. Therefore, there are little backscattering for water which appears black in SAR images (Kussul et al., 2011). We then applied Range Doppler terrain correction by using Shuttle Radar Topography Mission 1 Sec Height (SRTM 1 Sec HGT).

Three different dates of images were displayed into red, green, and blue channels to create an RGB composite image. With changes of amplitude, it can describe visually the progress of floods in specific duration (Sanyal and Lu, 2004). Images of before, during and after the flood were fed into the red, green, and blue channel, respectively, for visual inspection.

We then obtained a land cover map of Khartoum through unsupervised classification. Unsupervised classification is a process of classifying pixels with similar spectral properties into unknown classes without any prior information (Xue et al., 2015). We used Expectation-Maximization Cluster Analysis (EMCA) as an unsupervised classification. EMCA algorithm repeats between computing probabilities for assignments to each cluster (E-step) and re-calculating the cluster mean and covariance so that the resulting data likelihood function is maximized (M-step) (Do and Batzoglou, 2008). We set the number of clusters to 20, the number of iterations to 30, and input all 35 of the preprocessed images. We then reclassified the classification result into 8 classes by merging the similar land cover type.

2) Random Forest Classifier

Supervised classification is a process of classifying pixels with the most similar properties to known objects into a single labeled class (Xue et al., 2015). As we used amplitude-based GRD images in this study, the classifier would classify the pixels by thresholding of amplitude.

We used Random Forest classifier as a supervised classification. Random Forest is an algorithm in which several decision tree models are trained differently. The results are predicted independently by each model and are derived as a single prediction result through multiple results (Breiman, 2001).

Supervised classification requires training datasets. To obtain an elaborate training dataset, we used the land cover map, an RGB image, and Google Earth high resolution optical satellite images. For dry areas which are not related to flood, such as bare soil, vegetation, urban and buildings, Google Earth high-resolution optical satellite images were referred. For flooded areas and water, we determined the training area where the surface is homogenous and as the most overlapping area in superimposing the unsupervised classification image and the RGB composite image. Random Forest was performed in two independent stages. First, we created 7 different training datasets to observe changes in classification results while varying the size of the training dataset in homogeneous regions. The number of pixels of training datasets was increased from 10 × 10 to 70 × 70 with a 10-pixel step. Second, we performed Random Forest by increasing the number of images in regular temporal intervals. We observed changes in classification results and accuracy according to variance of the period of flood duration in dataset. The three images which contain the outbreak of the flood in July 2020 and two other images before and after the flood were used. The number of image is then added to increase the number of images as shown in Table 2.

Table 2. Number of images and periods used in classification

After the classification, we calculated accuracy to evaluate quantitatively the performance of Random Forest classifier. The accuracy is constituted the ratio of the number of pixels that are exactly classified over the whole number of pixels as shown below Eq. (1) (Fawcett, 2006; Vakili et al., 2020).

\(\text {Accuracy}= \frac { T P + T N } { T P + T N + F P + F N }\)       (1)

The reference data used in this study, however, is not a ground truth actually because we used the land cover map derived from unsupervised classification and visual inspection of SAR images. Therefore, the accuracy dealt with in this study could not indicate how similar the classification result is to actual land cover, but to the training dataset in the following Section 4-1. We then compared the classification result according to the change of each condition and observed the trend of accuracy in Section 4-2 and 4-3.

4. Results and Discussion

1) Training Dataset for Flood

To make a training dataset, a visual inspection was performed by using multi-temporal Sentinel-1B SAR images. We defined the flooded area as the area which had high digital number (DN) value before the flood but became low after the flood. Areas where amplitude was low and then raised again were designated as the drained areas after the flood. The point of completion of drainage in the river was defined as the point when restored to the similar DN value before the flood. We divided events and their period depending on whether it was flooded or drained. Table 3 shows events and the number of images during each event and their period. The flood occurred on July 25, 2020 and drainage started on October 17, 2020. After October 17, the water level of the river decreases until December 28, 2020. The duration of the flood and the continuous drainage are approximately 73 days.

Table 3. Description of number of images and events by period

Table 4 describes states of the surface according to the color of the RGB composite image. Fig. 4(a), (b), and (c) are SAR amplitude images of July 13, October 17, and November 22, 2020, which represent the before, after, and completion of flood respectively. Fig. 4(d) is the RGB composite image of them. Before the flood, the area had higher amplitude than water body to have high value in the red channel. After the outbreak of flood, the amplitude of the flooded area became low due to water body so the green and blue channel have low value. Hence, the flooded area shows red in RGB composite image in Fig. 4(d).

Table 4. Description of each land surface according to the color of the RGB composite image

Fig. 4. SAR amplitude images acquired on (a) July 13, 2020,(b) October 17, 2020,(c) November 22, 2020, and (d) RGB composite image of (a-c).

Unsupervised classification was then performed by using EMCA. Total 20 classified clusters were reclassified into 8 land cover type according to the similarity of pixel properties. The result of EMCA and reclassified image are shown in Fig. 5. Water was classified into one cluster. Flooded areas were reclassified into 3 types by the duration of floods: flooded area, slow drainage, and fast drainage. In amplitude image, fast drainage was occurred from October 17 to October 29, 2020. Slow drainage was occurred from November 10 to December 28, 2020. Within fast and slow drainage, the water level decreased 50 cm respectively. The flooded area indicates areas that have not been drained even long after the floods. The class named water represents areas that always are water regardless of the flood event such as a river. Table 5 shows land covers, their rate, and the number of clusters derived from unsupervised classification. With the help of EMCA image and the RGB composite image, we created 7- different-sized training datasets (Fig. 6).

Fig. 5. Results of EMCA. (a) EMCA image classified into 20 clusters, (b) reclassified image of (a) into 8 classes.

Table 5. Land cover of the study area derived from EMCA

Fig. 6. Training dataset with various classes. Each box has a size of 70 × 70 pixels.

2) Random Forest - in terms of the size of training datasets

Fig. 7 shows the accuracy of Random Forest classifier in terms of the size of the training dataset. The accuracy of flooded area is the highest at about 99.8% on average. Fast and slow drainage area showed high accuracy of 99.87% and 99.85% on average. The class of water was about 99.62% on average. All of the accuracy was high and there was not much difference between them. The range between the accuracy is only 0.73% point max. Fig. 8 is a result of the classification. The image was marked only for flooded areas and water, which are the regions of interest in this study. The size of training dataset has little effect on the classification result as long as the training dataset was well-crafted to represent homogeneous region. This result suggests that once the training region is decided, the burden on window size selection would be reduced. It is also expected to reduce the processing time when a rapid flood mapping is needed.

Fig. 7. Accuracy ofRandom Forest classification according to the size of training dataset.

Fig. 8. The result of Random Forest classification using 40 x 40-sized training dataset.

3) Random Forest –in terms of the number of images

Contrary to the result of Section 4-2, there was a notable difference in the results of classification while changing the number of images. Accuracy of flooded, drained area and water are shown in Fig. 9. The overall accuracy increased as the number of images increases. Starting from the lowest 90.54% with using 3 images, accuracy rose to 94.96% with 7 images and 96.06% with 11 images. After this point, the dataset contains images of the beginning of the drainage. For 15 and 19 images, the accuracy was 98.20% and 98.43% respectively. After this point, the dataset contains images of the completion of drainage. The range of accuracy converges to about 98.80% with the increase of the number of image. Hence, we determined the point of convergence of all accuracy was when using 15 images. This results indicate that accuracy saturates and would not increase significantly afterward if the dataset contains sufficient images of drainage.

Fig. 9. Accuracy of classification according to the number of images.

When using 3 images in Fig. 9, the accuracy of all classes was the lowest. As a result of using 7 images, the accuracy of flooded area and water (we call it FW) sharply increased by 13.08% points and 8% points respectively, when compared with those using 3 images. As the number of images increased from 11, the accuracy of the two classes gradually increased and started to converge. For fast drainage area and slow drainage area (we call it DA), their accuracy increased slowly until 11 images. Then it jumped by 6.8% points and converged when using 15 images.

As shown in Fig 9, FW and DA show different trends in accuracy: FW increased steadily while DA increased sharply as the number of images changed from 11 to 15. FW need 23 images to be saturated while DA need 15 only. This is due to the difference in the stability of shapes and the SAR amplitude of the FW and DA in time-series images. FW have a relatively constant boundary before and after the flood while the shape and size of DA changed rapidly during the drainage process. For FW, the signal to noise ratio (SNR) is low due to flooded water and is a more difficult target for the classifier. This leads to the steady increase of FW accuracy as more information is added. After the drainage for DA, however, the SAR amplitude increases rapidly so does the SNR of the drained area. No further information is required for DA after the drainage.

Fig. 10 shows classification result images. For the result of 3 images, the boundary of the river was divided, but some of water surfaces were recognized as flooded areas (Fig. 10(a)). In addition, boundaries between flooded areas are not clear. It is possibly attributed to the reason of not including enough images to distinguish the boundaries. Until using 11 images, the boundaries between flooded area and water were clearly divided. But drainage areas were still indistinguishable (Fig. 10(b)). From the result using 15 images, all boundaries of each class are separated distinctly (Fig. 10(c)). There was no large difference between the classified images as the accuracies were converged (Fig. 10(d)). In summary, when the dataset includes enough information about events, no additional dataset is necessary.

Fig. 10. Random Forest classification images using (a) 3 SAR images, (b) 7 images, (c) 15 images, and (d) 35 images.

By the way, to the left of the river, there is an area classified as a water class. This area is not actually water but bare soil. This region has a very smooth surface to have a similar backscattering surface as the waterbody in the amplitude in a SAR image. This would be a problem in studies using SAR images and classification techniques. Therefore, care must be taken not to misinterpret these classes and a combination of usage with SAR and optical images would always be beneficial.

5. Conclusions

In this study, we analyzed the effect of temporal and spatial conditions to the accuracy of Random Forest supervised classifier for detecting flooded areas. To make a training dataset, we used land cover obtained from EMCA unsupervised classification and visual inspection from RGB composite image before and after the flood. The result of Random Forest classification while increasing the size of training dataset showed all high accuracy and no difference between classification images. This indicates that training dataset derived from unsupervised classification was well-crafted containing homogeneously enough characteristics that are representative of each class. The result of Random Forest classification while increasing the number of images showed significant difference in accuracies. Generally, the accuracy increased as the number of images increased. However, after the point of the drainage begun, the accuracy saturated and remained constant even though the number of images increased. The accuracy of classes that represent flooded area was improved when the time-series imagery dataset contains flood events. Classification images also showed more clear boundary when using more images.

The advantage of classification using multi-temporal images is that it can detect the shape or extent of flooded areas efficiently according to the duration or velocity of drainage. Knowing exactly which areas have fast drainage and which areas have long flood duration could be the basis for flood risk management. In addition, when using multi-temporal images in classification, there is a limit to increase accuracy even though the large amount of the number of images. We concluded that the point of convergence of accuracy is when the dataset covers well before the starting of the flood to the completion of the drainage.

Instead of using a training dataset from unsupervised classification and visual inspection, a land cover map would have produced better results for this study which requires elaborate manpower or time and is not always available. For such circumstances, it is expected that the result of this study can be applied to flood mapping based on SAR imagery using various supervised classification algorithm including Random Forest in future.