1. Introduction

1.2. Deep Neural Networks

Deep learning models have become more and more popular nowadays and are constantly evolving. They are also increasingly being applied in recent research, including for tasks such as flood segmentation. Notably [8] does an overview of existing deep learning methods for flood segmentation, and it suggests the preference towards specific model architectures based on their “inductive biases”. The inductive bias refers to the set of assumptions a learning algorithm makes or is expected to make to predict outputs (labels) based on inputs (features). Based on the assumption that in flood analysis, SAR data often exhibit spatial correlations between adjacent pixels, NN (Neuronal Network) layers can be configured to exploit these data structures, by introducing constraints on the connections and interactions between inputs and outputs and favouring certain model architectures over others [8]. One way to exploit spatial information is the use of a convolutional layer, which is the core component of Convolutional Neural Network (CNN).

1.2.1. CNN (Convolutional Neural Network)

Convolutional Neural Networks (CNNs) are a type of model in the field of deep learning, particularly useful at processing data with a grid shape, such as images. Central to CNNs are convolutional layers, which exploit spatial information by capturing the local relationships among pixels through the use of filters. In a convolutional layer, traditional neuron weights are replaced with filters. These filters compute new values for each pixel by considering the local relationships with its neighbors.

Figure 1. Illustration of the convolutional operation, visualized through a 3x3 kernel applied to an input image, that produces a feature map highlighting detected patterns. This process, core of CNNs, enables the network’s to automatically learn spatial relationships of features [14].

Figure 1. Illustration of the convolutional operation, visualized through a 3x3 kernel applied to an input image, that produces a feature map highlighting detected patterns. This process, core of CNNs, enables the network’s to automatically learn spatial relationships of features [14].

The architecture is designed to progressively abstract higher-level features from the raw input in the initial layers to more abstract representations in deeper layers, making CNNs particularly powerful for tasks like image classification.

1.2.2. Segmentation Models: An Overview

The standard CNN architecture can’t be directly used for image segmentation. In fact, as shown in Figure 2 the fully connected layers link the pooling layers to a 1D output, which loses spatial context. To retain the spatial context in the output layer, other segmentation specific architectures have been proposed in the literature. The Fully Convolutional Network (FCN) introduced in [15] utilizes pre-existing CNN architectures designed for classification tasks (such as AlexNet, VGG, etc.). These architectures are adapted into fully convolutional networks and their learned representations are transferred to the segmentation task through fine-tuning. FCNs introduced the idea of replacing fully connected layers with convolutional layers, to produce pixel-wise predictions for segmentation. The architecture consists of an encoder path, which downsamples the input image to extract features, and a decoder path, which upsamples the features to produce the segmentation map. FCNs have been effectively applied in tasks such as flood segmentation, as demonstrated in [16].

In 2015, a novel paper [18] proposed the U-Net architecture which builds upon the ideas introduced by FCNs but adds the concept of skip connections. The idea behind it is that each encoder and decoder layer is symmetrically linked with a skip connection, that is achieved by concatenating the output from encoding layers to corresponding decoding layers. The goal was to provide a more detailed segmentation mask by exploiting global and local features of an image. This concept, initially proposed in [18], was further reproposed a few months later by a new model architecture called ResNet (Residual Nets) [19]. ResNet builds upon the concept of skip connections from Unet [18] but approaches it from a different angle. ResNet [19] solves a common issue encountered in deep neural networks: as the network gets deeper with more layers, the gradient (used to update the weights based on the error) diminishes significantly during backpropagation, leading to difficulties in updating the weights of earlier layers. This is known as the gradient vanishing problem. Recent advances in segmentation combine the power of ResNet deep features extraction into the Unet Encoder, which significantly improve the segmentation performance. This hybrid architecture has been applied for sea-land optical segmentation in [20], demonstrating that replacing ResNet into the Unet architecture significantly improves feature representation. This enhancement, together with fully connected Conditional Random Fields (CRFs) as post-processing predictions refinement, leads to more accurate delineation of coastlines and better-preserved regions in sea-land segmentation tasks as shown in [20].

1.2.3. Conditional Random Fields (CRFs) in Segmentation

Conditional Random Fields (CRFs) have been widely used as a post-processing technique to refine segmentation results. Initially developed as a probabilistic framework for segmenting and labelling sequential data [21], CRFs were later adapted into fully connected CRFs to address the challenges of semantic segmentation. However, the widespread adoption of fully connected CRFs was slowed down by the computational demands of their inference processes. The scenario changed with the introduction of an efficient inference algorithm shown in [22], making fully connected CRFs more efficient and viable for practical use especially as a common technique for post-processing in segmentation tasks. By modelling the spatial dependencies between pixels, CRFs improve the coherence and accuracy of the segmentation output, particularly in distinguishing between closely related classes and preserving boundary definitions. Their application, combined with advanced neural network architectures like U-Net and ResNet, represents a comprehensive approach to tackling complex segmentation tasks with improved precision. Notably in [20], it is demonstrated that the use of CRF as a post-processing stage enhances the segmentation predictions results. In our research, we adopt the U-Net architecture integrated with an encoder block derived from the ResNet encoder. Furthermore, our exploration will include the application of CRFs, in alignment with current trends indicated in the literature.

2. Materials and Methods

2.1. Dataset Creation

The Copernicus Programme, through Sentinel satellites, provides free access to satellite data, including Sentinel-1 SAR data. Pre-processing steps are necessary to improve image quality. The article in [29] discusses the challenges and necessary steps involved in preprocessing Sentinel-1 Ground Range Detected (GRD) data. MMFlood [25] uses the Sentinel Hub API, which aggregates multiple satellite data, enabling real-time processing and access to specific AOIs within seconds, thus avoiding extensive preprocessing.

A significant challenge in building the training dataset is the precise and reliable delineation of flood areas, as most flooded pixels are usually selected manually or adjusted based on operator knowledge. The Sen1Floods11 dataset [24] comprises Sentinel-1 and Sentinel-2 data, with 446 labeled 512x512 chips covering 11 flood events, but it does not consider various flood types and relies on human intervention, introducing potential errors. MMFlood [25] uses the Copernicus Emergency Management Service (EMS) to retrieve flood delineations in vector files for their dataset, consisting of 8,522 labeled 512x512 patches from Sentinel-1, covering 95 flood events globally. EMS has operated since 2021, providing maps for environmental disasters based on a closed-source algorithm and then reviewed by humans, but their labels are not always accurate.

Following the approach in [25], we use the Sentinel Hub API to directly obtain satellite products. We then pair Sentinel-1 and Sentinel-2 images to create a new dataset, SAR-Optical, for training our model. Since the dataset only contains uni-temporal flood and permanent water patches, we create a new dataset derived from the MMFlood dataset [25], enhanced with pre-flood images, which we call MMFlood-CD (change detection). This will be used to qualitatively assess the model performance on flood events through change detection.

2.1.1. Images Acquisition and Preprocessing

With the use of Sentinel Hub API, we can specify the pre-processing steps directly in the parameters of the requests. In our experiments we use Ground Range Detected (GRD) imagery, which preserves signal amplitude while eliminating phase information and reducing speckle noise through multi-look processing [29]. All the images are provided with a spatial resolution of 20 m and a pixels spacing of 10 m, and have a 32-bit floating point precision. As per preprocessing steps, we apply calibrations and thermal noise removal as outlined in the standard workflow proposed in [29]. Moreover, we apply Radiometric Terrain Correction (RTC) and Orthorectification using the MapZen Digital Elevation Model to more accurately geocode the images. We purposely avoid applying speckle filtering, based on the hypothesis, aligned with the findings of [25], that CNNs models inherently possess the capabilities to smooth images to extract relevant information.

Figure 4. Summary of the preprocessing steps involved to prepare the GRD and Optical products applied directly in Sentinel Hub. The API accounts for both data retrieval and preprocessing.

Figure 4. Summary of the preprocessing steps involved to prepare the GRD and Optical products applied directly in Sentinel Hub. The API accounts for both data retrieval and preprocessing.

2.1.2. SAR-Optical

The SAR-Optical dataset contains paired SAR and Optical patches, and it is used to train the SAR model using the Optical ground truth. The image pairing was not straightforward, since it’s unusual to have both Optical and SAR images taken on the same day, we made sure that the Optical image acquisitions were taken within a maximum of three days after the flood event. This is a good compromise that allows a balanced number of samples while accounting for rapid changes in floods. Clouds may obstruct the regions of Interest (ROIs), therefore we selected only those that don’t have clouds and that have at least 0.01% of water pixels. The Optical ground truth is derived from the SCL layer contained in the Sentinel-2 L2A Optical product, provided directly by Copernicus. Their layer automatically classifies various land covers such as water, land, clouds, etc., and is consistently updated. We extracted only the water classification from this layer to generate a binary raster image that marks pixels as 1 where water is present and 0 where it is not. Unlike the approach taken by DeepAcqua [27], we opt for the SCL layer as it presents a more reliable alternative to the NDWI calculation, which can occasionally miss water regions.

Figure 5. The paired Optical and SAR image, to facilitate the understanding of the SAR-Optical dataset structure. The dataset contains paired Optical and SAR images. The Optical image contains the SCL band from Sentinel-2 mission, Copernicus. This band provides a classification of the water extent which will be used as ground truth to train the SAR model.

Figure 5. The paired Optical and SAR image, to facilitate the understanding of the SAR-Optical dataset structure. The dataset contains paired Optical and SAR images. The Optical image contains the SCL band from Sentinel-2 mission, Copernicus. This band provides a classification of the water extent which will be used as ground truth to train the SAR model.

2.1.3. MMFlood-CD

The MMFlood-CD dataset contains the same flood events delineations and naming conventions of MMFlood dataset [25], with the addition of pre-flood images. This dataset is utilized solely for a qualitative visual comparision of our SAR model with the EMS ground truth, included in MMFlood-CD. Given that the objective of the SAR model is to detect water, regardless of it being floodwaters or permanent water, we take advantage of both pre- and post-flood images and use a change detection module to detect flood. A more detailed explaination of this process can be found in Section 2.5.

Figure 6. Diagram to facilitate the understanding of the MMFlood-CD (CD stands for change detection) dataset structure. The dataset contains the same flood events delineation from MMFlood but with the addition of pre-flood. It also contains the EMS ground truth, that will be used just to qualitatively compare our model with their performance

Figure 6. Diagram to facilitate the understanding of the MMFlood-CD (CD stands for change detection) dataset structure. The dataset contains the same flood events delineation from MMFlood but with the addition of pre-flood. It also contains the EMS ground truth, that will be used just to qualitatively compare our model with their performance

In the following table a summary of the 3 datasets discussed in the previous sections with a comparison highlighting the differences.

Table 1. Dataset Comparison.

Table 1. Dataset Comparison.

Dataset	N. Events	Ground Truth	Pre-Flood Images	Labels
MMFlood	1,748	EMS	No	Flood
MMflood-CD	1,748	EMS	Yes	Flood
SAR-Optical	823	SCL Optical	No	Water and Flood

2.2. Training Pipeline

The task of flood delineation involves categorizing each pixel within an image as either flood or non-flood (background). Specifically, we define a collection of image pairs, x, where each pair, $x_s$, consists of SAR images with uniform dimensions $H\times W$. These images are paired with corresponding labels, $y\in Y$, also of the same dimensions, derived from optical images that serve as the ground truth. For every pixel i, its label $y_i$ can be either 0 or 1, with 0 indicating non-flood (background) regions and 1 indicating flood-affected areas. The goal of binary segmentation training is to develop a model, $f_{\theta }$, parameterized by $\theta$, which accurately predicts flood pixels across the image, mapping the input, SAR image, to a corresponding label, $f_{\theta }:X\to {\mathbb{R}}^{H\times W}$. Here, ${\mathbb{R}}^{H\times W}$ denotes the output of the model, represented by a matrix of size $H\times W$ containing the raw scores (logits). These scores are subsequently converted into probabilities using a SoftMax function. Further details on SoftMax functionalities can be found in Section 2.5. The model is trained on the labels from the SAR-Optical dataset, which contains SAR images paired with the Optical ground truth. The Optical SCL Layer transfers the knowledge to the SAR model through a weighted cross-entropy function that minimizes the difference in the distributions of the two models output, making the SAR model predictions as close as possible to the Optical one.

Figure 7. Diagram of the training pipeline. The SAR model is trained on the paired SAR-Optical dataset. The predictions generated from the Optical Model are used as ground truth for the SAR Model, with the use of a Cross-Entropy Loss function.

Figure 7. Diagram of the training pipeline. The SAR model is trained on the paired SAR-Optical dataset. The predictions generated from the Optical Model are used as ground truth for the SAR Model, with the use of a Cross-Entropy Loss function.

2.2.1. Loss Functions for Data Imbalance

We observed a significant imbalance with a water-to-land ratio of 1:11. To tackle this imbalance without altering the dataset with down sampling or up sampling techniques, we opted for specific loss functions that mitigate this issue by adjusting the weights between water and non-water classes. We explored two specific loss functions for this purpose: Focal Tversky Loss [30] and Weighted Cross-Entropy Loss.

Weighted Cross-Entropy Loss

The Weighted Cross-Entropy Loss is an adaptation of the standard Cross-Entropy Loss, introducing a weighting factor to account for class imbalances. This is particularly useful in datasets where one class is significantly more frequent than the other, which can lead to a model bias towards the majority class. The weighted version is defined as:

$$WeithedCrossEntrophyLoss=-\sum _{i=1}^N{w}_i·y_i·log\left(p_i\right)$$

(1)

Here, w represents the weight assigned to the positive class. This weight is typically set inversely proportional to the class frequencies, meaning that a higher weight is assigned to the less frequent class. This adjustment aims to amplify the loss contributed by the minority class, encouraging the model to pay more attention to correctly classifying these instances.

Focal Tversky Loss

In our experiments, inspired from previous studies [25,31] we also explored the Focal Tversky Loss [30], which unlike the traditional Cross-Entropy Loss, aims directly at optimizing the Intersection over Union (IoU) score, aligning closely with our segmentation objectives. The Focal Tversky Loss is defined as:

$$FTL={(1-TI)}^{{\displaystyle \frac{1}{\gamma }}}$$

(2)

Here, $TI$ represents the Tversky index, and $\gamma$ is a modulating factor parameter that diminishes the influence of easily classified examples, thereby making the model focus more on hard to classify examples. The Tversky index $TI$ is a generalization of the IoU metric that introduces two parameters, $\alpha$ and $\beta$, to control the relative importance of false positives and false negatives. Specifically, is given by:

$$TI={\displaystyle \frac{TP}{TP+\alpha FP+\beta FN}}$$

(3)

In this formula, $TP$, $FP$, and $FN$ stand for true positives, false positives, and false negatives, respectively. This loss is specifically designed to account for the imbalance in the labels, in this case between water and non-water classes, by applying a dynamic balancing mechanism. The weighting factor $\alpha \in \left[0,1\right]$ for water and $\beta$ for non-water moderates the significance of respective pixels, with a higher $\alpha$ emphasizing water pixels

2.5. Change Detection Module for Flood Analysis

Our model is designed to identify areas covered by water, without specifically distinguishing between flooded and non-flooded regions. To highlight the extent of flooding, we compute the difference between the predictions made before and after the flood event, through the pre- and post-flood images. The model’s predictions are first processed through a SoftMax function to convert the raw logits from the output layer into probabilities. The SoftMax function in simple terms, is a mathematical operation that transforms a vector of real-valued logits into a vector of probabilities by exponentiating each logit and then normalizing those values. It ensures that the output values are in the range (0, 1) and sum up to 1, making them a valid probability distribution.

Figure 8. Example of how SoftMax converts the raw logits from a vector output to probabilities ranging between 0-1.

Figure 8. Example of how SoftMax converts the raw logits from a vector output to probabilities ranging between 0-1.

After applying the SoftMax function, we calculate the difference between the post-flood and pre-flood predictions probabilities. Pixels with a difference exceeding a predetermined threshold, empirically set at 0.5, are classified as flooded, while those below this threshold are identified as not flooded. Mathematically this methodology can be expressed as follow:

Let $Z_{water}\left(i,j\right)$ and $Z_{nowater}\left(i,j\right)$ , represent the SoftMax outputs for the classes "water" and "non-water" at each pixel location $\left(i,j\right)$, where i and j denote the pixel coordinates in terms of width and height, respectively. The SoftMax function ensures that for any given pixel, the sum of these probabilities equals one, such that:

$$Z_{water}\left(i,j\right)+Z_{nowater}\left(i,j\right)=1$$

(4)

Let’s denote the pre- and post-flood scenarios as $f_{pre}\left(i,j\right)$ and $f_{post}\left(i,j\right)$. The difference between the post-flood and pre-flood predictions indicated as $\Delta f$, is computed as:

$$\Delta f\left(f_{pre},f_{post}\right)=f_{post}\left(i,j\right)-f_{pre}\left(i,j,\right)$$

(5)

A pixel is classified as "flooded" if $\Delta f>0.5$ ; otherwise, it is classified as "not flooded". This methodology ensures that the classification is based on a probabilistic approach, with the threshold value determining the sensitivity of flood detection. Rather than relying solely on absolute binary values of 0 and 1, this approach is expected to be more robust and can be tuned through a threshold, so that it can capture water pixels classified with lower probabilities.

2.6. Inference and Evaluation

The model’s performance in segmentation tasks is often evaluated using a variety of metrics, among which the Intersection Over Union (IoU), also known as the Jaccard Index, is the most used. This metric is particularly crucial because it provides a comprehensive measure of how accurately the predicted segmentation aligns with the ground truth. The IoU metric is defined as the ratio of the intersection area between the predicted segmentation and the ground truth to the union area of these two segments [34].

$$IoU={\displaystyle \frac{AreaofIntersection}{AreaofUnion}}$$

(6)

This equation quantifies the overlap between the predicted and actual segmentations. The values of IoU range from 0 to 1, where a higher value indicates a greater overlap and, consequently, a higher model accuracy. Unlike simple accuracy, which might be misleading especially in cases with class imbalance, IoU provides a more balanced and accurate evaluation of model performance.

The SAR model selection process is conducted through a k-fold cross-validation in the SAR-Optical dataset. To choose the best performing model, we utilize 5 fold cycles to train and test the model, allocating 70% for training and 30% for testing in each cycle. Each training cycle is trained for 100 epochs and we implement an early stopping mechanism set to stop when the model does not improve over 10 consecutive epochs. At this point, the model with the best performance is selected, to be tested on MMFlood-CD dataset.

In this phase, the SAR model’s predictions from the change detection module, applied to pre- and post-flood events, are compared with the EMS ground truth to evaluate performance. The process involves feeding the model two distinct images separately to generate pre- and post-flood predictions. These predictions are then processed by the change detection (CD) module to compute the differences between them. The Conditional Random Field (CRF) algorithm is then applied for post-processing refinement, to generate the final flood prediction. This prediction is visually compared with the EMS ground truth for performance assessment.

Figure 9. Diagram to facilitate the understanding of the inference pipeline in the MMFlood-CD dataset.

Figure 9. Diagram to facilitate the understanding of the inference pipeline in the MMFlood-CD dataset.

3. Results

3.1. Models Benchmark Results

In the table below, various models are compared against the Otsu Thresholding baseline, with the Unet + ResNet50 combination, highlighted in gray, outperforming the other models. To provide a comprehensive overview, four metrics are considered: Precision, Recall, IoU, and F1 Score. The reported values are the mean results from 5-fold cross validations.

Table 2. Model benchmark results.

Table 2. Model benchmark results.

Model Architecture	Accuracy	Recall	IoU
Otsu Thresholding	0.78 ± 0.03	0.51 ± 0.02	0.27 ± 0.00
FCN16	0.79 ± 0.01	0.50 ± 0.01	0.41 ± 0.02
Unet	0.86 ± 0.03	0.82 ± 0.04	0.62 ± 0.01
Unet + ResNet50	0.96 ± 0.02	0.95 ± 0.02	0.72 ± 0.02
Unet + DenseNet201	0.85 ± 0.01	0.83 ± 0.01	0.63 ± 0.03
Unet + VGG19	0.79 ± 0.05	0.78 ± 0.04	0.59 ± 0.03

The Otsu Thresholding method, being used as the baseline, demonstrates relatively low performance across all metrics. It achieves an accuracy of 0.78, recall of 0.51, and IoU of 0.27, highlighting its limitations in more complex segmentation tasks. Among the tested models, the Fully Convolutional Network (FCN16) shows a slight improvement over the baseline with an accuracy of 0.79, recall of 0.50, and IoU of 0.41. This indicates its potential but also its limitations in effectively segmenting the data compared to more advanced architectures. The Unet model alone significantly outperforms both the baseline and FCN16, achieving an accuracy of 0.86, recall of 0.82, and IoU of 0.62. This underscores Unet’s capability to handle segmentation tasks efficiently, particularly with relatively small amounts of data. The highlight of the benchmark results is the Unet + ResNet50 combination, which outperforms all other models with an accuracy of 0.96, recall of 0.95, and IoU of 0.72. This combination is also the only one which benefits from the second image processing (b) detailed in Section 2.4. Other Unet combinations, such as Unet + DenseNet201 and Unet + VGG19, also show commendable results but they achieve poor performance when trained with the same image processing technique (b). This underperformance may be attributed to their deep and complex architectures, which are more prone to overfitting. While these models perform well, they do not match the exceptional performance of the Unet + ResNet50 combination.

3.2. Results on SAR-Optical Dataset

This section analyzes the performance of the top model, Unet + ResNet50. During the EMSR442-0-6 flood event, despite cloud cover obscuring the Optical ground truth, the SAR model successfully identified the water regions. Additionally, the model excels in EMSR280-4-17 event by distinguishing snow-covered areas, which typically appear similar to water in SAR images. Here’s a detailed explanation of the labels in the images:

• SAR Image – The input SAR image.
• Optical (SCL) Ground Truth – Optical ground truth from Sentinel-2 SCL layer, provided by Copernicus.
• Model Prediction - The model prediction (presence of water, whether it’s due to flooding or permanent sources).

An example where the model underperforms the Optical, SCL ground truth, is given by the flood event in EMSR258-13-1. As shown in Figure 10, it’s evident that the model predictions are less accurate along the borders, likely due to the lower resolution of SAR resolution compared to the Optical ground truth.

Figure 10. Performance Analysis of EMSR258-13-1 Permanent Water in Albania.

Figure 10. Performance Analysis of EMSR258-13-1 Permanent Water in Albania.

Figure 11. Performance Analysis of EMSR407-1-4 Flood Event in the United Kingdom. Event Date: November 12, 2019 at 21:15.

Figure 11. Performance Analysis of EMSR407-1-4 Flood Event in the United Kingdom. Event Date: November 12, 2019 at 21:15.

Figure 12. Performance Analysis of EMSR277-2-0 Flood Event in Greece. Event Date: March 29, 2018 at 11:27.

Figure 12. Performance Analysis of EMSR277-2-0 Flood Event in Greece. Event Date: March 29, 2018 at 11:27.

Figure 13. Performance Analysis of EMSR501-0-3 Flood Event in Albania. Event Date: February 12, 2021 at 16:50.

Figure 13. Performance Analysis of EMSR501-0-3 Flood Event in Albania. Event Date: February 12, 2021 at 16:50.

Overall, the model has demonstrated great performance. An example is shown in flood event EMSR442-0-6. Illustrated in Figure 14, this event highlights a section of water missing from the Optical (SCL) ground truth due to cloud cover obstruction. However, given SAR capacity of penetrating through clouds, the model was able to accurately detect those flood areas.

Another notable performance is seen in the flood event EMSR280-4-17. As illustrated in Figure 15, this event demonstrates the model’s capability to identify and exclude snow-covered areas. In SAR images, regions covered in snow closely resemble water, making them indistinguishable to the human eye. However, the model successfully distinguished these areas avoiding the misclassification of snow as water.

Figure 15. Performance Analysis of EMSR280-4-17 Flood Event in Sweden. Event date: April 21, 2018 at 15:15.

Figure 15. Performance Analysis of EMSR280-4-17 Flood Event in Sweden. Event date: April 21, 2018 at 15:15.

Figure 16. RGB image of the EMSR442-0-6 flood event, from the Optical Sentinel-2 Satellite. It demonstrates the presence of cloud cover.

Figure 16. RGB image of the EMSR442-0-6 flood event, from the Optical Sentinel-2 Satellite. It demonstrates the presence of cloud cover.

Figure 17. RGB image of the EMSR280-4-17 flood event, from the Optical Sentinel-2 Satellite. It demonstrates the presence of snow

Figure 17. RGB image of the EMSR280-4-17 flood event, from the Optical Sentinel-2 Satellite. It demonstrates the presence of snow

4. Discussion

5. Conclusions

References

1. On the Assessment and Management of Flood Risks (Text with EEA relevance), 2007. Available: https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:32007L0060&from=EN.
2. Lendering, K.T.; Jonkman, S.N.; Kok, M. Effectiveness of Emergency Measures for Flood Prevention. Journal of Flood Risk Management 2015, 9, 320–334. [CrossRef]
3. Tsokas, A.; Rysz, M.; Pardalos, P.M.; Dipple, K. SAR data applications in earth observation: An overview. Expert Systems with Applications 2022, 205, p. 117342. [CrossRef]
4. Shao, Y. Effect of dielectric properties of moist salinized soils on backscattering coefficients extracted from RADARSAT image. IEEE, 2003. Available: https://ieeexplore.ieee.org/abstract/document/1221799.
5. Amitrano, D.; Di Martino, G. Unsupervised Rapid Flood Mapping Using Sentinel-1 GRD SAR Images. IEEE Journals & Magazine, 2018. Available: https://ieeexplore.ieee.org/abstract/document/8291019.
6. Schumann, G.; Di Baldassarre, G.; Alsdorf, D.; Bates, P.D. Near real-time flood wave approximation on large rivers from space: Application to the River Po, Italy. Water Resources Research 2010, 46, no–5. [Google Scholar] [CrossRef]
7. de, W. et al. Analysis of Environmental and Atmospheric Influences in the Use of SAR and Optical Imagery from Sentinel-1, Landsat-8, and Sentinel-2 in the Operational Monitoring of Reservoir Water Level. Remote Sensing 2022, 14, 2218–2218. [Google Scholar] [CrossRef]
8. Bentivoglio, R.; Isufi, E.; Jonkman, S.N.; Taormina, R. Deep learning methods for flood mapping: a review of existing applications and future research directions. Hydrology and Earth System Sciences 2022, 26, 4345–4378. [Google Scholar] [CrossRef]
9. Qiu, J.; Cao, B.; Park, E.; Yang, X.; Zhang, W.; Tarolli, P. Flood Monitoring in Rural Areas of the Pearl River Basin (China) Using Sentinel-1 SAR. Remote Sensing 2021, 13, 1384. [Google Scholar] [CrossRef]
10. Liang, J.; Liu, D. A local thresholding approach to flood water delineation using Sentinel-1 SAR imagery. ISPRS Journal of Photogrammetry and Remote Sensing 2020, 159, 53–62. [Google Scholar] [CrossRef]
11. Chini, M. A Hierarchical Split-Based Approach for Parametric Thresholding of SAR Images: Flood Inundation as a Test Case. IEEE, 2017. Available: https://ieeexplore.ieee.org/document/8017436.
12. Amitrano, D.; Di Martino, G.; Di Simone, A.; Imperatore, P. Flood Detection with SAR: A Review of Techniques and Datasets. Remote Sensing 2024, 16(4), p. 656. [CrossRef]
13. Sarker, C.; Mejias, L.; Maire, F.; Woodley, A. Flood Mapping with Convolutional Neural Networks Using Spatio-Contextual Pixel Information. Remote Sensing 2019, 11, no–19. [Google Scholar] [CrossRef]
14. Syahdani, F.A. Implementasi Convolutional Neural Network Dalam Mendeteksi Cuaca Dari Foto. Master’s Thesis, Universitas Multimedia Nusantara, 2023. Available: https://kc.umn.ac.id/id/eprint/23969/4/BAB_II.pdf.
15. Long, J.; Shelhamer, E.; Darrell, T. Fully Convolutional Networks for Semantic Segmentation. 2015. Available: https://arxiv.org/pdf/1411.4038.pdf.
16. Nemni, E.; Bullock, J.; Belabbes, S.; Bromley, L. Fully Convolutional Neural Network for Rapid Flood Segmentation in Synthetic Aperture Radar Imagery. Remote Sensing 2020, 12, 2532. [Google Scholar] [CrossRef]
17. Oliveira, H.; Silva, G.; Machado, K.; Nogueira, K.; Santos. Fully Convolutional Open Set Segmentation. Machine Learning 2021, 112, 1733–1784. [CrossRef]
18. Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv.org, 18 May 2015. Available: https://arxiv.org/abs/1505.04597.
19. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. arXiv.org, 10 December 2015. Available: https://arxiv.org/abs/1512.03385.
20. Chu, Z. Sea-Land Segmentation With Res-UNet And Fully Connected CRF. In IEEE Conference Publication, IEEE Xplore, 1 January 2019. Available online: https://ieeexplore.ieee.org/document/8900625 (accessed on 26 February 2024).
21. Lafferty, J.; McCallum, A.; Pereira, F. Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data. January 2001. Available: http://www.cs.columbia.edu/~jebara/6772/papers/crf.pdf (accessed on 26 February 2024).
22. Krähenbühl, P.; Koltun, V. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. arXiv.org, 20 October 2012. Available: https://arxiv.org/abs/1210.5644.
23. Yadav, R.; Nascetti, A.; Azizpour, H.; Ban, Y. Unsupervised Flood Detection on SAR Time Series Using Variational Autoencoder. International Journal of Applied Earth Observation and Geoinformation 2024, 126, 103635. [Google Scholar] [CrossRef]
24. Bonafilia, D.; Tellman, B.; Anderson, T.; Issenberg, E. Sen1Floods11: A Georeferenced Dataset to Train and Test Deep Learning Flood Algorithms for Sentinel-1. openaccess.thecvf.com, 2020. Available: https://openaccess.thecvf.com/content_CVPRW_2020/html/w11/Bonafilia_Sen1Floods11_A_Georeferenced_Dataset_to_Train_and_Test_Deep_Learning_CVPRW_2020_paper.html (accessed on 28 October 2021).
25. MMFlood: A Multimodal Dataset for Flood Delineation From Satellite Imagery. Available online: https://ieeexplore.ieee.org/document/9882096 (accessed on 19 February 2024).
26. Paul, S.; Carted, S.; Ganju. Flood Segmentation on Sentinel-1 SAR Imagery with Semi-Supervised Learning. Accessed: 27 February 2024. Available: https://arxiv.org/pdf/2107.08369.pdf.
27. Peña, F.J.; Hübinger, C.; Payberah, A.H.; Jaramillo, F. DeepAqua: Semantic Segmentation of Wetland Water Surfaces with SAR Imagery Using Deep Neural Networks Without Manually Annotated Data. International Journal of Applied Earth Observation and Geoinformation 2024, 126, 103624. [Google Scholar] [CrossRef]
28. Main-Knorn, M.; Pflug, B.; Louis, J.; Debaecker, V.; Müller-Wilm, U.; Gascon, F. Sen2Cor for Sentinel-2. Image and Signal Processing for Remote Sensing XXIII, 2017. [CrossRef]
29. Filipponi, F. Sentinel-1 GRD Preprocessing Workflow. In Proceedings of the Proceedings, vol. 18, no. 1, p. 11, June 2019. [CrossRef]
30. A Novel Focal Tversky Loss Function With Improved Attention U-Net for Lesion Segmentation. In IEEE Conference Publication, IEEE Xplore. Available online: https://ieeexplore.ieee.org/document/8759329 (accessed on 1 March 2024).
31. Bai, Y. et al. Enhancement of Detecting Permanent Water and Temporary Water in Flood Disasters by Fusing Sentinel-1 and Sentinel-2 Imagery Using Deep Learning Algorithms: Demonstration of Sen1Floods11 Benchmark Datasets. Remote Sensing 2021, 13, 2220. [Google Scholar] [CrossRef]
32. offe, S.; Szegedy, C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. arXiv.org, 2015. Available: https://arxiv.org/abs/1502.03167.
33. Buslaev, A.; Iglovikov, V.I.; Khvedchenya, E.; Parinov, A.; Druzhinin, M.; Kalinin, A.A. Albumentations: Fast and Flexible Image Augmentations. Information 2020, 11, 125. [Google Scholar] [CrossRef]
34. Rezatofighi, H.; Tsoi, N.; Gwak, J.; Sadeghian, A.; Reid, I.; Savarese, S. Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression. Thecvf.com, 2019, pp. 658–666. Available: https://openaccess.thecvf.com/content_CVPR_2019/html/Rezatofighi_Generalized_Intersection_Over_Union_A_Metric_and_a_Loss_for_CVPR_2019_paper.html.