Estimating chlorophyll-a concentration (Chla) in global inland waters is challenged by their diverse material compositions and complex optical properties. Single Chla models often lack universality and may not adequately address the needs of large-scale, diverse water bodies. We collected remote sensing reflectance and Chla data from 23 (n=1383) inland and offshore water bodies worldwide. Following Neil et al.’s classification methodology, we categorized the collected data. Then, we conducted algorithm selection and parameter optimization for Chla estimation in 10 different types of water bodies, focusing on the application of the Sentinel-3 Ocean and Land Color Instrument sensor. After validation with measured data, our models achieved a notable R2 value of 0.93, with a root mean square error of 6.25 (mg/m3) and an average unbiased relative error of 30.10%. These results indicate a significant enhancement in algorithm accuracy, demonstrating wide applicability across complex water bodies globally.

Low-cost and efficient observation devices/methods are always desirable to supplement rainfall observation networks to get high-quality rainfall data. Ubiquitous urban surveillance cameras, recording rainfall video and audio, have shown their huge potential for rainfall estimation. Audio has gained more attention because it can be recorded in weather and has less volume than video, and some surveillance audio-based rainfall estimation (SARE) methods were proposed. However, in urban acoustic scenarios where noise is unavoidable, the current procedure lacks a noise processing part, potentially affecting estimation accuracy. A parallel neural network is put forward to address this noise challenge by employing an attention mechanism. First, one channel in our network is designed to process noises and later cooperates with another to estimate rainfall information. Then, a divide-and-conquer strategy is employed to calculate rainfall intensity. In experiments on the urban surveillance audio dataset, our method achieves a root mean absolute error of 0.949 mm h−1 and a coefficient of correlation of 0.745, outperforming the state-of-the-art method, which demonstrates its effectiveness in mitigating noise in SARE and bringing new insights to the rainfall observation system.

Extraction of water body information from synthetic aperture radar (SAR) images plays a crucial role in urban flood monitoring. Traditional threshold segmentation methods are commonly employed in water body extraction due to the advantages of not requiring labeled samples and high computational efficiency. However, in complex urban terrains, the optimal threshold may be offset by data quality. To address this challenge, we introduce an adaptive iterative thresholding segmentation method guided by optical prior water body information. First, optical images captured before the disaster are used to identify inherent water body areas within the city. Second, the SAR data coverage areas corresponding to the inherent areas are taken as prior information on water bodies during the disaster. Finally, an adaptive iterative thresholding segmentation method based on prior water body information is constructed to automatically extract urban water bodies from SAR images. To validate the effectiveness of this approach, water body extraction experiments in urban inundation zones are conducted using Sentinel-1 and GF-3 SAR data in Beijing, Shijiazhuang, and Zhengzhou. The results show that the overall accuracy of the method used in this study is 79%, 95%, and 97.8% on Sentinel-1 images in Beijing, Shijiazhuang, and Zhengzhou, respectively, and 91.1% on GF-3 images in Zhengzhou. The experimental results in different regions are favorable and have a certain universality. Meanwhile, compared with traditional threshold segmentation methods, this method improves the accuracy of urban water extraction by at least 2% on Sentinel-1 and GF-3 SAR images, providing a more effective technical means for water extraction in urban flood inundation areas.

We aimed to provide an efficient and user-friendly methodology for mapping surface water in rice paddies and wetlands by utilizing a novel combination of high-resolution satellite images from Sentinel-1, the random forests (RF) machine learning algorithm using Sentinel Application Platform software provided by the European Space Agency and texture surface features. We focused on the Thessaloniki plain, with dominant surface water bodies being the rice fields in the broader Chalastra region and Kalochori lagoon. The optimal set of seven features yielding the best results regarding the highest classification accuracy included vertical − vertical (VV) and vertical − horizontal) (VH) polarizations, a combination of these polarizations VV + VH and VV − VH, and texture features based on entropy on both VV and VH polarizations and dissimilarity on VV polarization. Various values for the optimal number of trees in the RF algorithm were tested, including 70, 100, 128, 400, and 500 trees. The most suitable one was found to be the value 128, achieving accuracy levels ranging from 85.3% to 98.3%. Despite potential constraints, such as the occurrence of floating algae on the water’s surface, the proposed methodology can be effectively utilized as an enhanced approach to integrating machine learning algorithms, including texture features into surface water mapping techniques achieving high accuracy.

Wetlands provide important ecosystem functions for water alteration and conservation of biodiversity. Mapping wetland plant communities at a fine scale is crucial for assessing wetland functions and for better management. However, optical remote sensing images are limited by clouds and fog, making it difficult to obtain timely information on wetlands over large areas. Synthetic aperture radar (SAR) time series are widely utilized for mapping and characterizing wetlands by examining the correlation between the distribution of plant community formations and the duration of flooding. We aim to identify the optimal number and crucial dates of SAR images for classification to map wetland plant communities. Using a time series of Sentinel-1 SAR data on Google Earth Engine, a long-term remote sensing images dataset of Poyang Lake in 2019 was constructed. Combining the phenological features of wetland plant communities, we calculated and selected the separability of their monthly backscatter features using the threshold separation method. With the optimal temporal radar classification features identified, the random forest model was then applied to delineate the spatial distribution of wetland plant communities. The results show that time-series SAR images can be used to classify plant communities. SAR backscatter intensity under VH polarization is more sensitive to wetland plant communities than VV polarization, with optimal radar data acquisition dates being January, May, September, and November. The good result achieves a classification accuracy of 88.5% for Poyang Lake wetland plant communities in 2019, along with a Kappa coefficient of 0.821. We introduce a new approach for herbaceous wetland monitoring using SAR imagery, which is of great significance to the ecological protection and restoration of Poyang Lake.

Detection of flooding in rural areas is now commonly performed by high-resolution synthetic aperture radar (SAR) sensors. However, flooding in urban areas causes a greater risk to lives and property. Urban flood detection is more challenging due to SAR shadow, layover, and double scattering effects. Nevertheless, it may often be identified using the fact that in a post-flood image, double scattering between building walls and adjacent floodwater generally exceeds that in a pre-flood image, where double scattering occurs between buildings and adjacent ground. However, in the event of the urban region being deeply flooded, only a part of the building walls may remain above the flood level so that the post-flood double scattering may be reduced and a flooded region may be misclassified as non-flooded. We investigate whether flood detection can be improved for deeper urban floods using interferometric coherence as an adjunct to double scattering. An urban area that is not flooded should often exhibit high coherence between image pairs, whereas if there is flooding in one of the images the coherence should be low. An urban flood in Japan that contained deep-flooded, shallow-flooded, and non-flooded areas was used as a test example. It was imaged by Sentinel-1, and WorldDEM Digital Surface Model data was used to estimate flood depth and building orientation. An analysis of double scatterers of low post-/pre-flood brightness ratio was carried out for deeply flooded and non-flooded urban double scatterers. It was shown that using coherence, 58% of the deeply flooded buildings could be detected at the cost of a 16% false-positive rate. Without the use of coherence to supplement the brightness ratio, all these deeply flooded buildings would be misclassified as non-flooded. This finding could be of use in automating the detection of urban flooding as an aid to flood risk management.

The precise detection of a surface water body using synthetic aperture radar (SAR) images is crucial for flood mitigation, disaster reduction, and water resource planning applications. Although SAR has been proven to have the ability to provide information for water-body detection, a single SAR feature is still insufficient to achieve high-precision classification. To fully leverage the backscatter intensity and polarimetric features of SAR images, we propose the dual-branch fusion network (DBFNet), an innovative semantic segmentation model that integrates the backscatter intensity and polarimetric features. Specifically, the DBFNet employs a distinctive dual-branch architecture that integrates the complementary information of both feature types using the layer feature fusion module and refines multiscale features at various levels through the intermediate feature refinement module. The performance of the proposed DBFNet is evaluated by conducting comparative experiments with five deep learning models: FCN, U-Net, DeepLabv3+, FWENet, and FFEDN. The experimental results demonstrate that the DBFNet achieves the highest accuracy in water-body detection, with an intersection over union of 89.28% and an F1-score of 94.34%.