Super Resolution by Deep Learning Improves Boulder Detection in Side Scan Sonar Backscatter Mosaics

Feldens, Peter

doi:10.3390/rs12142284

Cited by 11 publications

(7 citation statements)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This situation may improve as synthetic aperture sonar surveys (Hayes and Gough, 2009) become more widely available, increasing resolution to the centimetre scale. Further underestimation of the number of counted boulders occurs for small boulders represented by few pixels, where detection performance drops off sharply for objects represented by only a few pixels (Ren et al, 2018;Feldens, 2020).…”

Section: Discussionmentioning

confidence: 99%

“…Basinwide boulder detection relies on the manual interpretation of acoustic images (BSH, 2016;Heinicke et al, 2021). To automate the detection of individual boulders, convolutional neural networks (CNN) have been trained to detect boulder-sized and larger objects on the seafloor (Feldens et al, 2019;Michaelis et al, 2019;Feldens, 2020;Feldens et al, 2021;Steiniger et al, 2022). However, these models have not yet been applied to larger-scale investigation sites.…”

mentioning

confidence: 99%

See 1 more Smart Citation

Distribution of boulders in coastal waters of Western Pomerania, German Baltic Sea

et al. 2023

Self Cite

View full text Add to dashboard Cite

This study contributes to a better understanding of geogenic reef distribution in the southern Baltic Sea and highlights the implications of survey-related factors on automated boulder classification when utilizing data from multiple surveys. The distribution of hard grounds and reefs is needed as a baseline for geological and biological studies, but also for offshore construction, navigation and coastal management. In this study we provide maps of the distribution of geogenic reefs for about 750 km2 in the southern Baltic Sea, at the sites Wismar Bay, Darss Sill and Plantagenet Ground. The maps are based on full-coverage backscatter surveys with different side scan sonar and multibeam echo sounder systems. The distribution and number of boulders in the backscatter maps was determined using a convolutional neural network combined with quality control by human experts. The extent of the geogenic reefs was calculated on the basis of the number of boulders in 50 m x 50 m grid cells. We compare the results with previous reef maps based on point sampling, which show reefs of either biogenic or geogenic origin. According to the earlier maps, 11% of the Plantagenet Ground seabed was classified as reef habitat type. This is similar to the result of our study (12%), although we only considered reefs of geogenic origin. In the Darss Sill, geogenic reefs are larger in this study than in previous maps (30% versus 23%). In both regions, the spatial distribution of reefs differs significantly between old and new maps. For Wismar Bay, previous maps classify 3% of the seafloor as habitat type reef, whereas this study classifies 35% as geogenic reef. The use of automated classification during seafloor mapping allowed large areas to be interpreted in a few days. It also provided more information on the distribution of boulders within the geogenic reef. However, the boulder distribution maps show the negative effects of survey geometry, frequency and environmental conditions on automated boulder classification when data from different surveys are combined.

show abstract

Section: Discussionmentioning

confidence: 99%

mentioning

confidence: 99%

Distribution of boulders in coastal waters of Western Pomerania, German Baltic Sea

et al. 2023

Self Cite

View full text Add to dashboard Cite

show abstract

“…Various groups have experimented with some of the available codes to test their ability to detect boulders in various contexts (Hood et al, 2020, von Rönn et al, 2019Feldens, 2020;Feldens et al, 2021), while others continue to manually collect the required statistics (Golombke et al, 2012;Golombke et al, 2021). In view of the relevance of the abovementioned questions, one is not surprised to observe that among the two communities, a heated debate exists: can machine learning replace an experienced geomorphologist who classifies these images according to a given criteria?…”

Section: Artificial Intelligence For Boulder Detectionmentioning

confidence: 99%

Artificial intelligence in remote sensing geomorphology—a critical study

Mall,

Kloskowski,

Laserstein

2023

Front. Astron. Space Sci.

View full text Add to dashboard Cite

Planetary geomorphological maps over a wide range of spatial and temporal scales provide important information on landforms and their evolution. The process of producing a geomorphological map is extremely time-consuming and maps are often difficult to reproduce. The success of deep learning and machine learning promises to drastically reduce the cost of producing these maps and also to increase their reproducibility. However, deep learning methods strongly rely on having sufficient ground truth data to recognize the wanted surface features. In this study, we investigate the results from an artificial intelligence (AI)–based workflow to recognize lunar boulders on images taken from a lunar orbiter to produce a global lunar map showing all boulders that have left a track in the lunar regolith. We compare the findings from the AI study with the results found by a human analyst (HA) who was handed an identical database of images to identify boulders with tracks on the images. The comparison involved 181 lunar craters from all over the lunar surface. Our results show that the AI workflow used grossly underestimates the number of identified boulders on the images that were used. The AI approach found less than one fifth of all boulders identified by the HA. The purpose of this work is not to quantify the absolute sensitivities of the two approaches but to identify the cause and origin for the differences that the two approaches deliver and make recommendations as to how the machine learning approach under the given constraints can be improved. Our research makes the case that despite the increasing ease with which deep learning methods can be applied to existing data sets, a more thorough and critical assessment of the AI results is required to ensure that future network architectures can produce the reliable geomorphological maps that these methods are capable of delivering.

show abstract

“…The highest mean average precision (mAP-50) was 64 % for MBES (slope rasters), and 37 % to 43 % for two different detection models for SSS data. Feldens (2020) used deep learning super-resolution to address the limited resolution of many available side-scan sonar datasets.…”

Section: Introductionmentioning

confidence: 99%

AI-based boulder detection in sonar data – bridging the gap from experimentation to application

Hinz,

Westfeld,

Feldens

et al. 2024

IHR

View full text Add to dashboard Cite

The detection of boulders in hydroacoustic data is essential for a range of environmental, economic and marine planning applications. The manual interpretation of hydroacoustic data for object detection is a non-trivial, tedious and subjective task. Using the conventional means accessible to hydrographic professionals, it is nearly impossible to locate all boulders or rule out their presence for extended areas of interest. Although it has been shown that AI can do the job quickly and reproducibly, earlier work has not progressed beyond scientific experiments. As a result, AI software have not been routinely integrated into the workflows of institutions involved in hydrographic data acquisition and processing, or oceanographic analysis. This paper presents a workflow for fully automated boulder detection in hydroacoustic data. A graphical user interface enables training and evaluation of detection models, boulder detection model execution, and post-processing of detection results without programming. The workflow is demonstrated on data from the southern Baltic Sea. Validation results of the detection for various data inputs include a mAP-50 of 77.83 % for raster images of backscatter intensities based on side-scan sonar, a mAP-50 of 70.46 % for raster images of slope angles based on multibeam echosounder and a mAP-50 of 44.02 % for backscatter and bathymetric data given as 3D point clouds.

show abstract

Super Resolution by Deep Learning Improves Boulder Detection in Side Scan Sonar Backscatter Mosaics

Cited by 11 publications

References 35 publications

Distribution of boulders in coastal waters of Western Pomerania, German Baltic Sea

Distribution of boulders in coastal waters of Western Pomerania, German Baltic Sea

Artificial intelligence in remote sensing geomorphology—a critical study

AI-based boulder detection in sonar data – bridging the gap from experimentation to application

Contact Info

Product

Resources

About