Multi-scale,object-based image analysis for mapping geomorphic and ecological zones on coral reefs

Providing accurate maps of coral reefs where the spatial scale and labels of the mapped features correspond to map units appropriate for examining biological and geomorphic structures and processes is a major challenge for remote sensing. The objective of this work is to assess the accuracy and relevance of the process used to derive geomorphic zone and benthic community zone maps for three western Pacific coral reefs produced from multi-scale, object-based image analysis (OBIA) of high-spatial-resolution multi-spectral images, guided by field survey data. Three Quickbird-2 multi-spectral data sets from reefs in Australia, Palau and Fiji and georeferenced field photographs were used in a multi-scale segmentation and object-based image classification to map geomorphic zones and benthic community zones. A per-pixel approach was also tested for mapping benthic community zones. Validation of the maps and comparison to past approaches indicated the multi-scale OBIA process enabled field data, operator field experience and a conceptual hierarchical model of the coral reef environment to be linked to provide output maps at geomorphic zone and benthic community scales on coral reefs. The OBIA mapping accuracies were comparable with previously published work using other methods; however, the classes mapped were matched to a predetermined set of features on the reef.     

Change detection of coral reef habitat using Landsat-5 TM,Landsat 7 ETM+ and Landsat 8 OLI data in the Red Sea (Hurghada,Egypt)

Owing to continuing touristic developments in Hurghada, Egypt, several coral reef habitats have suffered major deterioration between 1987 and 2013, either by being bleached or totally lost. Such alterations in coral reef habitats have been well observed in their varying distributions using change detection analysis applied to a Landsat 5 image representing 1987, a Landsat 7 image representing 2000, and a Landsat 8 image representing 2013. Different processing techniques were carried out over the three images, including but not limited to rectification, masking, water column correction, classification, and change detection statistics. The supervised classifications performed over the three scenes show five significant marine-related classes, namely coral, sand subtidal, sand intertidal, macro-algae, and seagrass, in different degrees of abundance. The change detection statistics obtained from the classified scenes of 1987 and 2000 reveal a significant increase in the macro-algae and seagrass classes (93 and 47%, respectively). However, major decreases of 41, 40, and 37% are observed in the sand intertidal, coral, and sand subtidal classes, respectively. On the other hand, the change detection statistics obtained from the classified scenes of 2000 and 2013 revealed increases in sand subtidal and macro-algae classes by 14 and 19%, respectively, while major decreases of 49%, 46% and 74% are observed in the sand intertidal, coral, and seagrass classes, respectively.     

基于递推规范变量残差和核主元分析的微小故障检测

微小故障由于故障征兆不明显从而很难在故障发生早期对其进行检测. 针对该问题, 本文提出了一种基于递推规范变量残差和核主元分析(RCVD–KPCA)的微小故障检测方法. 首先构造规范变量残差, 从中提取数据的线性特征. 利用指数加权滑动平均法对规范变量残差进行递推滤波处理, 提高规范变量残差对微小故障的敏感程度;然后使用KPCA提取规范变量残差中的非线性主成分作为非线性特征, 根据提取的特征提出了两个新的故障检测统计量; 此外, 利用核密度估计确定故障检测统计量的控制限. 由于同时提取了过程数据的线性和非线性特征, 有效地提高了非线性动态过程中微小故障的可检测性. 以闭环连续搅拌釜式反应器过程为例进行了仿真分析, 仿真结果表明本文所提方法具有较好的故障检测性能.     

Sparse multinomial kernel discriminant analysis (sMKDA)

Dimensionality reduction via canonical variate analysis (CVA) is important for pattern recognition and has been extended variously to permit more flexibility, e.g. by “kernelizing” the formulation. This can lead to over-fitting, usually ameliorated by regularization. Here, a method for sparse, multinomial kernel discriminant analysis (sMKDA) is proposed, using a sparse basis to control complexity. It is based on the connection between CVA and least-squares, and uses forward selection via orthogonal least-squares to approximate a basis, generalizing a similar approach for binomial problems. Classification can be performed directly via minimum Mahalanobis distance in the canonical variates. sMKDA achieves state-of-the-art performance in terms of accuracy and sparseness on 11 benchmark datasets.     

基于核规范变量分析的非线性故障诊断方法

提出一种基于核规范变量分析（KCVA）的非线性过程故障诊断方法．该方法使用核函数完成非线性空间到高维线性空间的映射,避免了高维空间中的数据处理和非线性映射函数的使用．在线性空间中使用规范变量分析（CVA）来辨识状态空闻模型,从数据中提取状态信息．3个监测量（Tr^2,Ts^2,Q）用来进行故障检测,同时使用贡献图分离故障变量,并判断故障原因．在CSTR系统上的仿真结果表明,KCVA方法比主元分析法（PCA）和CVA方法能更灵敏地检测到故障的发生,更有效地监控过程变化．     

Airborne lidar sensing of massive stony coral colonies on patch reefs in the northern Florida reef tract

In this study we examined the ability of the NASA Experimental Advanced Airborne Research Lidar (EAARL) to discriminate cluster zones of massive stony coral colonies on northern Florida reef tract (NFRT) patch reefs based on their topographic complexity (rugosity). Spatially dense EAARL laser submarine topographic soundings acquired in August 2002 were used to create a 1-m resolution digital rugosity map for adjacent NFRT study areas characterized by patch reefs (Region A) and diverse substratums (Region B). In both regions, sites with lidar-sensed rugosities above 1.2 were imaged by an along-track underwater videography system that incorporated the acquisition of instantaneous GPS positions. Subsequent manual interpretation of videotape segments was performed to identify substratum types that caused elevated lidar-sensed rugosity. Our study determined that massive coral colony formation, modified by subsequent physical and biological processes that breakdown patch reef framework, was the primary source of topographic complexity sensed by the EAARL in the NFRT. Sites recognized by lidar scanning to be topographically complex preferentially occurred around the margins of patch reefs, constituted a minor fraction of the reef system, and usually reflected the presence of massive coral colonies in cluster zones, or their derivatives created by mortality, bioerosion, and physical breakdown.     

Detection limits of coral reef bleaching by satellite remote sensing: Simulation and data analysis

Monitoring of coral reef bleaching has hitherto been based on regional-scale, in situ data. Larger-scale trends, however, must be determined using satellite-based observations. Using both a radiative transfer simulation and an analysis of multitemporal Landsat TM images, the ability of satellite remote sensing to detect and monitor coral reef bleaching is examined. The radiative transfer simulation indicates that the blue and green bands of Landsat TM can detect bleaching if at least 23% of the coral surface in a pixel has been bleached, assuming a Landsat TM pixel with a resolution of 30×30 m on shallow (less than 3 m deep) reef flats at Ishigaki Island, Japan. Assuming an area with an initial coral coverage of 100% and in which all corals became completely bleached, the bleaching could be detected at a depth of up to 17 m. The difference in reflectance of shallow sand and corals is compared by examining multitemporal Landsat TM images at Ishigaki Island, after normalizing for variations in atmospheric conditions, incident light, water depth, and the sensor's reaction to the radiance received. After the normalization, a severe bleaching event when 25-55% of coral coverage was bleached was detected, but a slight bleaching event when 15% of coral coverage was bleached was not detected. The simulation and data analysis agreed well with each other, and identified reliable limits for satellite remote sensing for detecting coral reef bleaching. Sensitivity analysis on solar zenith angle, aerosol (visibility) and water quality (Chl a concentration) quantified the effect of these factors on bleaching detection, and thus served as general guidelines for detecting coral reef bleaching. Spatial misregistration resulted in a high degree of uncertainty in the detection of changes at the edges of coral patches mainly because of the low (∼30 m) spatial resolution of Landsat TM, indicating that detection of coral reef bleaching by Landsat TM is limited to extremely severe cases on a large homogeneous coral patch and shallow water depths. Satellite remote sensing of coral reef bleaching should be encouraged, however, because the development and deployment of advanced satellite sensors with high spatial resolution continue to progress.     

Quantification of two decades of shallow-water coral reef habitat decline in the Florida Keys National Marine Sanctuary using Landsat data (1984–2002)

The loss of coral reef habitats has been witnessed at a global scale including in the Florida Keys and the Caribbean. In addition to field surveys that can be spatially limited, remote sensing can provide a synoptic view of the changes occurring on coral reef habitats. Here, we utilize an 18-year time series of Landsat 5/TM and 7/ETM+ images to assess changes in eight coral reef sites in the Florida Keys National Marine Sanctuary, namely Carysfort Reef, Grecian Rocks, Molasses Reef, Conch Reef, Sombrero Reef, Looe Key Reef, Western Sambo and Sand Key Reef. Twenty-eight Landsat images (1984–2002) were used, with imagery gathered every 2 years during spring, and every 6 years during fall. The image dataset was georectified, calibrated to remote sensing reflectance and corrected for atmospheric and water-column effects. A Mahalanobis distance classification was trained for four habitat classes (‘coral’, ‘sand’, ‘bare hardbottom’ and ‘covered hardbottom’) using in situ ground-truthing data collected in 2003–2004 and using the spectral statistics from a 2002 image. The red band was considered useful only for benthic habitats in depths less than 6 m. Overall mean coral habitat loss for all sites classified by Landsat was 61% (3.4%/year), from a percentage habitat cover of 19% (1984) down to 7.6% (2002). The classification results for the eight different sites were critically reviewed. A detailed pixel by pixel examination of the spatial patterns across time suggests that the results range from ecologically plausible to unreliable due to spatial inconsistencies and/or improbable ecological successions. In situ monitoring data acquired by the Coral Reef Evaluation and Monitoring Project (CREMP) for the eight reef sites between 1996 and 2002 showed a loss in coral cover of 52% (8.7%/year), whereas the Landsat-derived coral habitat areas decreased by 37% (6.2%/year). A direct trend comparison between the entire CREMP percent coral cover data set (1996–2004) and the entire Landsat-derived coral habitat areas showed no significant difference between the two time series (ANCOVA; F-test, p = 0.303, n = 32), despite the different scales of measurements.     

一种基于DIOCVA的过程监控方法

传统基于典型变量分析的过程监控方法无法判断故障是否影响产 品质量.为此,本文提出一种基于动态输入输出典型变量分析(Dynamic input-output canonical variate analysis, DIOCVA)的过程监控方法.该方法利用典型变量分析提取数据之间的相关性,并进一步考虑方差信息和时序相关性, 将过程数据和质量数据映射到5个子空间:输入输出相关子空间,不相关输入主元子空间, 不相关输入残差子空间,不相关输出主元子空间和不相关输出残差 子空间.所提方法能够精细区分影响质量的过程故障和不影响质量的过程故障.以Tennessee Eastman过程为例对所提方法的有效性进行了验证.     

The semi-automated classification of acoustic imagery for characterizing coral reef ecosystems

Coral reef habitat maps describe the spatial distribution and abundance of tropical marine resources, making them essential for ecosystem-based approaches to planning and management. Typically, these habitat maps have been created from optical and acoustic remotely sensed imagery using manual, pixel- and object-based classification methods. However, past studies have shown that none of these classification methods alone are optimal for characterizing coral reef habitats for multiple management applications because the maps they produce (1) are not synoptic, (2) are time consuming to develop, (3) have low thematic resolutions (i.e. number of classes), or (4) have low overall thematic accuracies. To address these deficiencies, a novel, semi-automated object- and pixel-based technique was applied to multibeam echo sounder imagery to determine its utility for characterizing coral reef ecosystems. This study is not a direct comparison of these different methods but rather, a first attempt at applying a new classification technique to acoustic imagery. This technique used a combination of principal components analysis, edge-based segmentation, and Quick, Unbiased, and Efficient Statistical Trees (QUEST) to successfully partition the acoustic imagery into 35 distinct combinations of (1) major and (2) detailed geomorphological structure, (3) major and (4) detailed biological cover, and (5) live coral cover types. Thematic accuracies for these classes (corrected for proportional bias) were as follows: (1) 95.7%, (2) 88.7%, (3) 95.0%, (4) 74.0%, and (5) 88.3%, respectively. Approximately half of the habitat polygons were manually edited (hence the name ‘semi-automated’) due to a combination of mis-classifications by QUEST and noise in the acoustic data. While this method did not generate a map that was entirely reproducible, it does show promise for increasing the amount of automation with which thematically accurate benthic habitat maps can be generated from acoustic imagery.     

Mapping shallow nearshore benthic features in a Caribbean marine-protected area: assessing the efficacy of using different data types (hydroacoustic versus satellite images) and classification techniques

Various benthic mapping methods exist but financing and technical capacity limit the choice of technology available to developing states to aid with natural resource management. Therefore, we assessed the efficacy of using a single-beam echosounder (SBES), satellite images (GeoEye-1 and WorldView-2) and different image (pixel-based Maximum Likelihood Classifier (MLC), and an object-based image analysis (OBIA)) and hydroacoustic classification and interpolation techniques, to map nearshore benthic features at the Bluefields Bay marine protected area in western Jamaica (13.82 km² in size). A map with three benthic classes (submerged aquatic vegetation (SAV), bare substrate, and coral reef) produced from a radiometrically corrected, deglinted and water column-corrected WorldView-2 image had a marginally higher accuracy (3%) than that of a map classified from a similarly corrected GeoEye-1 image. However, only one of the two extra WorldView-2 image bands (coastal) was used because the yellow band was completely attenuated at depths ≥3.7 m. The coral reef class was completely misclassified by the MLC and had to be contextually edited. The contextually edited MLC map had a higher overall accuracy (OA) than the OBIA map (86.7% versus 80.4%) and maps that were not contextually edited. But, the OBIA map had a higher OA than a MLC map without edits. Maps produced from the images also had a higher accuracy than the SAV map created from the acoustic data (OAs >80% and kappa >0.67 versus 76.6% and kappa = 0.32). SAV classification was comparable among the classified SBES SAV data points and all the final maps. The total area classified as SAV was marginally larger for satellite maps; however, the total area classified as bare substrate using the images was twice as large. A substrate map with three classes (silt, sand, and coral/hard bottom) produced from the SBES data using a random forest classifier and a Markov chain interpolator had a higher accuracy than a substrate map produced using a fractal dimension classifier and an indicator krig (the default choice) (72.4% versus 53.5%). The coral reef class from the SBES, OBIA, and contextually edited maps had comparable accuracies, but covered a much smaller area in the SBES maps because data points were lost during the interpolation process. The use of images was limited by turbidity levels and cloud cover and it yielded lower benthic detail. Despite these limitations, satellite image classification was the most efficacious method. If greater benthic detail is required, the SBES is more suitable or more effort is required during image classification. Also, the SBES can be operated in areas with turbid waters and greater depths. However, it could not be used in very shallow areas. Also, processing and interpolation of data points can result in a loss of resolution and introduces spatial uncertainty.     

A new image classification approach for mapping coral density in State of Kuwait using high spatial resolution satellite images

Remote sensing technology can be a valuable tool for mapping coral reef ecosystems. However, the resolution capabilities of remote sensors, the diversity and complexity of coral reef ecosystems, and the low reflectivity of marine environments increase the difficulties in identifying and classifying their features. This research study explores the capability of high spatial resolution (WorldView-2 (WV-2) and Pleiades-1B) and low spatial resolution (Land Remote-Sensing Satellite (Landsat 8)) multispectral (MS) satellite sensors in quantitatively mapping coral density. The Kubbar coral reef ecosystem, located in Kuwait’s southern waters, was selected as the research site. The MS imagery of WV-2, Pleiades-1B and Landsat 8 were, after geometric and radiometric assessment and corrections, subjected to new image classification approach using a Multiple Linear Regression (MLR) analysis. The new approach of MLR coral density analysis used the dependent variable of coral density percentage from ground truth and independent variables of spectral reflectance from selected imagery, depth (as estimated from a surface derived from bathymetric charts) and distance to land or reef unit centre. Accuracy assessment using independent ground truth was performed for the selected approach and satellite sensors to determine the quality of the information derived from image classification processes. The results showed that coral density maps developed using the MLR coral density model proved to have some level of reliability (radiometrically corrected WV-2 image (the coefficient determination denoted as R-squared (R²) = 0.5, Root-Mean-Square Error (RMSE) = 10) and radiometrically corrected Pleiades-1B image (R² = 0.8, RMSE = 10)). This study suggested using high spectral resolution data and including additional factors (variables) (e.g. water turbidity, temperature and salinity) could contribute to improving the accuracy of coral density maps produced by application of the MLR model; however, all of these would add cost and effort to the mapping process. The outcomes of this research study provide coral reef ecosystem researchers, managers, and decision makers a tool to determine and map coral reef density in more detail than in the past. It will help quantify coral density at particular points in time leading to estimates of change, and allow coral reef ecologists to identify the current coral reef habitat health status, distribution and extent.     

Time-lapse side-scan sonar imaging of bleached coral reefs: A case study from the Seychelles

High-resolution (675 kHz) side-scan sonar surveys were collected six and thirty months after a major coral bleaching event in the Seychelles. The surveyed areas contain four different reef morphologies, the distribution of which depends on water depth and distance from the shoreline. These four reef types also have different coral densities and colony morphologies (branching or mixed massive/branching). Textural analysis, based on first order statistics, unsupervised cluster analysis and Mann-Whitney U-tests, showed a correlation between recorded backscatter response and reef type that is attributed to the link between community composition and rugosity on a millimetre to tens of metre scale. Branching coral colonies are found to be relatively good scatterers of acoustic energy, and result in a broad intermediate to high intensity response. Massive coral colonies and reefs with a hard carbonate pavement are found to be principally reflectors of the acoustic energy, resulting in a narrow low intensity response. Comparison of the two surveys separated in time by two years, showed seabed texture to change most significantly over the reef areas that contained the highest coral abundances and mortality rates. In particular, the disintegration of dense branching colonies that suffered almost 100% mortality during the bleaching event resulted in a characteristic loss of intermediate to high backscatter intensity. The work demonstrates the contribution that side-scan sonar could make in the assessment of loss of rugosity following a bleaching episode, which has important implications to both the recovery of the reef itself and the abundance and distribution of associated reef organisms.     

Mapping coral reefs at reef to reef-system scales, 10s–1000s km2, using object-based image analysis

Coral reef maps at various spatial scales and extents are needed for mapping, monitoring, modelling, and management of these environments. High spatial resolution satellite imagery, pixel <10 m, integrated with field survey data and processed with various mapping approaches, can provide these maps. These approaches have been accurately applied to single reefs (10–100 km²), covering one high spatial resolution scene from which a single thematic layer (e.g. benthic community) is mapped. This article demonstrates how a hierarchical mapping approach can be applied to coral reefs from individual reef to reef-system scales (10–1000 km²) using object-based image classification of high spatial resolution images guided by ecological and geomorphological principles. The approach is demonstrated for three individual reefs (10–35 km²) in Australia, Fiji, and Palau; and for three complex reef systems (300–600 km²) one in the Solomon Islands and two in Fiji. Archived high spatial resolution images were pre-processed and mosaics were created for the reef systems. Georeferenced benthic photo transect surveys were used to acquire cover information. Field and image data were integrated using an object-based image analysis approach that resulted in a hierarchically structured classification. Objects were assigned class labels based on the dominant benthic cover type, or location-relevant ecological and geomorphological principles, or a combination thereof. This generated a hierarchical sequence of reef maps with an increasing complexity in benthic thematic information that included: ‘reef’, ‘reef type’, ‘geomorphic zone’, and ‘benthic community’. The overall accuracy of the ‘geomorphic zone’ classification for each of the six study sites was 76–82% using 6–10 mapping categories. For ‘benthic community’ classification, the overall accuracy was 52–75% with individual reefs having 14–17 categories and reef systems 20–30 categories. We show that an object-based classification of high spatial resolution imagery, guided by field data and ecological and geomorphological principles, can produce consistent, accurate benthic maps at four hierarchical spatial scales for coral reefs of various sizes and complexities.     

SPOT5多光谱图像对南沙珊瑚礁信息提取方法的探讨

珊瑚礁在海洋生态系统中具有重要的作用,国外对珊瑚礁遥感信息提取研究较早,但国内起步较晚,目前仍以目视判读为主。使用SPOT5 10 m多光谱数据对我国南沙群岛中弹丸礁和光星礁进行了分类实验,并以QuickBird融合影像判读结果及岛礁调查资料为参考,探讨了SPOT5对珊瑚礁水下信息的辨识能力。首先利用阈值分割方法将水下和水上信息分离;对于水下信息,尝试进行Deglint纠正并引入衍生波段,用4种波段组合方案分别进行最大似然法分类后进行结果比较。结果表明,Deglint纠正能够有效消除水面噪声,从而提高分类精度;近红外波段对水下信息提取有辅助作用,可帮助纠正一些由水深导致的错分;衍生波段替代近红外波段分类效果略差。总之,SPOT5 10 m多光谱数据经过图像纠正后能够有效提取珊瑚岛礁水下成分信息,其分类精度可达80%以上。     

Characterisation of ecotones using membership degrees computed with a fuzzy classifier

Ecotones zones lie between homogeneous ecological systems. They are characterised on images by heterogeneous pixels in a specific neighbourhood. Fuzzy classifiers output membership degrees that better represent the heterogeneity within a pixel and can be further processed within the context of a local neighbourhood. This Letter formalises these notations. Two coral reef systems examples are presented. They illustrate the use of possibility measurement to characterise ecotones, and the use of information on well-known ecotones to increase the accuracy of image classification.     

回转窑烧成系统故障诊断方法研究

针对水泥烧成系统过程变量繁多、变量间静态关系耦合强等特点,采用因子分析方法建立静态过程监控模型。针对系统时序相关问题,结合经典动态主元分析DPCA方法和典型变量分析CVA方法,提出典型变量动态主元分析CVDPCA过程监控方法,有效解决了DPCA方法扩展后的数据矩阵维度大等不足之处。将算法用于水泥烧成系统故障检测,结果表明该算法能准确识别故障和更早检测到微小渐变故障。将CVA和DPCA算法相结合,可以同时监控动态过程和静态关系,且不需要大量的故障数据建立故障模型池,具有一定研究价值。     

Using bathymetric lidar to define nearshore benthic habitat complexity: Implications for management of reef fish assemblages in Hawaii

Habitat complexity plays a major role in determining the distribution and structure of fish assemblages in the aquatic environment. These locations are critical for ecosystem function and have significant implications for conservation and management. In this study, we evaluated the utility of remotely sensed lidar (light detection and ranging) data for deriving substrate rugosity (a measure of habitat complexity) on a coral reef in Hawaii. We also assessed the potential application of lidar data for examining the relationship between habitat complexity and Hawaiian reef fish assemblage characteristics. Lidar-derived rugosity (4 m grid size) was found to be highly correlated with in-situ rugosity and was concluded to be a viable method for measuring rugosity in analogous coral reef environments. We established that lidar-derived rugosity was a good predictor of fish biomass and demonstrated a strong relationship with several fish assemblage metrics in hard bottom habitat at multiple spatial resolutions. This research demonstrates (i) the efficacy of lidar data to provide substrate rugosity measures at scales commensurate with the resources and their environment (ii) the applicability of lidar-derived rugosity for examining fish–habitat relationships on a coral reef in Hawaii and (iii) the potential of lidar to provide information about the seascape structure that can ultimately be used to prioritize areas for conservation and management.     

Lithologic mapping in the Mountain Pass, California area using Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) data

Evaluation of an Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) image of the Mountain Pass, California area indicates that several important lithologic groups can be mapped in areas with good exposure by using spectral-matching techniques. The three visible and six near-infrared bands, which have 15-m and 30-m resolution, respectively, were calibrated by using in situ measurements of spectral reflectance. Calcitic rocks were distinguished from dolomitic rocks by using matched-filter processing in which image spectra were used as references for selected spectral categories. Skarn deposits and associated bright coarse marble were mapped in contact metamorphic zones related to intrusion of Mesozoic and Tertiary granodioritic rocks. Fe-muscovite, which is common in these intrusive rocks, was distinguished from Al-muscovite present in granitic gneisses and Mesozoic granite.Quartzose rocks were readily discriminated, and carbonate rocks were mapped as a single broad unit through analysis of the 90-m resolution, five-band surface emissivity data, which is produced as a standard product at the EROS Data Center. Three additional classes resulting from spectral-angle mapper processing ranged from (1) a broad granitic rock class (2) to predominately granodioritic rocks and (3) a more mafic class consisting mainly of mafic gneiss, amphibolite and variable mixtures of carbonate rocks and silicate rocks.     

CoralSeg: Learning coral segmentation from sparse annotations

Robotic advances and developments in sensors and acquisition systems facilitate the collection of survey data in remote and challenging scenarios. Semantic segmentation, which attempts to provide per‐pixel semantic labels, is an essential task when processing such data. Recent advances in deep learning approaches have boosted this task's performance. Unfortunately, these methods need large amounts of labeled data, which is usually a challenge in many domains. In many environmental monitoring instances, such as the coral reef example studied here, data labeling demands expert knowledge and is costly. Therefore, many data sets often present scarce and sparse image annotations or remain untouched in image libraries. This study proposes and validates an effective approach for learning semantic segmentation models from sparsely labeled data. Based on augmenting sparse annotations with the proposed adaptive superpixel segmentation propagation, we obtain similar results as if training with dense annotations, significantly reducing the labeling effort. We perform an in‐depth analysis of our labeling augmentation method as well as of different neural network architectures and loss functions for semantic segmentation. We demonstrate the effectiveness of our approach on publicly available data sets of different real domains, with the emphasis on underwater scenarios—specifically, coral reef semantic segmentation. We release new labeled data as well as an encoder trained on half a million coral reef images, which is shown to facilitate the generalization to new coral scenarios.