摘要:Since the 1960s, remote sensing science and technology has emerged as a competitive high-tech field, with major countries striving to advance their capabilities. It has become a fundamental tool for human research in the earth system science and the comprehensive application of aerospace information across multiple domains. Recently, two significant developments warrant attention: First, in 2022, the Ministry of Education of China officially recognized Remote Sensing Science and Technology as a first-level interdisciplinary discipline within the graduate education framework, thereby strengthening foundational research in remote sensing and broadening its application areas. Second, the rise of artificial intelligence technologies, particularly deep learning, has ushered in a new paradigm for data-driven analysis and application of remote sensing data. While remote sensing fundamentally belongs to the domain of electromagnetic radiation physics, the associated physical models have been indispensable for the development of quantitative remote sensing. Nevertheless, the data-driven deep learning paradigm has introduced transformative ideas and methodologies to the field. Moving forward, the synergy between physical models and artificial intelligence will undoubtedly shape the future trajectory of remote sensing research and applications. In this context, a deeper exploration of the core concepts and fundamental issues in remote sensing science is crucial for achieving significant technological breakthroughs and scientific discoveries within this discipline.This article begins by examining the physical origins of remote sensing science, focusing on the interaction between ground objects and electromagnetic waves, which produces spectral radiation images under specific conditions. It explores the characteristics of various remote sensing methods across the electromagnetic spectrum, including solar reflected radiation in the visible to shortwave infrared remote sensing, daylight-induced chlorophyll fluorescence (SIF) remote sensing, laser remote sensing, both medium and longwave infrared remote sensing, and microwave remote sensing. The fundamental theoretical issues in remote sensing science are categorized into three primary characteristics: radiative, spectral, and temporal characteristics, along with five major effects: scale, atmospheric, angular, adjacent, and transfer effects. The former pertains to the intrinsic physical and chemical properties of ground objects within the electromagnetic spectrum, while the latter relates to factors such as imaging scale, atmospheric conditions, observation angle, and background environment. This discussion includes the expression and variation patterns of remote sensing features of land objects formed under diverse observation modes and conditions.The radiative characteristics reflect overall difference in term of radiation across different electromagnetic bands for various land covers, closely tied to geophysical and chemical properties. The spectral characteristics of land cover manifest as variations in the intensity of reflected and emitted signals with wavelength, highlighting significant differences in absorption, reflection, and emission behaviors among different materials, known as spectral characteristics. Temporal characteristics pertain to the systematic changes in spectral reflection or emission over time, aiding in remote sensing identification or feature inversion of land cover. The scale effect refers to the changes in remote sensing observation characteristics due to variations in pixel area size, influenced by spatial resolution or point scanning density (e.g., laser scanning spot density). The atmospheric effect describes how electromagnetic waves are impacted by the absorption, scattering, and emission from atmospheric particles during remote sensing imaging, leading to radiation distortion in image data. The angular effect highlights the directional nature of the interaction between land cover and electromagnetic waves, resulting in significant anisotropic characteristics and variations in radiation values based on the angles of incident radiation, remote sensing observation, and electromagnetic wave wavelength. The adjacent effect refers to the influence of spatial structure heterogeneity among land features, which can create cross-radiation contributions from non-target pixels to target pixels, dependent on spatial distribution and remote sensing observation mode. Finally, the transfer effect encompasses the changes in imaging quality after the electromagnetic signal of the ground objects entering the remote sensing system, including the processes such as photoelectric conversion, signal transmission, and digital recording.The review and discussion presented in this article on the fundamental issues of remote sensing science aim to deepen theoretical research in the field, particularly in the context of artificial intelligence. This exploration is intended to foster innovative methods in remote sensing technology and applications, promote the collaborative evolution of AI for Science and Science for AI in remote sensing, and encourage profound cross-disciplinary integration between remote sensing and other fields.
摘要:Land Use/Cover Change (LUCC) is a direct driver of the carbon balance in terrestrial ecosystems, and its impact on global warming is second only to fossil fuel and industrial emissions. The forest ecosystem is the largest carbon pool in terrestrial ecosystems and has an important role to play in addressing global climate change and achieving carbon neutrality targets. However, the limited availability of LUCC data has led to a significant underestimation of its impact on carbon emissions, and the lack of spatiotemporal LUCC data under future climate scenarios also introduced considerable uncertainty in exploring the response of the forest carbon cycle to LUCC. How to simulate LUCC and analyze the impact of LUCC on the carbon cycle of forest ecosystems have become key research focuses both domestically and internationally. This study systematically reviewed the progress of research on past LUCC extraction, spatiotemporal LUCC simulations, forest carbon balance estimation methods, and the impact of LUCC on the forest carbon cycle. The advantages, applicability, and existing challenges of different LUCC simulation models and forest carbon balance estimation models were listed and analyzed.First, this review summarized historical LUCC extraction methods and highlights the urgent need to integrate deep learning techniques to improve the accuracy of LUCC change detection, thereby providing more reliable data for future LUCC simulations. Second, this review generalized the mainstream models for future LUCC spatiotemporal simulations. It emphasized the importance of coupling deep learning algorithms with the SD model while integrating meteorological and socioeconomic driving factors. This approach would more comprehensively account for feedback mechanisms between natural and anthropogenic factors, thereby enhancing the accuracy and applicability of simulations. Subsequently, this review organized the commonly used methods in current forest carbon cycle modeling and highlights recent developments in the field. It noted that carbon balance estimation has increasingly shifted toward remote sensing-driven process models, gradually replacing the traditional approach of combining remote sensing with parameterized models. This shift allows for the incorporation of more comprehensive and detailed ecosystem processes based on previous methodologies. Finally, this review examined the research on the effects of LUCC on carbon cycles. It pointed out that most current studies loosely couple LUCC simulation results with process-based ecosystem models, neglecting the dynamic effects of LUCC on key physiological and biochemical parameters of the forest carbon cycle, such as LAI and chlorophyll. Future research should leverage remote sensing and other technologies to strengthen simulations of the spatiotemporal LAI distribution. This could reduce uncertainties in carbon sink estimation and improve the precision of assessments of carbon sink potential driven by LUCC.Despite great progress in recent years, future research should focus on optimizing the reconstruction of historical LUCC, spatial-temporal simulations, and the parameterization coupling of carbon cycles. This would provide a more comprehensive understanding of the response mechanisms of forest carbon cycles to LUCC. In summary, leveraging remote sensing data as a basis to simulate LUCC and drive process-based models to achieve accurate spatial and temporal simulation of forest ecosystem carbon cycle remains critical research directions in future LUCC and carbon cycle related to it.
摘要:Forest age is a critical parameter determining forest carbon sequestration capacity and its temporal trends. Quantifying spatiotemporal variations in forest age is essential for predicting forest ecosystem carbon dynamics. While traditional forest age assessments were limited to forest plots, the development of remote sensing technology has expanded the estimation from plots to regional and global scales. Research related to forest age has gaining increasing attention across fields of forestry, ecology, and geography, etc. This article aims to review the process in forest age estimation by summarizing the main methods and their applications from related literatures and datasets published since the year of 2000.Remote sensing-based approaches fall into three main categories: 1. regression from image spectral and texture features, 2. time series change detection, and 3. tree height or biomass growth equation modeling. 1.The spectral image regression method is straightforward but often has limited regression accuracy due to the saturation effect in the spectral image information-forest age relationship. 2. The time series change detection method can achieve high accuracy but only applicable to forests with continuous remote sensing observations. 3. The tree height or biomass growth equations based strategy can broaden the limits of forest age estimation, but its estimation accuracy is sensitive to the selections of model equation and input parameters. Consequently, integrating multisource datasets and combining multiple modeling approaches have become the predominant strategy for forest age estimation. This strategy has been successfully implemented in high-resolution forest age mapping at national scales across China and Canada.The advancement of remote sensing technology has substantially improved the efficiency and accuracy of forest age estimation, extending its applicability from individual plots to regional and global scales. Large-scale forest age data have great potential for applications in forest carbon cycle modeling, biodiversity assessment, and forest management. Future research should focus on improving and updating forest measurement datasets, fully leveraging multisource and multispatial-temporal remote sensing information, and enhancing the transferability and generality of estimation models.
摘要:The forest canopy structure plays a crucial role in regulating the exchange of substances and energy between plants and the atmosphere, thereby influencing regional microclimate and ecosystem functionality. Accurate characterization of vegetation canopy structure is of significant importance for forest ecosystem research, such carbon storage estimation, carbon cycle simulation etc. Canopy structural complexity, also known as canopy structural biodiversity, which describes the spatial distribution of branches and leaves within the canopy, has emerged as a key attribute in forest ecosystems and has found wide application in related research. For example, carbon cycle, mechanisms of community composition, sustainable forest management, wildlife conservation, forest disturbance monitoring and restoration, forest microclimate research and so on.Traditional ground-based survey methods have limitations as they only provide partial information through statistical values, which primarily involve plot-based surveys using tools such as diameter tapes, clinometers, and angle gauges to obtain individual tree information such as tree position, diameter at breast height, tree height, and crown width. The heterogeneity of these measured tree attributes and their distribution, such as diameter at breast height and tree height, or combinations of tree height, diameter at breast height, and tree density, are used to quantify canopy structure complexity, including the standard deviation, coefficient of variation, and Gini coefficient of survey attributes. However, these indices may not fully represent canopy structural complexity.The rapid development of lidar technology has enabled the rapid acquisition of three-dimensional structural information for entire forests, offering new opportunities for comprehensive and accurate characterization of canopy structure complexity. In addition to the indicators used in traditional ground-based survey methods, existing quantitative indices for canopy structure complexity based on lidar data can generally be divided into three categories: horizontal distribution indices, vertical distribution indices, and integrated distribution indices. Horizontal distribution indices primarily quantify the horizontal spatial distribution of canopy elements, without considering their vertical distribution, such as canopy cover, canopy closure, and leaf area index. Vertical distribution indices mainly describe the heterogeneity of canopy element distribution in the vertical direction while neglecting their horizontal distribution including canopy effective layers and leaf height diversity and so on. Integrated distribution indices consider both the horizontal and vertical distribution heterogeneity of canopy structure, thereby overcoming the limitations of solely considering a single direction in horizontal or vertical distribution indices, for example canopy fractal dimension, canopy roughness, and canopy entropy.Finally, we summarize the current applications of canopy structure complexity in regulating forest ecosystem functions, including light resource utilization, precipitation interception, microclimate modulation, productivity, and ecosystem stability. Additionally, there are key issues and directions that require emphasis in forest ecosystem research related to canopy structure complexity. These include investigating the cross-platform generality of lidar-based indicators, addressing scale issues, and establishing long-term monitoring methods. While the concept of forest canopy structure complexity is relatively new and has limited application in China, we anticipate that advancements in characterization methods and a deeper understanding of its implications will be facilitated by the increasing availability of long-term, multi-source remote sensing data and the utilization of various deep learning methods.
摘要:In recent years, quantitative studies on CO2 fertilization effects(β) have gradually become a hotspot in global carbon cycle research. The rising atmospheric CO2 concentration largely affects the changes in gross primary productivity (GPP), which in turn may have an impact on net ecosystem productivity (NEP). Therefore, in the context of global climate change, accurately quantifying the response of GPP to the rise in atmospheric CO2 () and the response of NEP to the rise in atmospheric CO2 () and exploring their spatial and temporal trends are of great significance to fully understand the mechanism of CO2 fertilization effects. Method In this study, the global forest was taken as the research object, and the EC-LUE GPP dataset based on an eddy covariance-light use efficiency model and the Jena CarboScope NEP dataset based on an atmospheric transport model were employed to quantify the spatial and temporal trends of the effects of GPP and NEP on CO2 fertilization effects during 1982-2015 and to investigate the effects of and on CO2 fertilization effects in different forest stand types and climatic zones. This study used the random forest regression algorithm and established two CO2 concentration scenario models for GPP and NEP datasets to reconstruct the changes in GPP and NEP in the global forest-growing seasons during the period of 1982-2015. The differences in GPP and NEP between the two scenarios were calculated and combined with the differences in CO2 concentration to quantitatively estimate and . The spatial and temporal trends of and over the period 1982-2015 were analyzed by setting a moving window of 15 years and using the Mann–Kendall trend test on a pixel-by-pixel basis. Result The mean values of and in global forests from 1982 to 2015 were (18.3±14.9)%/100 ppm and (7.4±4.0)%/100 ppm, respectively, and showed a significant decreasing trend with annual average rates of 0.14%/100 ppm/yr and 0.11%/100 ppm/yr, respectively. The increase in atmospheric CO2 concentration exerted a significant contributing effect on GPP (80.1%) and NEP (81.6%) in most regions of global forests; however, most of them showed a significant decreasing trend in (52.4%) and (59.2%). As the latitude of the climatic zone increased, the rate of decline in indicated a gradually decreasing trend, whereas showed a gradually increasing trend. Evergreen broadleaf forests demonstrated the slowest decline in , whereas deciduous broadleaf forests exhibited the fastest decline in . The decline rate of in broadleaf forests was greater than that in coniferous forests. Conclusion The response of GPP and NEP in global forests to the increasing atmospheric CO2 concentration shows a declining trend. Neglecting this decline would impact the future estimation of global forest carbon sequestration potential and the achievement of carbon neutrality goals. Therefore, in the process of achieving carbon neutrality, timely adjustments should be made to optimize forest age structure, improve forest site conditions, and implement forest management practices aimed at enhancing forest carbon sequestration capacity.
摘要:Cold-temperate forests, recognized as the most extensive terrestrial ecosystems, cover vast areas around the globe and hold important ecological and social values. Accurate mapping of forest type and tree species cover fraction in these forests across space and time is crucial for quantifying ecosystem services and formulating effective forest management policies to ensure their sustainable conservation. However, despite the increasing development of remote sensing technologies, studies exploring the feasibility of inverting forest type and tree species cover fraction using medium-resolution multispectral satellite-based data, such as Landsat, in China’s cold-temperate forests are limited. This limitation is primarily attributed to the scarcity of reference data and the restricted spectral information available in multispectral images. Moreover, the quantitative impact of the temporal frequency of data acquisitions (e.g., single-date, multidate) on mapping forest type and tree species cover fraction remains largely unexplored. The timing and frequency of satellite data acquisition can significantly influence the detection and characterization of dynamic changes in forests, which in turn affects the accuracy of mapping forest attributes. To address these gaps, our study aims to map the forest type and tree species cover fraction in Mengjiagang Forest, Heilongjiang Province by employing synthetically mixed data and a random forest regression model. We extend our analysis to three decades (from 1986 to 2020) of Landsat data, mapping the cover fractions of broadleaf and needleleaf forests in Mengjiagang Forest by using an optimal broadleaf and needleleaf random forest regression model. The results of our study reveal the following key findings: (1) For forest type cover fraction inversion, the random forest regression model based on the growing season median index (including spectral bands, NDTI, and TCT indices) is the optimal model (achieving =0.76 for broadleaf and =0.71 for needleleaf). (2) For tree species cover fraction inversion, the random forest regression model based on multidate spectral features (including the spectral bands, NDTI, and TCT indices of growth and leaf off seasons) is the optimal model (achieving =0.40 for Larch, =0.23 for Korean pine, and =0.61 for Mongolian pine). (3) Increasing the temporal frequency of data acquisition can enhance tree species cover fraction inversion accuracy (achieving =0.04 for Larch, =0.07 for Korean pine, and=0.27 for Mongolian pine), whereas its effect on improving forest type cover fraction inversion accuracy is limited. By effectively combining the advantages of synthetically trained data and random forest regression, we have successfully mapped the forest type and tree species cover fraction of Mengjiagang Forest. Moreover, our study provides a comprehensive analysis that accurately quantifies the influence of temporal data acquisition frequency on mapping forest type and tree species cover fraction. This study offers valuable insights into the future mapping of forest type and tree species cover fraction across space and time, particularly for regions with similar species composition. The outcomes of this research will make a significant contribution to the understanding and management of cold-temperate forests, thereby supporting their conservation and sustainable use.
关键词:remote sensing;forest type coverage;tree species coverage;synthetically training data;long time series;machine learning
摘要:Forests constitute the largest carbon pools in terrestrial ecosystems, and elucidating their baseline carbon stock and carbon sink potential is crucial for attaining the nation’s “dual carbon” strategic objectives. Remote sensing, owing to its macroscale, comprehensive, dynamic, rapid, and reproducible nature, has addressed the limitations in carbon accounting for forest vegetation. The estimation of global/regional forest vegetation carbon sinks using spectral information from satellite remote sensing images, vertical structure information collected by airborne/satellite LiDAR, and ground observation data has become a popular topic. Establishing a forest vegetation carbon accounting method based on remote sensing is urgently needed to serve the national “dual carbon” goals and the demand of carbon trading market.This paper introduces a novel system for forest vegetation carbon accounting, grounded in the structure and growth equations of individual trees. Specifically, the system has the following characteristics: integrating airborne and terrestrial LiDAR data to extract structural parameters, including Diameter at Breast Height (DBH), tree height, and crown width, thereby establishing a carbon stock calculation method for forest plots at the individual tree level; developing a pixel-level regional forest carbon stock model with physical interpretability, utilizing canopy height and crown closure as key variables, to mitigate the uncertainty of machine/deep learning remote sensing regression inversion; accurately estimating regional forest carbon sinks by forecasting future forest canopy height and crown closure on the basis of the pixel-level forest carbon stock model and individual tree growth equations.Calculation of forest plot carbon stock. The DBH and tree height parameters are extracted by ground stations and unmanned aerial vehicle LiDAR, the biomass of individual trees is obtained using the allometric growth equation, and the carbon stock of forest plots is calculated. Calculation of regional forest carbon stock. The carbon stock density of forests is closely related to height. A pixel-level forest carbon stock explicit model is established using forest canopy density and height, and the model parameters are automatically calculated from the structure equation of individual trees (DBH-tree height-crown width). Prediction of regional forest carbon sinks. Future forest canopy closure and height are derived using the structure and growth (tree height-tree age) equation of individual trees, and forest carbon stock model and the latest remote sensing data are combined to calculate and update forest carbon sinks.This study adheres to the overarching theme of “forest plot carbon stock-regional forest carbon stock-regional forest carbon sink,” expanding from plot to regional spatial scales and extending from carbon stock to carbon sink across temporal scales, thereby establishing a novel remote sensing-based system for forest vegetation carbon accounting.
摘要:This study aims to accurately estimate the carbon stock in coastal wetland salt marsh vegetation areas, which is essential for understanding the status of wetland carbon pool resources and assessing the carbon sequestration potential of wetlands. Given the limitations of traditional field inventory methods and remote sensing inversion methods, this research explores a two-step inversion modeling approach to improve the estimation process. The methodology involves constructing models that translate measured Aboveground Biomass (AGB) data of three types of salt marsh vegetation to UAV imagery and then to Sentinel-2 satellite imagery. The models are developed using a two-step inversion process, which has allowed for the calculation of vegetation carbon storage, soil carbon storage, and total carbon storage by applying carbon coefficients. This approach is used to achieve synergistic estimation and mapping of AGB and carbon storage. The determination coefficients () for the AGB inversion models of Spartina alterniflora, Phragmites australis, and other salt marsh vegetations are 0.48, 0.42, and 0.45, respectively, with Root-Mean-Square Errors () of 613.89, 650.6, and 624.03 g/m2, respectively. This study reveals that the Ningbo coastal wetland has a total area of salt marsh vegetation of 111.47 km2, with a total AGB of 3.09×105 t and a total carbon stock of 1.68×106 t, equating to a carbon sequestration value of 178 million RMB. The integration of UAV and satellite remote sensing in this study has demonstrated its effectiveness in compensating for the lack of ground sample data, significantly improving the precision and scope of carbon stock estimation in coastal wetlands. This method enables monitoring of wetland carbon storage characterized by low cost, large scale, and high precision, offering a valuable tool for environmental management and policy making related to carbon sequestration in coastal ecosystems.
摘要:Swampy wetlands (forest, scrub, and herbaceous swamps) are among the most important carbon reservoirs on earth and play a pivotal role in the global carbon cycle. The proportion of marshy wetlands in China is nearly 40% of the total wetland area, which is of great significance for maintaining regional biodiversity and ecosystem carbon balance. The Net Primary Productivity (NPP) of vegetation refers to the amount of organic matter accumulated by green plants through photosynthesis minus the remaining part of autotrophic respiration per unit of time and space. It is one of the most important indicators of the carbon sequestration potential of marsh wetlands, which plays a significant role in reflecting the ecological changes of vegetation in the context of climate change.Aiming at the relatively weak research on NPP estimation in China’s swampy wetlands and the saturation problem in the process of NPP estimation, this study estimated the NPP of China’s swampy wetlands in the last 20 years on the basis of MODIS remote sensing data products (MOD13Q1 and MCD12Q1) using the kernel Normalized Difference Vegetation Index (kNDVI) constructed by the radial basis function kernel with the CASA model. Additionally, the spatiotemporal evolution of China’s swampy wetlands and its driving mechanism from 2001 to 2020 were quantitatively analyzed and discussed.The results of the study showed that the coefficient of determination (R2) of NPP_kNDVI estimated using the kNDVI index with the measured value of NPP was 0.854, and the root-mean-square error was 14.46 g C/m2month, which was closer to the real NPP value compared with NPP_NDVI. Compared with the saturation phenomenon of NDVI in highly vegetated areas, kNDVI mitigated the saturation effect of the vegetation index itself, adapted to densely and sparsely vegetated areas, and improved the accuracy of the estimation of NPP of vegetation to a certain extent. The regional pattern of multiyear NPP mean values in China’s swampy wetlands was obvious, showing a decreasing and then increasing trend from low latitude to high latitude. This pattern was the result of a combination of factors, such as the distribution of swampy wetlands, air temperature, precipitation, and solar radiation. The annual mean change in NPP in the study area from 2001 to 2020 ranged from 162.73 g C/m2a to 189.34 g C/m2a, indicating a fluctuating upward trend, with a growth rate of 1.215 g C/m2a (R2=0.82) and a mean value of 177.17 g C/m2a. Between 2001 and 2020, the proportions of areas with increasing and decreasing NPP trends in China’s swampy wetlands were 72.96% and 26.27%, respectively, and were mainly concentrated in the northeastern plains, the northeastern and southwestern parts of Qinghai Province, and the northern part of Sichuan. Compared with human activities, climate change is the main driving factor affecting the spatial and temporal evolution of China’s swampy wetlands, with 66.23% and 33.76% of the influence area respectively.
摘要:High-resolution and noncontact water depth monitoring is crucial for the management and protection of tufa lake landscapes. Satellite-derived bathymetry cannot capture the subtle underwater sedimentary characteristics of tufa lakes. In recent years, the remote sensing technology of light and small Unmanned Aerial Vehicles (UAVs) has gradually been applied to ultrahigh-resolution bathymetric mapping in shallow water areas. However, the classic logarithmic model in water depth inversion is difficult to adapt to the widespread Rayleigh scattering phenomenon in tufa lakes. Therefore, in this study, machine learning methods are used to construct bathymetric inversion models of tufa lakes based on UAV imagery.Taking Spark Lake in Jiuzhaigou National Nature Reserve, Sichuan Province, China as the experimental area, this study extracts aerial image data for bathymetric model construction from UAV platforms. On the basis of pre- and post-earthquake UAV images, a pre-earthquake orthophoto with water and a post-earthquake surface model without water are generated using the structure-from-motion algorithm. After the exclusion of anomalous areas, sample points for the bathymetric inversion are randomly selected. Each sample datum includes the red, green, and blue band digital number values of the pre-earthquake orthophoto and the relative depth values of the post-earthquake exposed terrain relative to the pre-earthquake water surface. Through this dataset, machine learning regression models based on Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP) are constructed. The above machine learning models are trained repeatedly to determine their respective optimal parameters. The accuracy of the estimated bathymetry is verified using the exposed lake terrain after the earthquake.Results indicate that the water depth distribution of the three models has small differences in shallow water areas, and the areas with significant differences are mainly deep-water regions. The bathymetric map simulated by the RF model is susceptible to high-frequency signals, while the bathymetric maps simulated by the SVM and MLP models suffer from localized overestimation of water depth distribution. In terms of accuracy assessment, the RF, SVM, and MLP models have Root-Mean-Square Errors (RMSE) of 0.816, 0.945, and 0.832 m and coefficients of determination (R2) of 0.948, 0.930, and 0.946, respectively. The RF and MLP models have relatively good consistency across the entire depth range, while the SVM model has an overestimation of depth in general within the interval of 6—9 m.To sum up, machine learning models exhibit higher accuracy in water depth retrieval compared with traditional logarithmic models. Among them, the RF and MLP models are more suitable than the SVM model for water depth retrieval in tufa lakes based on UAV imagery. Unlike the models only utilizing blue and green bands, the introduction of red bands into machine learning models improves the accuracy of shallow-water bathymetry while increasing local bathymetry uncertainty. In the future, further research using UAV multispectral imagery with the coastal blue band is necessary. Given an adequate dataset, constructing a deep convolutional neural network-based bathymetric inversion model for tufa lakes is proposed.
摘要:Digital Elevation Models (DEMs) are indispensable data sources for natural resource investigation, climate change analysis, and disaster monitoring and assessment. TanDEM-X mission, as the first twin-satellite Interferometric Synthetic Aperture Radar (InSAR) system, has successfully obtained a high-precision global DEM with 12 m resolution. However, limited by the penetration capability of shortwave signals, DEMs acquired in dense forest areas are usually contaminated by forest canopy signals and are difficult to meet practical applications. The Phase Center Height (PCH) caused by forest volume scattering needs to be removed from InSAR-derived DEM to obtain sub-canopy topography. Unfortunately, TanDEM-X acquires single-baseline, single-polarization data in the global standard mode, which is difficult to meet the needs of existing model solutions and requires the introduction of external data. In this study, we propose a machine learning-based method to estimate sub-canopy topography by combining TanDEM-X InSAR, ICESat-2, and Landsat 8 OLI data. The effectiveness of the proposed method was tested and validated in the Gabon rainforest and the Spanish boreal forest. In the Gabon rainforest test site, compared with that of the airborne LiDAR Digital Terrain Model (DTM), the Root-Mean-Square errors (RMSEs) of the InSAR DEMs corresponding to two locations are 14.70 m and 18.58 m. After PCH removal, the accuracy is improved to 5.54 m and 5.86 m, which represents an improvement of over 60%. In the Spanish northern forest test site with complex terrain, the RMSE of sub-canopy topography decreased from 6.05—9.10 m to 3.06—4.42 m. In addition, we investigate the necessity of the proposed method to use InSAR observations and the effect of the accuracy of the ICESat-2 control points used on the sub-canopy topography estimation. These satisfactory results demonstrate the potential of the proposed method in estimating sub-canopy topography for future spaceborne InSAR missions (e.g., TanDEM-L and LT-1) when only single-baseline, single-polarization data are available. Furthermore, by combining the high resolution of TanDEM-X and the strong penetration of BIOMASS, the proposed method has the potential to estimate sub-canopy topography with higher accuracy and resolution in the future.
关键词:sub-canopy topography;InSAR;phase center height;machine learning;TanDEM-X;ICESat-2
摘要:Mountain ecosystems, covering approximately 24% of the terrestrial surface, are the key component of earth’s carbon cycle in terrestrial ecosystems. Vegetation in mountain ecosystems can regulate the energy budget via mediating the exchange of energy and substance and thus has been regarded as an essential bioindicator for the global climate change over the past decades. Accurate estimation of mountain vegetation Gross Primary Productivity (GPP) plays a vital role in understanding the function of mountain ecosystems and characterizes the ecosystem responses to climate change. Owing to the effect of complex mountainous conditions and the limitations from spatial resolutions, obvious topographic errors and spatial scaling errors in mountain vegetation GPP estimates occur. Thus, evaluating the error sources in the estimation of mountain vegetation GPP across multiple spatial scales is crucial.In this study, we selected the Wanglang National Nature Reserve—a typical mountainous ecosystem in southwest China—as the study area. An eco-hydrological model called Boreal Ecosystems Productivity Simulator-TerrainLab was used to obtain the vegetation GPP and analyze the topographic and spatial scaling errors at fine, medium, and coarse spatial scales (i.e., 30, 480, and 960 m). First, the topographic errors in estimating vegetation GPP were evaluated across four scenarios that characterized the effects of different topographic features at the fine, medium, and coarse spatial scales. Then, spatial scaling errors were illustrated at the scales of 480 and 960 m. Finally, the agreement index (d), determination coefficient (R2), Root-Mean-Square Error (RMSE), and Mean Bias Error (MBE) were used to evaluate the topographic and spatial scaling errors in modeling mountain vegetation GPP at the fine, medium, and coarse spatial scales.Results showed that the multiscale vegetation GPP estimates across different simulation conditions presented obvious spatial differences (the difference among regional mean values upped to 198 g/(m2·a)). The topographic errors of vegetation GPP estimates showed a decreasing trend with the decrease in spatial resolution, suggesting that considerable attention should be paid to high spatial resolutions (the MBE value was 200 g/(m2·a)). Specifically, the error caused by ignoring the redistribution of soil water was observed to be the largest source of topographic errors. As for the spatial scaling errors, an increasing trend with the decrease in spatial resolution was found, highlighting the necessity of reducing the spatial scaling errors in middle- and coarse-spatial-resolution GPP estimates (161 and 210 g/(m2·a), respectively).In generating multiscale mountain vegetation GPP products, the topographic effects on high-spatial-resolution GPP estimation should be eliminated. Attention should also be given to the spatial scaling errors of GPP products at middle and coarse spatial resolutions. In consideration of the evident topographic errors caused by ignoring water redistribution, accurate estimation of soil moisture would improve the quality of GPP products over mountainous areas, especially the products at high spatial resolutions.
摘要:The MicroWave Radiation Imager (MWRI) on China’s Fengyun-3 (FY-3) satellite can provide effective data for various fields, such as meteorological research and weather forecasting. When using MWRI data for coinversion of multifrequency detection channels, considering resampling is necessary to make the data of different detection frequency points have a consistent spatial resolution. The Backus-Gilbert (BG) method is a widely used resampling method for spaceborne microwave radiometers. The traditional resampling method uses the BG method to calculate the weight coefficient of the sampling points obtained by a single rotation scan of the radiometer (precalculated weight coefficient) and directly applies it to all sampling points generated by the rotation scan, thereby completing the resampling of all data. However, during the sampling process of satellite instruments, the influence of factors such as the ellipsoid of the Earth will lead to changes in the relative positions between the observation fields, and directly applying the precalculated weight coefficients to all the data will induce errors in resampling. Therefore, in this paper, on the basis of the orbit parameters and pattern information of FY-3D MWRI, we propose a resampling method combined with antenna pattern projection positioning and analyze the resampling effect of MWRI at different orbit positions in detail.In the resampling experiments with fixed scan lines, we implement resampling for different combinations of source and target channels on the 200th scan line of the sample data. Results show that the traditional resampling method can achieve a better resampling effect by applying the BG method on a fixed scan line. In resampling experiments at different scan lines, we randomly select sample data in the time range from 2019-04-01 to 2019-12-11 and calculate the positional changes of the adjacent points in the source pattern of the 1st, 44th, and 133rd scan columns with the increase in the scan lines of the sample data. The fit error due to positional changes is then calculated using precomputed weighting factors. The maximum change in the position of the adjacent points can lead to a deterioration in the fit error of about 0.09. In the error analysis experiment, we verify the correction effect of the resampling method adopted in this study on the Brightness Temperature (BT) error on the MWRI sample data and long-term data by calculating the distribution of the BT difference between channels. The adopted resampling method can correct the average BT error of 1.32 K on 10 different channel combinations of MWRI.In conclusion, this study analyzes the resampling effect of the MWRI instrument at different orbital positions in detail. A source of error in resampled BT is identified and corrected. In the future, when multifrequency detection channel data are used to collaboratively retrieve geophysical parameters, more accurate retrieval results may be obtained by applying the resampled data generated in this study on the basis of the sensitivity of the parameters to changes in BT.Result:Conclusion
摘要:The observation of geostationary orbit satellites, with the characteristics of wide range, high frequency, and fixed point observation, provides an important way to obtain land surface and atmospheric parameters and can monitor the change in land surface temperature over long time series. As the second generation of China’s geostationary meteorological satellites, Fengyun-4 A and B satellites’ constellation observation greatly expand the scope of meteorological observation and improve data utilization efficiency. Before the joint use of the two satellites’ data, the radiometric consistency between the same-band observations of the two satellites needs to be explored.Taking the thermal infrared band data as an example, this study assessed the radiometric consistency between three thermal infrared bands (centered at 8.5 µm, 10.8 µm, and 12.0 µm) of Fengyun-4 A and B satellites in four experimental areas: Dunhuang calibration field, Hulunbuir Grassland, Chaohu Lake, and South China Sea. The heterogeneity of the Dunhuang calibration field and Hulunbuir Grassland study areas was evaluated first. A method was then proposed to correct the angular effects and spectral response function differences in observation from geostationary orbit satellites. This method corrected the angular effects by establishing empirical relationships between simulated radiance data at different viewing angles and eliminated spectral response function differences by correcting brightness temperature and radiance on the basis of lookup tables. Finally, this method was used to correct the thermal infrared data of the two satellites at the same observation time, and the brightness temperatures after correction were compared to analyze the radiometric consistency of the corresponding thermal infrared bands.Based on the data analysis, after the removal of the differences in space-time, observation angle, and spectral response function of the two satellites, the results show a strong positive correlation between the brightness temperatures of the three thermal infrared bands on Fengyun-4 A and B satellites. The brightness temperature errors are small, indicating good radiometric consistency. However, a slight variation in brightness temperature exists among those different thermal infrared bands. The consistency of brightness temperature of the second thermal infrared band is better than that of the third band, and the third band performs better than the first band. The root-mean-square errors of brightness temperatures for the three bands range from 0.28 K to 1.51 K, with deviations between -1.13 K and 0.85 K. The brightness temperature deviations in the second and third bands exhibit a noticeable positive skewness, whereas the deviations in the first band show a negative skewness.Comparison of the brightness temperature data before and after correction suggests that the method proposed in this paper has good applicability for the study of radiometric consistency in thermal infrared bands of geostationary orbit satellites. The results show that the radiometric consistency of thermal infrared radiation between the two satellites is generally good, although it might be influenced by land cover types. The findings presented in this paper provide important guidance for the joint utilization of thermal infrared data from Fengyun-4 A and B satellites.
摘要:Band selection is a crucial task in the dimensionality reduction of hyperspectral remote sensing imagery. Its objective is to choose a subset of bands with minimal redundant information, high information content, and class discriminability. To address the issues of existing band selection methods based on nearest neighbor subspace partitioning, which do not consider the spatial distribution of objects and neglect the effect of noisy bands when computing cluster centers, this study proposes a hyperspectral image band selection method that integrates a spatial-spectral structure with improved local density, referred to as ISSS-ELD.This method first performs image segmentation on the hyperspectral image using an entropy-based approach to acquire homogeneous regions. A composite region-level neighboring band correlation coefficient vector is obtained by integrating the correlation coefficient matrix of these homogeneous regions. Subsequently, a Gaussian kernel is applied to globally smooth the neighboring band correlation coefficient vector, reducing the influence of noisy bands. Bands are grouped on the basis of extremum points in the smoothed vector. The product of the maximized improved local density and band information entropy serves as the criterion for selecting representative bands.This study conducts experiments on hyperspectral image datasets, including Indian Pines, Botswana, and Salinas. Different band selection methods are evaluated by calculating metrics such as classification accuracy, average correlation coefficient, and noise robustness of the selected bands. The results are as follows: (1) Compared with pixel-level correlation-based partitioning methods, the utilization of region-level correlation coefficients results in more reasonable grouping of neighboring bands, reducing band redundancy while retaining some potential characteristic bands. The classification performance on the three datasets is improved by 2.63%, 0.68%, and 0.16%. (2) In contrast with methods solely using information entropy for band assessment, the proposed approach of maximizing the product of improved local density and information entropy proves effective. On the three datasets, the Overall Accuracy (OA) is increased by 4.13%, 0.5%, and 0.21%. (3) Compared with six other advanced band selection methods, the proposed method achieves significant performance improvements: OA is increased from 62.34% to 75.03%, from 86.74% to 88.28%, and from 86.04% to 92.36% on the three datasets. Furthermore, the selected subset of bands by our method is dispersed, concentrating in regions with higher information entropy and effectively avoiding the inclusion of noisy bands.In summary, the band subset selected by the proposed band selection method exhibits low redundancy, high information content, strong class separability, and robustness against noise, effectively addressing the challenges in hyperspectral image band selection.
摘要:Wetland is an important ecosystem and plays a vital role in maintaining regional ecological security. Wetland structure changes respond sensitively to natural and human activities, and flood wetlands experience drastic seasonal water and vegetation changes due to intermittent flood inundations. Mapping high-accuracy wetland structures is challenging because of frequent water and vegetation alternations, which cause spectral confusion and misclassification in optical satellite images. Several wetland extraction methods are available today, including object-oriented methods, whose parameters need to be decided subjectively, and machine learning methods, which have relatively low accuracy. With the continuous development of deep learning in image semantic segmentation, a precise and automatic remote sensing image binary classification becomes possible. Recent studies have suggested that deep learning semantic segmentation methods show great potential for mapping wetland changes in high-resolution images. However, the extraction of wetland structures in complex floodplain scenarios places high demands on models in terms of mining deep spatial information. The deformable U-Net (D-UNet) semantic segmentation model is improved to enhance the accuracy of the extraction of floodplain wetland structure.In this study, the Taitema Lake in Xinjiang, China was selected as the study area because it is a typical floodplain wetland in the arid zone. A multiscene and multitemporal wetland sample dataset was collected using Sentinel-2 remote sensing images in the study area. The D-UNet for wetland structure extraction used VGG16 to build the encoding/decoding network and focused on improving the convolutional layer in the network. D-UNet was improved by replacing the convolution block before dimensionality reduction with multiscale dilated convolutions to enhance the network’s receptive field, fuse features of different scales, and avoid loss of detailed information in high-resolution remote sensing images. After pretraining D-UNet, we determined that a multiscale convolution module consisting of three scales with dilation rates of 1, 2, and 3 would maximize the network’s receptive field. We eventually input multiple remote sensing images from multiple scenes to fine-tune our model.The applicability of the improved D-UNet model, traditional index-thresholding methods, and four classical semantic segmentation networks for extracting wetland structural information in floodplains was compared. Results showed that the improved D-UNet had an overall accuracy of 96.3% in single-temporal image wetland structure extraction, with a kappa coefficient of 0.839. Moreover, it demonstrated better transferability on time-series images, with an overall multitemporal accuracy of 92.3%. Compared with five models and the index-thresholding method, the improved D-UNet model showed better application potential in the extraction of floodplain wetland structure. It reduced misclassification and omission of wetland water bodies and vegetation by 7.2% and 48.9% compared with the index-thresholding method and by 0.6% and 5.4% compared with D-UNet, respectively.This study proposes an effective classification method for the identification of fine structures in floodplain wetlands. It verifies the excellent performance of the semantic segmentation model in the extraction of complex feature information from remote sensing images. The improved D-UNet model can be used for information extraction for floodplain wetlands in similar environments. It provides a reference for rapid automated mapping of large-scale wetlands.
摘要:Change Detection (CD) is the process of identifying the land cover changes of interest by analyzing bitemporal remote sensing images acquired in the same geographical area at different times. As a popular research topic in remote sensing, CD plays an important role in many practical applications, such as urban research, environmental monitoring, and disaster assessment. With the rapid development of deep learning, its applications have expanded to the remote sensing CD task, leading to the development of numerous deep learning-based CD methods. Although the existing deep learning-based CD methods have their own advantages and can implement the CD task effectively, their applicability and robustness need to be further analyzed. How to design an appropriate network model to obtain more reliable and accurate CD results is still an open problem. Currently, two main challenges remain for CD: (1) The changed objects existing in remote sensing images have various scales and shapes. (2) The proportions of the changed and unchanged classes are often significantly unbalanced, i.e., the number of the changed pixels is usually much less than that of the unchanged pixels. To deal with these two challenges, this study proposes a novel end-to-end CD network for high-resolution remote sensing images based on a residual network, i.e., Hybrid Spatial Pyramid Pooling Network (HSPPNet). First, a Hybrid Spatial Pyramid Pooling (HSPP) module is built by integrating atrous convolutions and attention mechanism-guided convolutions in parallel to effectively extract the changed objects with different shapes and scales from high-resolution remote sensing images. Then, an adaptive balancing loss function is presented to alleviate the effects of the serious imbalance between changed and unchanged classes on CD. The loss function is constructed by defining a truncation–compensation weighting cross entropy function and a class-level IoU function and integrating them. Finally, a simple but effective input module that considers the bitemporal remote sensing images and their difference image is designed to enhance the change information. Owing to the above three points, the proposed HSPPNet method enhances the performance of deep learning CD. Experiments are carried out with two public CD datasets, namely, the CDD and Google datasets, to evaluate the performance of the proposed HSPPNet method. Six state-of-the-art deep learning methods are used as the comparative methods. Quantitative analysis of CD results is done on four evaluation metrics, namely, precision, recall, overall accuracy, and F1-score. For the CDD and Google datasets, the overall accuracy/F1-score values of the proposed HSPPNet method are 0.9943/0.9758 and 0.9399/0.9076, respectively, which are superior to those of the six comparative methods. The experimental results on two public CD datasets demonstrate that the proposed HSPPNet is effective and feasible. In addition, the results of ablation experiments show that the HSPP module, the adaptive balancing loss function, and the input module proposed in this work can enhance the CD performance effectively.
摘要:Superpixel segmentation offers significant advantages for information extraction from SAR images. First, it effectively reduces data volume and enhances the efficiency of subsequent applications. Second, it effectively reduces noise interference in SAR images, thereby improving data quality. Third, it preserves the edge features of images, which is beneficial to the SAR image postprocessing stages, such as deep learning-based classification. Lastly, the results of superpixel segmentation can be directly used as inputs for graph convolutional networks to explore the application of superpixel-based graph convolutional networks. As a result, SAR image superpixel segmentation has found extensive application in ship monitoring, water body extraction, and various other fields. Existing superpixel segmentation algorithms for SAR images predominantly rely on local clustering methods; however, they exhibit certain shortcomings, including a predefined number of superpixels, limited adaptability, and the necessity for multiple iterations. To overcome these limitations, this study proposes a novel adaptive superpixel segmentation algorithm (ASSA). This algorithm maximizes the benefits derived from Gaussian mixture models, neighborhood properties, and priority queues.Method First, this study proposes an adaptive adjustment strategy for seeds to overcome the challenges associated with predefined number of superpixels and limited adaptability. The strategy is based on Gaussian mixture models, involving seed adjustment and generation using homogeneity discrimination criteria. Second, the algorithm solves the issue of multiple iterations by implementing single-iteration superpixel segmentation using neighborhood properties and priority queues under the neighborhood compulsory connection. Lastly, the algorithm tackles severe speckle noise in SAR images by employing a Gaussian kernel function to smooth the unmarked pixels and a postprocessing algorithm to eliminate isolated superpixels. In this study, we use nine Sentinel-1 images to evaluate ASSA in terms of visualization effect, quantitative accuracy, and runtime efficiency. Results show that compared with existing superpixel segmentation algorithms, ASSA achieves higher boundary adherence and internal homogeneity while improving segmentation efficiency. In particular, the boundary recall rate is improved by 11.3% and 15.9% compared with SLIC and ESOM, respectively, while the undersegmentation error rate is reduced by 33.3% and 29.4%, respectively. This study proposes a single-iteration superpixel adaptive segmentation algorithm based on neighborhood characteristics and an adaptive adjustment strategy for seeds. This algorithm combines Gaussian mixture models with superpixel homogeneity discrimination to achieve adaptive segmentation. The experimental results demonstrate that ASSA is an effective and efficient method for SAR image superpixel segmentation.
关键词:SAR;superpixel segmentation;priority queue;adaptive adjustment strategy for seeds;Gaussian Mixture Model
摘要:Wetland is an important ecosystem and plays a vital role in maintaining regional ecological security. Wetland structure changes respond sensitively to natural and human activities, and flood wetlands experience drastic seasonal water and vegetation changes due to intermittent flood inundations. Mapping high-accuracy wetland structures is challenging because of frequent water and vegetation alternations, which cause spectral confusion and misclassification in optical satellite images. Several wetland extraction methods are available today, including object-oriented methods, whose parameters need to be decided subjectively, and machine learning methods, which have relatively low accuracy. With the continuous development of deep learning in image semantic segmentation, a precise and automatic remote sensing image binary classification becomes possible. Recent studies have suggested that deep learning semantic segmentation methods show great potential for mapping wetland changes in high-resolution images. However, the extraction of wetland structures in complex floodplain scenarios places high demands on models in terms of mining deep spatial information. The deformable U-Net (D-UNet) semantic segmentation model is improved to enhance the accuracy of the extraction of floodplain wetland structure.In this study, the Taitema Lake in Xinjiang, China was selected as the study area because it is a typical floodplain wetland in the arid zone. A multiscene and multitemporal wetland sample dataset was collected using Sentinel-2 remote sensing images in the study area. The D-UNet for wetland structure extraction used VGG16 to build the encoding/decoding network and focused on improving the convolutional layer in the network. D-UNet was improved by replacing the convolution block before dimensionality reduction with multiscale dilated convolutions to enhance the network’s receptive field, fuse features of different scales, and avoid loss of detailed information in high-resolution remote sensing images. After pretraining D-UNet, we determined that a multiscale convolution module consisting of three scales with dilation rates of 1, 2, and 3 would maximize the network’s receptive field. We eventually input multiple remote sensing images from multiple scenes to fine-tune our model.The applicability of the improved D-UNet model, traditional index-thresholding methods, and four classical semantic segmentation networks for extracting wetland structural information in floodplains was compared. Results showed that the improved D-UNet had an overall accuracy of 96.3% in single-temporal image wetland structure extraction, with a kappa coefficient of 0.839. Moreover, it demonstrated better transferability on time-series images, with an overall multitemporal accuracy of 92.3%. Compared with five models and the index-thresholding method, the improved D-UNet model showed better application potential in the extraction of floodplain wetland structure. It reduced misclassification and omission of wetland water bodies and vegetation by 7.2% and 48.9% compared with the index-thresholding method and by 0.6% and 5.4% compared with D-UNet, respectively.This study proposes an effective classification method for the identification of fine structures in floodplain wetlands. It verifies the excellent performance of the semantic segmentation model in the extraction of complex feature information from remote sensing images. The improved D-UNet model can be used for information extraction for floodplain wetlands in similar environments. It provides a reference for rapid automated mapping of large-scale wetlands.
摘要:Point cloud simplification is a prerequisite for efficient transmission and multiscale applications of massive airborne LiDAR ground point clouds. However, existing ground point cloud simplification methods suffer from poor applicability in complex environments and loss of terrain detail features. This study proposes an airborne LiDAR point cloud clustering simplification algorithm considering terrain features and boundary protection against contraction. First, the point cloud is segmented into initial point cloud clusters using k-means algorithm. Then, further subdivisions are performed on the basis of the terrain complexity of each cluster. Subsequently, terrain feature points at different terrains are identified using the point cloud normal vector information within the subdivided subclusters and the elevation differences of edge points between adjacent clusters. Finally, boundary feature points of the target area are preserved to prevent the contraction of the original point cloud boundary. In six groups of point cloud scenes with high terrain complexity, the proposed method is analyzed and compared with seven classical point cloud simplification methods, namely, random, voxel grid, curvature-based, maximum Z tolerance, graph-based, multi-index weighted, and iterative simplification methods. The experimental results demonstrate that compared with the traditional methods, the proposed method achieves a minimum reduction of 12.1% in the average root-mean-square error of the generated digital elevation models (DEMs) and a minimum reduction of 9.6% in the average mean absolute error. The derived products, including average slope and terrain roughness, also exhibit closer agreement with the reference values. The qualitative analysis results indicate that the DEM constructed by the proposed method aligns better with the reference DEM and provides more accurate and detailed terrain features. The above experimental results demonstrate that the proposed method effectively reduces the accuracy loss caused by simplification of DEMs while maintaining strong adaptability to terrain. This method can be applied to intelligent simplification of airborne LiDAR point cloud data, enabling the construction of high-precision DEMs to meet the requirements of geoscientific analysis for high accuracy and efficiency.
关键词:airborne LiDAR;point cloud simplification;K-means;terrain features;Digital Elevation Model (DEM)
摘要:Mobile Laser Scanning (MLS) systems are widely used in various fields owing to their ability of rapidly acquiring high-precision and high-density 3D point cloud data, particularly in the acquisition of urban spatial information. Given that urban MLS point clouds exhibit complex scenes, large data volumes, and uneven spatial distribution, accurate classification of large-scale urban point clouds presents significant challenges. Currently, many deep learning point cloud classification methods enhance feature representation of point clouds by adding a feature aggregation module. Nonetheless, this approach frequently results in increased training parameters and model overfitting.We propose an MLS point cloud classification method integrating resemble prediction constraints and error prediction entropy maximization. The proposed method can enhance the point cloud feature representation of the baseline network and improve the generalization ability of the model without increasing the training parameters. Our method consists of three main components: a basic supervision branch, an ensemble prediction constraint branch, and an error prediction entropy maximization branch. Specifically, we first employ RandLA-Net as the backbone network to obtain point cloud classification features. Then, a basic supervised branch calculates the weighted cross-entropy loss on the basis of true labels and predicts probability distributions and category weights to provide a basic fully supervised signal for model training. For the ensemble prediction constraint branch, we first generate ensemble predictions by recording the predicted values during the point cloud training process. Because the input to RandLA-Net is a random subpoint cloud, the ensemble predictions can be integrated for predictions not only at different stages but also at different relative positions. Thus, the ensemble prediction is highly robust to the current prediction. Afterward, we apply a consistency constraint to minimize the difference between the two predictions to improve the point cloud feature representation. Finally, we design an error prediction entropy maximization branch to maximize the entropy of error prediction point sets, increasing their confusion to reduce the model overfitting.The public MLS point cloud dataset Toronto3D is chosen as the primary dataset to validate the performance of the proposed method. The qualitative result (Fig. 5) and quantitative result (Table 1) on the Toronto3D dataset show that the proposed method can correctly classify most points. To verify the validity of the method, we compare the accuracy of the proposed method with those of other popular methods. The comparison results (Table 1 and 2) indicate that the proposed method obtains the best OA (97.71%) and mIoU (83.68%). A series of ablation experiments is carried out to verify the effectiveness of each branch of the proposed method. The results (Table 3 and Fig. 7) show that each branch can effectively improve the model classification performance. The complexity analysis (Table 4) indicates that the proposed method can enhance the accuracy of the baseline method without increasing the model parameters. Furthermore, the experimental results (Table 5) on other public MLS datasets (WHU-MLS and Paris datasets) demonstrate that the proposed method can obtain competitive results on multiple datasets.