Due to dynamic characteristic of demand response and stochastic nature of power generation, it brings great challenge to smart energy management. In this paper, a demand response model is created with two-level dynamic pricing transaction among grid operator, service provider and customers, which also involves customers’ active participation with load shifting issue. To effectively control system load on the demand side, an improved deep reinforcement learning approach is proposed with a recursive least square (RLS) technique to deal with the dynamic pricing demand response problem, which accelerates the on-line training and optimization efficiency. On the power generation side, a probabilistic penalty-based boundary intersection (PBI) based multi-objective optimization algorithm is improved to optimize the economic cost, emission rate and statistic voltage stability index (SVSI) simultaneously with generated stochastic scenarios, which can ensure energy conservation and environmental protection, as well as system security. The case results reveal that the proposed two-level optimization strategy successfully deals with energy management with dynamic pricing demand response.Note to Practitioners—This paper is motivated by solving stochastic energy management issue of isolated power system with dynamic pricing demand response. Those existing methods merely focus on the load demand or power generation side, and the methods for demand response issue lacks efficient on-line learning ability, while this work proposes a recursive least square based deep reinforcement learning approach to tackle with the two-level dynamic pricing demand response issue, scenario based PBI multi-objective optimization is proposed to solve the power dispatch issue on power generation side, and the numerical analysis results suggest that the proposed optimization strategy can deal with the whole energy management issue well. The future work will focus on the dynamic power-load coordination in the energy management issue.

Accurate understory terrain estimation is a key challenge in ecological modeling and forest resource management. Traditional optical remote sensing is affected by significant signal saturation in vegetated areas, while spaceborne LiDAR systems such as ICESat-2 are limited in supporting regional-scale continuous modeling due to their sparse and discrete footprint coverage. This study integrates filtered ICESat-2 understory elevation control points with optical remote sensing data to comprehensively assess the applicability and predictive accuracy of various machine learning models across different regions. By carefully selecting and optimizing models, the most suitable approach for each study area was identified, enabling precise regional-scale understory terrain estimation. Through multi-source remote sensing data fusion, a continuous surface elevation model was constructed, substantially enhancing overall estimation accuracy. Experimental results demonstrate a notable accuracy improvement, with ME = 0.42 m, RMSE = 2.80 m, and STD = 2.77 m. Furthermore, this study systematically quantifies the influence of environmental factors such as forest type, landform features, slope, aspect, and forest canopy height on estimation accuracy. Beyond advancing methodologies for high-precision understory terrain estimation, this study leverages machine learning optimization and multi-source data fusion to overcome the limitations of ICESat-2’s footprint coverage, providing robust technical support for global-scale understory terrain monitoring and ecosystem research.

To enhance the accuracy of the Shuttle Radar Topography Mission (SRTM) DEM, this study establishes a mathematical modeling approach based on Ice, Cloud, and land Elevation Satellite-2 (ICESat-2) data to interpolate and correct the SRTM DEM of understory terrain. Puerto Rico in the United States was selected as the study area. First, the elevation error was obtained by matching ICESat-2 with SRTM DEM. Then, interpolation models for SRTM DEM were built using the methods of IDW, Kriging, Spatial Analyst, and Trend for correction. Finally, the accuracy of the interpolate models was evaluated based on ICESat-2 data. In addition, to explore the factors influencing the elevation accuracy of SRTM DEM, linear regression analysis was conducted on two dimensions: canopy height and forest type. The results indicate that this method significantly improves the accuracy of the SRTM DEM in understory terrain, and the IDW interpolation model proved to be the most effective, reducing RMSE from 13.40 to 6.74 m, improving the elevation accuracy of SRTM DEM for understory terrain by 49%. This achievement not only provides an effective method for improving the quality of SRTM DEM but also lays a solid foundation for subsequent geospatial analysis and environmental monitoring work.

Advanced Topographic Laser Altimeter System (ATLAS) is a new micropulse photon-counting laser system that offers unprecedented options for the observation of forest ecosystems. However, the ATLAS system is sensitive to solar background noise, which poses a tremendous challenge to the photon cloud noise filtering for various observation time scenes in a forest environment. This article presents a multilevel adaptive photon cloud noise filtering algorithm (MLAPCNF) for different observation time scenes that integrate the improved density-based spatial clustering of applications with noise (DBSCAN) algorithm and the improved localized statistics algorithm. The MLAPCNF algorithm was tested at different observation time scenarios, laser intensities, and forest coverage using the ATLAS dataset for forests located in nine study areas in the United States. The results showed that the MLAPCNF algorithm was effective in identifying noise photons and preserving signal photons in the raw ATLAS data with an$R$value of 0.99 and$F$value of 0.79 which produced marginally superior results than the other existing filtering methods. The$F$values of the MLAPCNF algorithm under daytime observation conditions were 0.01–0.03 higher than those under nighttime observation conditions, indicating that the algorithm performed better under daytime observation conditions. Results demonstrated that the proposed method can eliminate the impact of observation time differences in forest environments. Overall, the MLAPCNF algorithm outperforms the other existing filtering techniques at the given test site and is capable of delivering accurate data for estimating forest structural parameters.

Accurate tumor segmentation is crucial for esophageal cancer radiotherapy treatment planning. The low contrast among the esophagus, tumors, and surrounding tissues, and irregular tumor shapes limit the performance of automatic segmentation methods. In this paper, we aim to exploit the irregular shapes of tumors to facilitate accurate segmentation. We propose a simple and pluggable shape-aware contrastive deep supervision network (SCDSNet) with shape-aware regularization and voxel-to-voxel contrastive deep supervision. Specifically, the shape-aware regularization with an uncertainty minimization strategy encourages the precise predictions of an additional shape-aware head. The voxel-to-voxel contrastive deep supervision enhances the multi-scale shape-tumor contrast for better voxel-to-voxel prediction of shapes. The proposed method is simple and highly pluggable, which can easily be extended to other frameworks. Further, we establish a large in-house dataset on esophageal cancer to validate the effectiveness of our proposed method. The quantitative and qualitative experimental results demonstrate the effectiveness of SCDSNet on the esophageal cancer dataset.

The Global Ecosystem Dynamics Investigation (GEDI) contains a full waveform, multi-beam laser altimeter, which provides vertical vegetation structure information. However, few studies have assessed the accuracy of using GEDI LiDAR data in retrieving ground topography from dense forest environments. There has been no in-depth study assessing the accuracy of GEDI data with different kinds of elevation index in footprint samples from densely forested areas (canopy cover >70%). To address this limitation, this study aims to assess the accuracy of ground topography estimates from dense forested terrain through the combination of airborne LiDAR data. The results show that under dense forest conditions, the mean elevation in the GEDI footprint performed more accurate than the max elevation, the min elevation and the median elevation in the GEDI footprint. GEDI provides reasonable estimates of mean terrain height, with R2of 0.99, root mean squared error (RMSE) of 6.08 m and mean absolute error (MAE) of 3.92 m. In addition, dense forest ecosystems also reduce the accuracy of the terrain height estimates. Overall, GEDI with its full waveform technology represents promising spaceborne LiDAR for terrain height retrieval in dense forest environments.

As a continuation of the Ice, Cloud, and land Elevation Satellite (ICESat), the ICESat-2 contains a micro-pulse photon-counting Advanced Topographic Laser Altimeter System (ATLAS), which is expected to provide comprehensive earth observation data. However, the abundant noise photons present in the ICESat-2 data pose a tremendous challenge to photon data noise filtering algorithms. There have been many studies on photon cloud noise filtering algorithms, including the official noise filtering algorithm of NASA named the differential, regressive, and Gaussian adaptive nearest neighbor (DRAGANN) algorithm. However, to data there has been no in-depth study on an evaluation criterion of the ICESat-2 reference signal photon data. To address this limitation, in this study a novel evaluation criterion of the ICESat-2 reference signal photon data is proposed through the combination of Goddard’s light detection and ranging (LiDAR) and hyperspectral and thermal imager (G-LiHT) data for evaluating the noise filtering performance of the DRAGANN algorithm on ATL08 data. The proposed evaluation criterion uses the G-LiHT digital terrain model (DTM) data as a signal photon lower boundary and the DTM + canopy height model (CHM) data as a signal photon upper boundary and thus can describe the signal photon range more accurately and systematically than the manually labeled reference data. In the study area, the DRAGANN algorithm can achieve mean recall, precision, accuracy, and$F$-value of 0.97, 0.66, 0.73, and 0.77, respectively. The results show that DRAGANN algorithm can filter noise photons from the ATLAS data effectively at different laser intensities and observation times. Also, the results demonstrate that the observation time has a greater influence than the laser beam intensity on the noise filtering accuracy of the DRAGANN algorithm.

In this paper, we develop a new library TenDeC++ for tensor decompositions in C++. TenDeC++ supports popular tensor decomposition functions including Canonical Polyadic, Tucker, tensor-train, and t-SVD, assisting C++ programmers to shorten the development cycle of deep learning applications. Compared with the resource-intensive Python and MATLAB, C++ has the nature advantages on fast running time and high compatibility. To further explore potentials of C++, we propose a novel underlying technology PointerDefomer leveraging the unique pointer. Since the transformation between tensor and size-specific matrix is indispensable in tensor decompositions, PointerDefomer can virtually achieve such a transformation by controlling the movement of pointer in memory address. As a result, the conventional transformation steps can be skipped to accelerate the decomposition process and there is no memory needed for saving the intermediate results of tensor transformation. In our experiment, TenDeC++ reduces decomposition time and support larger size of tensor compared with the classic Tensorly in Python and TensorLab in MATLAB, respectively.