Search Results (1,253)

Diffusion models are among the most common techniques used for image generation, having achieved state-of-the-art performance by implementing auto-regressive algorithms. However, multi-step inference processes are typically slow and require extensive computational resources. To address this issue, we propose the use of an information bottleneck to reschedule inference using a new sampling strategy, which employs a lightweight distilled neural network to map intermediate stages to the final output. This approach reduces the number of iterations and FLOPS required for inference while ensuring the diversity of generated images. A series of validation experiments were conducted involving the COCO dataset as well as the LAION dataset and two proposed distillation models, requiring 57.5 million and 13.5 million parameters, respectively. Results showed that these models were able to bypass 40–50% of the inference steps originally required by a stable U-Net diffusion model, which included 859 million parameters. In the original sampling process, each inference step required 67,749 million multiply–accumulate operations (MACs), while our two distillate models only required 3954 million MACs and 3922 million MACs per inference step. In addition, our distillation algorithm produced a Fréchet inception distance (FID) of 16.75 in eight steps, which was remarkably lower than those of the progressive distillation, adversarial distillation, and DDIM solver algorithms, which produced FID values of 21.0, 30.0, 22.3, and 24.0, respectively. Notably, this process did not require parameters from the original diffusion model to establish a new distillation model prior to training. Information theory was used to further analyze primary bottlenecks in the FID results of existing distillation algorithms, demonstrating that both GANs and typical distillation failed to achieve generative diversity while implicitly studying incorrect posterior probability distributions. Meanwhile, we use information theory to analyze the latest distillation models including LCM-SDXL, SDXL-Turbo, SDXL-Lightning, DMD, and MSD, which reveals the basic reason for the diversity problem confronted by them, and compare those distillation models with our algorithm in the FID and CLIP Score. Full article

While numerous studies have investigated the NDVI-LST relationship at local or regional scales, existing global analyses are outdated and fail to incorporate recent environmental changes driven by climate change and human activity. This study aims to address this gap by conducting an extensive global analysis of NDVI-LST correlations from 2000 to 2024, utilizing multi-source satellite data to assess latitudinal and ecosystem-specific variability. The MODIS dataset, which provides global daily LST data at a 1 km resolution from 2000 to 2024, was used alongside MODIS-derived NDVI data, which offers global vegetation indices at a 1 km resolution and 16-day temporal intervals. A correlation analysis was performed by extracting NDVI and LST values for each raster cell. The analysis revealed significant negative correlations in regions such as the western United States, Brazil, southern Africa, and northern Australia, where increased temperatures suppress vegetation activity. A total of 38,281,647 pixels, or 20% of the global map, exhibited statistically significant correlations, with 80.4% showing negative correlations, indicating a reduction in vegetation activity as temperatures rise. The latitudinal distribution of significant correlations revealed two prominent peaks: one in the tropical and subtropical regions of the Southern Hemisphere and another in the temperate zones of the Northern Hemisphere. This study uncovers notable spatial and latitudinal patterns in the LST-NDVI relationship, with most regions exhibiting negative correlations, underscoring the cooling effects of vegetation. These findings emphasize the crucial role of vegetation in regulating surface temperatures, providing valuable insights into ecosystem health, and informing conservation strategies in response to climate change. Full article

Large-scale and long-term landslide susceptibility assessments are crucial for revealing the patterns of landslide risk variation and for guiding the formulation of disaster prevention and mitigation policies at the national level. This study, through the establishment of a global dynamic landslide susceptibility model, uses the multi-dimensional analysis strategy and studies the development trend of China’s large-scale landslide susceptibility. First, a global landslide dataset consisting of 8023 large-scale landslide events triggered by rainfall and earthquakes between 2001 and 2020 was constructed based on the GEE (Google Earth Engine) platform. Secondly, a global dynamic landslide susceptibility model was developed using the ResNet18 (18-layer residual neural network) DL (deep learning) framework, incorporating both dynamic and static LCFs (landslide conditioning factors). The model was utilized to generate sequential large-scale landslide susceptibility maps for China from 2001 to 2022. Finally, the MK (Mann–Kendall) test was used to investigate the change trends in the large-scale landslide susceptibility of China. The results of the study are as follows. (1) The ResNet18 model outperformed SVMs (support vector machines) and CNNs (convolutional neural networks), with an AUC value of 0.9362. (2) SHAP (Shapley Additive Explanations) analyses revealed that precipitation played an important factor in the occurrence of landslides in China. In addition, profile curvature, NDVI, and distance to faults are thought to have a significant impact on landslide susceptibility. (3) The large-scale landslide susceptibility trends in China are complex and varied. Particular emphasis should be placed on Southwest China, including Chongqing, Guizhou, and Sichuan, which exhibit high landslide susceptibility and notable upward trends, and also consider Northwest China, including Shaanxi and Shanxi, which have high susceptibility but decreasing trends. These results provide valuable insights for disaster prevention and mitigation in China. Full article

High-resolution water depth data are of great significance in island research and coastal ecosystem monitoring. However, the acquisition of high-resolution imagery has been a challenge due to the difficulties and high costs associated with obtaining such data. To address this issue, this study proposes a water depth inversion method based on Gaofen-1 (GF-1) satellite data, which integrates multi-source satellite data to obtain high-resolution bathymetric data. Specifically, the research utilizes bathymetric data derived from Sentinel-2 and Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) as prior information, combined with high-resolution imagery obtained from the GF-1 satellite constellation (GF-1B/C/D). Then, it employs a scale-invariant effect to map bathymetry with a spatial resolution of 2 m, applied to four study areas in the Pacific Islands. The results are further evaluated using ICESat-2 data, which demonstrate that the water depth inversion results from this study possess high accuracy, with $R^2$ values exceeding 0.85, root mean square error (RMSE) ranging from 0.56 to 0.90 m, with an average of 0.7125 m, and mean absolute error (MAE) ranging from 0.43 to 0.76 m, with an average of 0.55 m. Additionally, this paper discusses the applicability of the scale-invariant assumption in this research and the improvements of the quadratic polynomial ratio model (QPRM) method compared to the classical linear ratio model (CLRM) method. The findings indicate that the integration of multi-source satellite remote sensing data based on the scale-invariant effect can effectively obtain high-precision, high-resolution bathymetric data, providing significant reference value for the application of GF-1 satellites in high-resolution bathymetry mapping. Full article

Fire risk prediction is of great importance for fire prevention. Fire risk maps are an effective tool to quantify regional fire risk. Most existing studies on forest fire risk maps mainly use a single machine learning model, but different models have varying degrees of feature extraction in the same spatial environment, leading to inconsistencies in prediction accuracy. To address this issue, this study proposes a novel integrated machine learning framework that systematically evaluates multiple models and combines their outputs through a weighted ensemble approach, thereby enhancing prediction robustness. During the feature selection stage, factors including socio-economic, climate, terrain, remote sensing data, and human factors were considered. Unlike previous studies that mainly use a single model, eight models were evaluated and compared using performance metrics. Three models were weighted based on Mean Squared Error (MSE) values, and cross-validation results showed an improvement in model performance. The integrated model achieved an accuracy of 0.8602, an area under the curve (AUC) of 0.772, and superior sensitivity (0.9234), outperforming individual models. Finally, the weighted framework was applied to generate a fire risk map. Compared with prior studies, this multi-model ensemble approach not only improves predictive accuracy but also provides a scalable and adaptable framework for fire risk mapping, and provides valuable insights to address future fire sustainability issues. Full article

The Albufera of Valencia (Spain) is a representative case of pressure on water quality, which caused the hypertrophic state of the lake to completely change the ecosystem that once featured crystal clear waters. PerúSAT-1 is the first Peruvian remote sensing satellite developed for natural disaster monitoring. Its high spatial resolution makes it an ideal sensor for capturing highly detailed products, which are useful for a variety of applications. The ability to change its acquisition geometry allows for an increase in revisit time. The main objective of this study is to assess the potential of PerúSAT-1′s multispectral images to develop multi-parameter algorithms to evaluate the ecological state of the Albufera lagoon. During five field campaigns, samples were taken, and measurements of ecological indicators (chlorophyll-a, Secchi disk depth, total suspended matter, and its organic-inorganic fraction) were made. All possible combinations of two bands were obtained and subsequently correlated with the biophysical variables by fitting a linear regression between the field data and the band combinations. The equations for estimating all the water variables result in the following R² values: 0.76 for chlorophyll-a (NRMSE: 16%), 0.75 for Secchi disk depth (NRMSE: 15%), 0.84 for total suspended matter (NRMSE: 11%), 0.76 for the inorganic fraction (NRMSE: 15%), and 0.87 for the organic fraction (NRMSE: 9%). Finally, the equations were applied to the Albufera lagoon images to obtain thematic maps for all variables. Full article

Studying natural hazards in the context of human-induced landscape transformation is complex, especially in regions with limited information. The narratives of the elderly can play a role in filling these knowledge gaps at the multi-decadal timescale. Here, we build upon a citizen-based elderly approach to understanding natural hazard patterns and landscape transformation in a tropical mountainous environment, the Kigezi Highlands (SW Uganda). We engaged 98 elderly citizens (>70 years old) living in eight small watersheds with different characteristics. Through interviews and focus group discussions, we reconstructed historical timelines and used participatory mapping to facilitate the interview process. We cross-checked the information of the elderly citizens with historical aerial photographs, archives, and field visits. Our results show that major land use/cover changes are associated with a high population increase over the last 80 years. We also evidence an increase in reported natural hazard events such as landslides and flash floods from the 1940s until the 1980s. Then, we notice a stabilization in the number of hazard events per decade, although the two most impacted decades (1980s and 2000s) stand out. Despite this new information, an increase in natural hazard frequency due to land use/cover change cannot yet be quantitatively validated, especially when the probable modulator effect of climate variability is considered. Nevertheless, the increase in the exposure of a vulnerable population to natural hazards is clear, and population growth together with poor landscape management practices are the key culprits that explain this evolution. This study demonstrates the added value of historical narratives in terms of understanding natural hazards in the context of environmental changes. This insight is essential for governments and non-governmental organizations for the development of policies and measures for disaster risk reduction that are grounded in the path dependence of local realities. Full article

Leaf chlorophyll content (LCC) is a key indicator of crop growth condition. Real-time, non-destructive, rapid, and accurate LCC monitoring is of paramount importance for precision agriculture management. This study proposes an improved method based on multi-source data, combining the Sentinel-2A spectral response function (SRF) and computer algorithms, to overcome the limitations of traditional methods. First, the equivalent remote sensing reflectance of Sentinel-2A was simulated by combining UAV hyperspectral images with ground experimental data. Then, using grey relational analysis (GRA) and the maximum information coefficient (MIC) algorithm, we explored the complex relationship between the vegetation indices (VIs) and LCC, and further selected feature variables. Meanwhile, we utilized three spectral indices (DSI, NDSI, RSI) to identify sensitive band combinations for LCC and further analyzed the response relationship of the original bands to LCC. On this basis, we selected three nonlinear machine learning models (XGBoost, RFR, SVR) and one multiple linear regression model (PLSR) to construct the LCC inversion model, and we chose the optimal model to generate spatial distribution maps of maize LCC at the regional scale. The results indicate that there is a significant nonlinear correlation between the VIs and LCC, with the XGBoost, RFR, and SVR models outperforming the PLSR model. Among them, the XGBoost_MIC model achieved the best LCC inversion results during the tasseling stage (VT) of maize growth. In the UAV hyperspectral data, the model achieved an R² = 0.962 and an RMSE = 5.590 mg/m² in the training set, and an R² = 0.582 and an RMSE = 6.019 mg/m² in the test set. For the Sentinel-2A-simulated spectral data, the training set had an R² = 0.923 and an RMSE = 8.097 mg/m², while the test set showed an R² = 0.837 and an RMSE = 3.250 mg/m², which indicates an improvement in test set accuracy. On a regional scale, the LCC inversion model also yielded good results (train R² = 0.76, test R² = 0.88, RMSE = 18.83 mg/m²). In conclusion, the method proposed in this study not only significantly improves the accuracy of traditional methods but also, with its outstanding versatility, can achieve rapid, non-destructive, and precise crop growth monitoring in different regions and for various crop types, demonstrating broad application prospects and significant practical value in precision agriculture. Full article

Small object detection in aerial imagery remains challenging due to sparse feature representation, limited spatial resolution, and complex background interference. Current deep learning approaches enhance detection performance through multi-scale feature fusion, leveraging convolutional operations to expand the receptive field or self-attention mechanisms for global context modeling. However, these methods primarily rely on spatial-domain features, while self-attention introduces high computational costs, and conventional fusion strategies (e.g., concatenation or addition) often result in weak feature correlation or boundary misalignment. To address these challenges, we propose a unified spatial-frequency modeling and multi-scale alignment fusion framework, termed USF-DETR, for small object detection. The framework comprises three key modules: the Spatial-Frequency Interaction Backbone (SFIB), the Dual Alignment and Balance Fusion FPN (DABF-FPN), and the Efficient Attention-AIFI (EA-AIFI). The SFIB integrates the Scharr operator for spatial edge and detail extraction and FFT/IFFT for capturing frequency-domain patterns, achieving a balanced fusion of global semantics and local details. The DABF-FPN employs bidirectional geometric alignment and adaptive attention to enhance the significance expression of the target area, suppress background noise, and improve feature asymmetry across scales. The EA-AIFI streamlines the Transformer attention mechanism by removing key-value interactions and encoding query relationships via linear projections, significantly boosting inference speed and contextual modeling. Experiments on the VisDrone and TinyPerson datasets demonstrate the effectiveness of USF-DETR, achieving improvements of 2.3% and 1.4% mAP over baselines, respectively, while balancing accuracy and computational efficiency. The framework outperforms state-of-the-art methods in small object detection. Full article

For the multi-UAV path planning problem, environmental modeling and an improved swarm intelligence-based optimization algorithm are discussed in this paper. Firstly, to align with reality, specific constraints of UAVs in motions, attitudes and altitudes, real-world threats such as radars and no-fly zones, and inter-UAV collisions are considered. Thus, multi-UAV path planning is transformed into a multi-objective constrained optimization problem. Accordingly, an improved nutcracker optimization algorithm is proposed to solve this problem. Through initializing with logistic chaotic mapping and the lens imaging inverse learning strategy, a more fit elite initialization population is produced to increase the efficiency of path planning. Furthermore, by adjusting adaptive parameters and integrating an improved sine-cosine search strategy, a balance between global exploration capability and local exploitation capability during path planning is achieved. Experimental results show that the improved nutcracker optimization algorithm surpasses other algorithms with respect to both convergence speed and convergence value, making it an effective method for multi-UAV path planning. Full article

Beach surface moisture (BSM) is crucial to studying coastal aeolian sand transport processes. However, traditional measurement techniques fail to accurately monitor moisture distribution with high spatiotemporal resolution. Remote sensing technologies have garnered widespread attention for providing rapid and non-contact moisture measurements, but a single method has inherent limitations. Passive remote sensing is challenged by complex beach illumination and sediment grain size variability. Active remote sensing represented by LiDAR (light detection and ranging) exhibits high sensitivity to moisture, but requires cumbersome intensity correction and may leave data holes in high-moisture areas. Using machine learning, this research proposes a BSM inversion method that fuses UAV (unmanned aerial vehicle) orthophoto brightness with intensity recorded by TLSs (terrestrial laser scanners). First, a back propagation (BP) network rapidly corrects original intensity with in situ scanning data. Second, beach sand grain size is estimated based on the characteristics of the grain size distribution. Then, by applying nearest point matching, intensity and brightness data are fused at the point cloud level. Finally, a new BP network coupled with the fusion data and grain size information enables automatic brightness correction and BSM inversion. A field experiment at Baicheng Beach in Xiamen, China, confirms that this multi-source data fusion strategy effectively integrates key features from diverse sources, enhancing the BP network predictive performance. This method demonstrates robust predictive accuracy in complex beach environments, with an RMSE of 2.63% across 40 samples, efficiently producing high-resolution BSM maps that offer values in studying aeolian sand transport mechanisms. Full article

Path planning is of great research significance as it is key to affecting the efficiency and safety of mobile robot autonomous navigation task execution. The traditional gray wolf optimization algorithm is widely used in the field of path planning due to its simple structure, few parameters, and easy implementation, but the algorithm still suffers from the disadvantages of slow convergence, ease of falling into the local optimum, and difficulty in effectively balancing exploration and exploitation in practical applications. For this reason, this paper proposes a multi-strategy improved gray wolf optimization algorithm (MSIAR-GWO) based on reinforcement learning. First, a nonlinear convergence factor is introduced, and intelligent parameter configuration is performed based on reinforcement learning to solve the problem of high randomness and over-reliance on empirical values in the parameter selection process to more effectively coordinate the balance between local and global search capabilities. Secondly, an adaptive position-update strategy based on detour foraging and dynamic weights is introduced to adjust the weights according to changes in the adaptability of the leadership roles, increasing the guiding role of the dominant individual and accelerating the overall convergence speed of the algorithm. Furthermore, an artificial rabbit optimization algorithm bypass foraging strategy, by adding Brownian motion and Levy flight perturbation, improves the convergence accuracy and global optimization-seeking ability of the algorithm when dealing with complex problems. Finally, the elimination and relocation strategy based on stochastic center-of-gravity dynamic reverse learning is introduced for the inferior individuals in the population, which effectively maintains the diversity of the population and improves the convergence speed of the algorithm while avoiding falling into the local optimal solution effectively. In order to verify the effectiveness of the MSIAR-GWO algorithm, it is compared with a variety of commonly used swarm intelligence optimization algorithms in benchmark test functions and raster maps of different complexities in comparison experiments, and the results show that the MSIAR-GWO shows excellent stability, higher solution accuracy, and faster convergence speed in the majority of the benchmark-test-function solving. In the path planning experiments, the MSIAR-GWO algorithm is able to plan shorter and smoother paths, which further proves that the algorithm has excellent optimization-seeking ability and robustness. Full article

Air temperature ($T_a$) measurements are crucial for characterizing phenomena like the urban heat island (UHI), which can create critical conditions in cities during summer. This study aims to develop a machine learning-based model, namely gradient boosting, to estimate $T_a$ from geostationary satellite LST data and to apply these estimates to investigate UHI dynamics. Using Rome, Italy, as a case study, the model was trained with $T_a$ data from 15 weather stations, taking multi-temporal LST values (instantaneous and lagged up to 4 h) and additional predictors. The model achieved an overall RMSE of 0.9 °C. The resulting $T_a$ fields, with a 3 km spatial and hourly temporal resolution, enabled a detailed analysis of UHI intensity and dynamics during the summers of 2019–2020, significantly enhancing the spatial and temporal detail compared to previous studies based solely on in situ data. The results also revealed a slightly higher nocturnal UHI intensity than previously reported, attributed to the inclusion of rural areas with near-zero imperviousness, thanks to the complete mapping of $T_a$ across the domain now accessible. Full article

An accurate and efficient detection method for harvesting is crucial for the development of automated harvesting robots in short-cycle, high-yield facility tomato cultivation environments. This study focuses on cherry tomatoes, which grow in clusters, and addresses the complexity and reduced detection speed associated with the current multi-step processes that combine target detection with segmentation and traditional image processing for clustered fruits. We propose YOLO-Picking Point (YOLO-PP), an improved cherry tomato picking point detection network designed to efficiently and accurately identify stem keypoints on embedded devices. YOLO-PP employs a C2FET module with an EfficientViT branch, utilizing parallel dual-path feature extraction to enhance detection performance in dense scenes. Additionally, we designed and implemented a Spatial Pyramid Squeeze Pooling (SPSP) module to extract fine features and capture multi-scale spatial information. Furthermore, a new loss function based on Inner-CIoU was developed specifically for keypoint tasks to further improve detection efficiency.The model was tested on a real greenhouse cherry tomato dataset, achieving an accuracy of 95.81%, a recall rate of 98.86%, and mean Average Precision ($mAP$) scores of 99.18% and 98.87% for mAP50 and mAP50-95, respectively. Compared to the DEKR, YOLO-Pose, and YOLOv8-Pose models, the mAP value of the YOLO-PP model improved by 16.94%, 10.83%, and 0.81%, respectively. The proposed algorithm has been implemented on NVIDIA Jetson edge computing devices, equipped with a human–computer interaction interface. The results demonstrate that the proposed Improved Picking Point Detection Network exhibits excellent performance and achieves real-time accurate detection of cherry tomato harvesting tasks in facility agriculture. Full article

Hydropower is a vital strategic component of China’s clean energy development. Its construction and optimized water resource allocation are crucial for addressing global energy challenges, promoting socio-economic development, and achieving sustainable development. However, the optimization scheduling of cascade hydropower stations is a large-scale, multi-constrained, and nonlinear problem. Traditional optimization methods suffer from low computational efficiency, while conventional intelligent algorithms still face issues like premature convergence and local optima, which severely hinder the full utilization of water resources. This study proposed an improved whale optimization algorithm, the Black-winged Differential-variant Whale Optimization Algorithm (BDWOA), which enhanced population diversity through a Logistic-Sine-Cosine combination chaotic map, improved algorithm flexibility with an adaptive adjustment strategy, and introduced the migration mechanism of the black-winged kite algorithm along with a differential mutation strategy to enhance the global search ability and convergence capacity. The BDWOA algorithm was tested using test functions with randomly generated simulated data, with its performance compared against five related optimization algorithms. Results indicate that the BDWOA achieved the optimal value with the fewest iterations, effectively overcoming the limitations of the original whale optimization algorithm. Further validation using actual runoff data for the cascade hydropower station optimization scheduling model showed that the BDWOA effectively enhanced power generation efficiency. In high-flow years, the average power generation increased by 8.3%, 6.5%, 6.8%, 4.1%, and 8.2% compared to the five algorithms while achieving the shortest computation time. Significant improvements in power generation were also observed in normal-flow and low-flow years. The scheduling solutions generated by the BDWOA can adapt to varying inflow conditions, offering an innovative approach to solving complex hydropower station optimization scheduling problems. This contributes to the sustainable utilization of water resources and supports the long-term development of renewable energy. Full article