Modeling global oceanic nitrogen deposition from food systems and its mitigation potential by reducing overuse of fertilizers.

Growing population and consumption pose unprecedented demands on food production. However, ammonia emissions mainly from food systems increase oceanic nitrogen deposition contributing to eutrophication. Here, we developed a long-term oceanic nitrogen deposition dataset (1970 to 2018) with updated ammonia emissions from food systems, evaluated the impact of ammonia emissions on oceanic nitrogen deposition patterns, and discussed the potential impact of nitrogen fertilizer overuse. Based on the chemical transport modeling approach, oceanic ammonia-related nitrogen deposition increased by 89% globally between 1970 and 2018, and now, it exceeds oxidized nitrogen deposition by over 20% in coastal regions including China Sea, India Coastal, and Northeastern Atlantic Shelves. Approximately 38% of agricultural nitrogen fertilizer was excessive, which corresponds to 15% of global oceanic ammonia-related nitrogen deposition. Policymakers and water quality managers need to pay increasingly more attention to ammonia associated with food production if the goal of reducing coastal nitrogen pollution is to be achieved for Sustainable Development Goals.

Exploring global changes in agricultural ammonia emissions and their contribution to nitrogen deposition since 1980.

SignificanceAgricultural systems are already major forces of ammonia pollution and environmental degradation. How agricultural ammonia emissions affect the spatio-temporal patterns of nitrogen deposition and where to target future mitigation efforts, remains poorly understood. We develop a substantially complete and coherent agricultural ammonia emissions dataset in nearly recent four decades, and evaluate the relative role of reduced nitrogen in total nitrogen deposition in a spatially explicit way. Global reduced nitrogen deposition has grown rapidly, and will occupy a greater dominant position in total nitrogen deposition without future ammonia regulations. Recognition of agricultural ammonia emissions on nitrogen deposition is critical to formulate effective policies to address ammonia related environmental challenges and protect ecosystems from excessive nitrogen inputs.

Global change progressively increases foliar nitrogen-phosphorus ratios in China's subtropical forests.

Globally increased nitrogen (N) to phosphorus (P) ratios (N/P) affect the structure and functioning of terrestrial ecosystems, but few studies have addressed the variation of foliar N/P over time in subtropical forests. Foliar N/P indicates N versus P limitation in terrestrial ecosystems. Quantifying long-term dynamics of foliar N/P and their potential drivers is crucial for predicting nutrient status and functioning in forest ecosystems under global change. We detected temporal trends of foliar N/P, quantitatively estimated their potential drivers and their interaction between plant types (evergreen vs. deciduous and trees vs. shrubs), using 1811 herbarium specimens of 12 widely distributed species collected during 1920-2010 across China's subtropical forests. We found significant decreases in foliar P concentrations (23.1%) and increases in foliar N/P (21.2%). Foliar N/P increased more in evergreen species (22.9%) than in deciduous species (16.9%). Changes in atmospheric CO2 concentrations ( P CO 2 $$ {\mathrm{P}}_{{\mathrm{CO}}_2} $$ ), atmospheric N deposition and mean annual temperature (MAT) dominantly contributed to the increased foliar N/P of evergreen species, while P CO 2 $$ {\mathrm{P}}_{{\mathrm{CO}}_2} $$ , MAT, and vapor pressure deficit, to that of deciduous species. Under future Shared Socioeconomic Pathway (SSP) scenarios, increasing MAT and P CO 2 $$ {\mathrm{P}}_{{\mathrm{CO}}_2} $$ would continuously increase more foliar N/P in deciduous species than in evergreen species, with more 12.9%, 17.7%, and 19.4% versus 6.1%, 7.9%, and 8.9% of magnitudes under the scenarios of SSP1-2.6, SSP3-7.0, and SSP5-8.5, respectively. The results suggest that global change has intensified and will progressively aggravate N-P imbalance, further altering community composition and ecosystem functioning of subtropical forests.

Nitrogen deposition accelerates soil carbon sequestration in tropical forests.

Terrestrial ecosystem carbon (C) sequestration plays an important role in ameliorating global climate change. While tropical forests exert a disproportionately large influence on global C cycling, there remains an open question on changes in below-ground soil C stocks with global increases in nitrogen (N) deposition, because N supply often does not constrain the growth of tropical forests. We quantified soil C sequestration through more than a decade of continuous N addition experiment in an N-rich primary tropical forest. Results showed that long-term N additions increased soil C stocks by 7 to 21%, mainly arising from decreased C output fluxes and physical protection mechanisms without changes in the chemical composition of organic matter. A meta-analysis further verified that soil C sequestration induced by excess N inputs is a general phenomenon in tropical forests. Notably, soil N sequestration can keep pace with soil C, based on consistent C/N ratios under N additions. These findings provide empirical evidence that below-ground C sequestration can be stimulated in mature tropical forests under excess N deposition, which has important implications for predicting future terrestrial sinks for both elevated anthropogenic CO2 and N deposition. We further developed a conceptual model hypothesis depicting how soil C sequestration happens under chronic N deposition in N-limited and N-rich ecosystems, suggesting a direction to incorporate N deposition and N cycling into terrestrial C cycle models to improve the predictability on C sink strength as enhanced N deposition spreads from temperate into tropical systems.

Nitrogen addition and mowing had only weak interactive effects on macronutrients in plant-soil systems of a typical steppe in Inner Mongolia.

Effective management of macronutrients is pivotal in the optimization and provisioning of ecosystem services in grassland areas, particularly in degraded grasslands. In such instances where mowing and nitrogen (N) fertilization have emerged as predominant management strategies, nutrient management is especially important. However, the precise effects of these concurrent practices on the distribution of macronutrients in plant-soil systems remain unclear. Here we evaluated the effects of 12 years of N addition (2, 10, and 50 g N m-2 year-1) and mowing on the concentrations and pools of six macronutrients (i.e., N; phosphorus P; sulfur S, calcium Ca, magnesium Mg, and potassium K) in three plant components (aboveground plants, litter, and belowground roots) at the community level and in the soil in a typical steppe in Inner Mongolia. Our results revealed that N addition generally raised the N concentration in the entire plant-soil system, regardless of whether plots were mowed. Higher N addition (10 and 50 g N m-2 year-1) also led to higher concentrations of P (+22%, averaging two N addition rates), S (+16%), K (+22%), Ca (+22%), and Mg (+24%) in plants but lower concentrations of these nutrients in the litter. Similar decreases in K (-9%), Ca (-46%), and Mg (-8%) were observed in the roots. In light of the observed increases in vegetation biomass and the lack of pronounced changes in soil bulk density, we found that the ecosystem N enrichment resulted in increased pools of all measured macronutrients in plants, litter, and roots (with the exception of Ca in the roots) while concurrently decreased the pools of P (-20%, averaging two higher N addition rates), S (-12%), K (-10%), Ca (-37%), and Mg (-19%) in the soil, with no obvious effect of the mowing practice. Overall, mowing exhibited a very limited capacity to alleviate the effects of long-term N addition on macronutrients in the plant-soil system. These findings highlight the importance of considering the distribution of macronutrients across distinct plant organs and the dynamic nutrient interplay between plants and soil, particularly in the context of long-term fertilization and mowing practices, when formulating effective grassland management strategies.

Drivers of foliar 15 N trends in southern China over the last century.

Foliar stable nitrogen (N) isotopes (δ15 N) generally reflect N availability to plants and have been used to infer about changes thereof. However, previous studies of temporal trends in foliar δ15 N have ignored the influence of confounding factors, leading to uncertainties on its indication to N availability. In this study, we measured foliar δ15 N of 1811 herbarium specimens from 12 plant species collected in southern China forests from 1920 to 2010. We explored how changes in atmospheric CO2 , N deposition and global warming have affected foliar δ15 N and N concentrations ([N]) and identified whether N availability decreased in southern China. Across all species, foliar δ15 N significantly decreased by 0.82‰ over the study period. However, foliar [N] did not decrease significantly, implying N homeostasis in forest trees in the region. The spatiotemporal patterns of foliar δ15 N were explained by mean annual temperature (MAT), atmospheric CO2 ( P CO 2 ), atmospheric N deposition, and foliar [N]. The spatiotemporal trends of foliar [N] were explained by MAT, temperature seasonality, P CO 2 , and N deposition. N deposition within the rates from 5.3 to 12.6 kg N ha-1  year-1 substantially contributed to the temporal decline in foliar δ15 N. The decline in foliar δ15 N was not accompanied by changes in foliar [N] and therefore does not necessarily reflect a decline in N availability. This is important to understand changes in N availability, which is essential to validate and parameterize biogeochemical cycles of N.

Latitudinal patterns of terrestrial phosphorus limitation over the globe.

Phosphorus limitation on terrestrial plant growth is being incorporated into Earth system models. The global pattern of terrestrial phosphorus limitation, however, remains unstudied. Here, we examined the global-scale latitudinal pattern of terrestrial phosphorus limitation by analysing a total of 1068 observations of aboveground plant production response to phosphorus additions at 351 forest, grassland or tundra sites that are distributed globally. The observed phosphorus-addition effect varied greatly (either positive or negative), depending significantly upon fertilisation regime and production measure, but did not change significantly with latitude. In contrast, phosphorus-addition effect standardised by fertilisation regime and production measure was consistently positive and decreased significantly with latitude. Latitudinal gradient in the standardised phosphorus-addition effect was explained by several mechanisms involving substrate age, climate, vegetation type, edaphic properties and biochemical machinery. This study suggests that latitudinal pattern of terrestrial phosphorus limitation is jointly shaped by macro-scale driving forces and the fundamental structure of life.

Plant acclimation to long-term high nitrogen deposition in an N-rich tropical forest.

Anthropogenic nitrogen (N) deposition has accelerated terrestrial N cycling at regional and global scales, causing nutrient imbalance in many natural and seminatural ecosystems. How added N affects ecosystems where N is already abundant, and how plants acclimate to chronic N deposition in such circumstances, remains poorly understood. Here, we conducted an experiment employing a decade of N additions to examine ecosystem responses and plant acclimation to added N in an N-rich tropical forest. We found that N additions accelerated soil acidification and reduced biologically available cations (especially Ca and Mg) in soils, but plants maintained foliar nutrient supply at least in part by increasing transpiration while decreasing soil water leaching below the rooting zone. We suggest a hypothesis that cation-deficient plants can adjust to elevated N deposition by increasing transpiration and thereby maintaining nutrient balance. This result suggests that long-term elevated N deposition can alter hydrological cycling in N-rich forest ecosystems.

Distinct storage mechanisms of soil organic carbon in coniferous forest and evergreen broadleaf forest in tropical China.

The impact of human activities on soil carbon (C) storage in tropical forests has aroused wide concern during the past decades, because these ecosystems play a key role in ameliorating global climate change. However, there remain uncertainties about how land-use history alters soil organic carbon (SOC) stability and storage in different forests. In this study, we measured the C content and mass distributions of soil aggregates, density fractions, mineral-bound C and microbial biomass C in the organic horizon, 0-10 cm and 10-20 cm soil layers in coniferous forest and evergreen broadleaf forest at Dinghushan Biosphere Reserve in tropical China. The broadleaf forest had larger SOC stocks than the coniferous forest, but the proportion of SOC stored in different density fractions at 0-10 cm soils was similar between forest types, while a greater proportion of SOC was stored in microaggregates in the coniferous forest. Most of the SOC was held as light fraction C in the organic horizon in the coniferous forest, whereas the concentrations of mineral-bound C were higher in the broadleaf forest. These findings indicate clear differences in the protection of SOC between broadleaf and coniferous forests growing on the same soil type. We propose that historic conversion of broadleaf forest to coniferous forest has reduced soil C sequestration capacity by altering the diversity and quality of plant inputs to the soil, which in turn affected macroaggregate formation, soil chemical properties and microbial biomass. Our results thus demonstrate that changes in forest tree species composition could have long-lasting effects on soil structure and carbon storage, providing crucial evidence for policy decisions on forest carbon sink management.

Species richness promotes ecosystem carbon storage: evidence from biodiversity-ecosystem functioning experiments.

Plant diversity has a strong impact on a plethora of ecosystem functions and services, especially ecosystem carbon (C) storage. However, the potential context-dependency of biodiversity effects across ecosystem types, environmental conditions and carbon pools remains largely unknown. In this study, we performed a meta-analysis by collecting data from 95 biodiversity-ecosystem functioning (BEF) studies across 60 sites to explore the effects of plant diversity on different C pools, including aboveground and belowground plant biomass, soil microbial biomass C and soil C content across different ecosystem types. The results showed that ecosystem C storage was significantly enhanced by plant diversity, with stronger effects on aboveground biomass than on soil C content. Moreover, the response magnitudes of ecosystem C storage increased with the level of species richness and experimental duration across all ecosystems. The effects of plant diversity were more pronounced in grasslands than in forests. Furthermore, the effects of plant diversity on belowground plant biomass increased with aridity index in grasslands and forests, suggesting that climate change might modulate biodiversity effects, which are stronger under wetter conditions but weaker under more arid conditions. Taken together, these results provide novel insights into the important role of plant diversity in ecosystem C storage across critical C pools, ecosystem types and environmental contexts.

Global response patterns of plant photosynthesis to nitrogen addition: A meta-analysis.

A mechanistic understanding of plant photosynthetic response is needed to reliably predict changes in terrestrial carbon (C) gain under conditions of chronically elevated atmospheric nitrogen (N) deposition. Here, using 2,683 observations from 240 journal articles, we conducted a global meta-analysis to reveal effects of N addition on 14 photosynthesis-related traits and affecting moderators. We found that across 320 terrestrial plant species, leaf N was enhanced comparably on mass basis (Nmass , +18.4%) and area basis (Narea , +14.3%), with no changes in specific leaf area or leaf mass per area. Total leaf area (TLA) was increased significantly, as indicated by the increases in total leaf biomass (+46.5%), leaf area per plant (+29.7%), and leaf area index (LAI, +24.4%). To a lesser extent than for TLA, N addition significantly enhanced leaf photosynthetic rate per area (Aarea , +12.6%), stomatal conductance (gs , +7.5%), and transpiration rate (E, +10.5%). The responses of Aarea were positively related with that of gs , with no changes in instantaneous water-use efficiency and only slight increases in long-term water-use efficiency (+2.5%) inferred from 13 C composition. The responses of traits depended on biological, experimental, and environmental moderators. As experimental duration and N load increased, the responses of LAI and Aarea diminished while that of E increased significantly. The observed patterns of increases in both TLA and E indicate that N deposition will increase the amount of water used by plants. Taken together, N deposition will enhance gross photosynthetic C gain of the terrestrial plants while increasing their water loss to the atmosphere, but the effects on C gain might diminish over time and that on plant water use would be amplified if N deposition persists.

Long-term nitrogen addition modifies microbial composition and functions for slow carbon cycling and increased sequestration in tropical forest soil.

Nitrogen (N) deposition is a component of global change that has considerable impact on belowground carbon (C) dynamics. Plant growth stimulation and alterations of fungal community composition and functions are the main mechanisms driving soil C gains following N deposition in N-limited temperate forests. In N-rich tropical forests, however, N deposition generally has minor effects on plant growth; consequently, C storage in soil may strongly depend on the microbial processes that drive litter and soil organic matter decomposition. Here, we investigated how microbial functions in old-growth tropical forest soil responded to 13 years of N addition at four rates: 0 (Control), 50 (Low-N), 100 (Medium-N), and 150 (High-N) kg N ha-1  year-1 . Soil organic carbon (SOC) content increased under High-N, corresponding to a 33% decrease in CO2 efflux, and reductions in relative abundances of bacteria as well as genes responsible for cellulose and chitin degradation. A 113% increase in N2 O emission was positively correlated with soil acidification and an increase in the relative abundances of denitrification genes (narG and norB). Soil acidification induced by N addition decreased available P concentrations, and was associated with reductions in the relative abundance of phytase. The decreased relative abundance of bacteria and key functional gene groups for C degradation were related to slower SOC decomposition, indicating the key mechanisms driving SOC accumulation in the tropical forest soil subjected to High-N addition. However, changes in microbial functional groups associated with N and P cycling led to coincidentally large increases in N2 O emissions, and exacerbated soil P deficiency. These two factors partially offset the perceived beneficial effects of N addition on SOC storage in tropical forest soils. These findings suggest a potential to incorporate microbial community and functions into Earth system models considering their effects on greenhouse gas emission, biogeochemical processes, and biodiversity of tropical ecosystems.

Stoichiometry controls asymbiotic nitrogen fixation and its response to nitrogen inputs in a nitrogen-saturated forest.

Lowland tropical forests with chronic nitrogen (N) deposition and/or abundant N-fixing organisms are commonly rich in N relative to other nutrients. The tropical N richness introduces a paradoxical relationship in which many tropical forests sustain high rates of asymbiotic N fixation despite the soil N richness and the higher energy cost of N fixation than of soil N uptake. However, the mechanism underlying this phenomenon remains unclear. Our study aims to test this phenomenon and examine potential mechanisms of nutrient concentrations vs. substrate stoichiometry in regulating N fixation using multiple linear regression models. We hypothesized that the rates of asymbiotic N fixation would be low in an N-rich forest under N deposition and substrate stoichiometry would explain the variation in N fixation better than nutrient concentrations. We conducted a chronic N-addition experiment in an N-saturated tropical forest in southern China and measured the N fixation rates, carbon (C), N, and phosphorus (P) concentrations, and stoichiometry in different substrates (soil, forest floor, mosses, and canopy leaves). Total N fixation rates were high (10.35-12.43 kg N·ha-1 ·yr-1 ) in this N-saturated forest because of the high substrate C:N and N:P stoichiometry (which explained 13-52% of the variation in N fixation, P < 0.037) rather than substrate nutrient concentrations (P > 0.05). Atmospheric N deposition (34-50 kg N·ha-1 ·yr-1 ) failed to down-regulate asymbiotic N fixation in this forest possibly because the N deposition rate was insufficient to inhibit N fixation or N deposition maintained high N fixation rates by increasing C sequestration in the substrates. Our N-addition experiment showed the insensitivity of N fixation in all the tested substrates to low N addition (50 kg N·ha-1 ·yr-1 ); however, medium and high N addition (100-150 kg N·ha-1 ·yr-1 ) stimulated the moss and foliar N fixation because of the increases in substrate C:N stoichiometry (which explained 30-34% of the variation in N fixation, P < 0.001). Overall, our results emphasize the importance of substrate (particularly mosses and foliage) stoichiometry as a driver of asymbiotic N fixation and sustained N richness in lowland tropical forests.

Effects of climate on soil phosphorus cycle and availability in natural terrestrial ecosystems.

Climate is predicted to change over the 21st century. However, little is known about how climate change can affect soil phosphorus (P) cycle and availability in global terrestrial ecosystems, where P is a key limiting nutrient. With a global database of Hedley P fractions and key-associated physiochemical properties of 760 (seminatural) natural soils compiled from 96 published studies, this study evaluated how climate pattern affected soil P cycle and availability in global terrestrial ecosystems. Overall, soil available P, indexed by Hedley labile inorganic P fraction, significantly decreased with increasing mean annual temperature (MAT) and precipitation (MAP). Hypothesis-oriented path model analysis suggests that MAT negatively affected soil available P mainly by decreasing soil organic P and primary mineral P and increasing soil sand content. MAP negatively affected soil available P both directly and indirectly through decreasing soil primary mineral P; however, these negative effects were offset by the positive effects of MAP on soil organic P and fine soil particles, resulting in a relatively minor total MAP effect on soil available P. As aridity degree was mainly determined by MAP, aridity also had a relatively minor total effect on soil available P. These global patterns generally hold true irrespective of soil depth (≤10 cm or >10 cm) or site aridity index (≤1.0 or >1.0), and were also true for the low-sand (≤50%) soils. In contrast, available P of the high-sand (>50%) soils was positively affected by MAT and aridity and negatively affected by MAP. Our results suggest that temperature and precipitation have contrasting effects on soil P availability and can interact with soil particle size to control soil P availability.

High retention of 15 N-labeled nitrogen deposition in a nitrogen saturated old-growth tropical forest.

The effects of increased reactive nitrogen (N) deposition in forests depend largely on its fate in the ecosystems. However, our knowledge on the fates of deposited N in tropical forest ecosystems and its retention mechanisms is limited. Here, we report the results from the first whole ecosystem 15 N labeling experiment performed in a N-rich old-growth tropical forest in southern China. We added 15 N tracer monthly as 15 NH415 NO3 for 1 year to control plots and to N-fertilized plots (N-plots, receiving additions of 50 kg N ha-1  yr-1 for 10 years). Tracer recoveries in major ecosystem compartments were quantified 4 months after the last addition. Tracer recoveries in soil solution were monitored monthly to quantify leaching losses. Total tracer recovery in plant and soil (N retention) in the control plots was 72% and similar to those observed in temperate forests. The retention decreased to 52% in the N-plots. Soil was the dominant sink, retaining 37% and 28% of the labeled N input in the control and N-plots, respectively. Leaching below 20 cm was 50 kg N ha-1  yr-1 in the control plots and was close to the N input (51 kg N ha-1  yr-1 ), indicating N saturation of the top soil. Nitrogen addition increased N leaching to 73 kg N ha-1  yr-1 . However, of these only 7 and 23 kg N ha-1  yr-1 in the control and N-plots, respectively, originated from the labeled N input. Our findings indicate that deposited N, like in temperate forests, is largely incorporated into plant and soil pools in the short term, although the forest is N-saturated, but high cycling rates may later release the N for leaching and/or gaseous loss. Thus, N cycling rates rather than short-term N retention represent the main difference between temperate forests and the studied tropical forest.

Divergent responses of soil buffering capacity to long-term N deposition in three typical tropical forests with different land-use history.

Elevated anthropogenic nitrogen (N) deposition has become an important driver of soil acidification at both regional and global scales. It remains unclear, however, how long-term N deposition affects soil buffering capacity in tropical forest ecosystems and in ecosystems of contrasting land-use history. Here, we expand on a long-term N deposition experiment in three tropical forests that vary in land-use history (primary, secondary, and planted forests) in Southern China, with N addition as NH4NO3 of 0, 50, 100, and 150 kg N ha(-1) yr(-1), respectively. Results showed that all three forests were acid-sensitive ecosystems with poor soil buffering capacity, while the primary forest had higher base saturation and cation exchange capacity than others. However, long-term N addition significantly accelerated soil acidification and decreased soil buffering capacity in the primary forest, but not in the degraded secondary and planted forests. We suggest that ecosystem N status, influenced by different land-use history, is primarily responsible for these divergent responses. N-rich primary forests may be more sensitive to external N inputs than others with low N status, and should be given more attention under global changes in the future, because lack of nutrient cations is irreversible.

Nitrogen deposition contributes to soil acidification in tropical ecosystems.

Elevated anthropogenic nitrogen (N) deposition has greatly altered terrestrial ecosystem functioning, threatening ecosystem health via acidification and eutrophication in temperate and boreal forests across the northern hemisphere. However, response of forest soil acidification to N deposition has been less studied in humid tropics compared to other forest types. This study was designed to explore impacts of long-term N deposition on soil acidification processes in tropical forests. We have established a long-term N-deposition experiment in an N-rich lowland tropical forest of Southern China since 2002 with N addition as NH4 NO3 of 0, 50, 100 and 150 kg N ha(-1)  yr(-1) . We measured soil acidification status and element leaching in soil drainage solution after 6-year N addition. Results showed that our study site has been experiencing serious soil acidification and was quite acid-sensitive showing high acidification (pH(H2O) <4.0), negative water-extracted acid neutralizing capacity (ANC) and low base saturation (BS,< 8%) throughout soil profiles. Long-term N addition significantly accelerated soil acidification, leading to depleted base cations and decreased BS, and further lowered ANC. However, N addition did not alter exchangeable Al(3+) , but increased cation exchange capacity (CEC). Nitrogen addition-induced increase in SOC is suggested to contribute to both higher CEC and lower pH. We further found that increased N addition greatly decreased soil solution pH at 20 cm depth, but not at 40 cm. Furthermore, there was no evidence that Al(3+) was leaching out from the deeper soils. These unique responses in tropical climate likely resulted from: exchangeable H(+) dominating changes of soil cation pool, an exhausted base cation pool, N-addition stimulating SOC production, and N saturation. Our results suggest that long-term N addition can contribute measurably to soil acidification, and that shortage of Ca and Mg should receive more attention than soil exchangeable Al in tropical forests with elevated N deposition in the future.

Quality-Aware Selective Fusion Network for V-D-T Salient Object Detection.

Depth images and thermal images contain the spatial geometry information and surface temperature information, which can act as complementary information for the RGB modality. However, the quality of the depth and thermal images is often unreliable in some challenging scenarios, which will result in the performance degradation of the two-modal based salient object detection (SOD). Meanwhile, some researchers pay attention to the triple-modal SOD task, namely the visible-depth-thermal (VDT) SOD, where they attempt to explore the complementarity of the RGB image, the depth image, and the thermal image. However, existing triple-modal SOD methods fail to perceive the quality of depth maps and thermal images, which leads to performance degradation when dealing with scenes with low-quality depth and thermal images. Therefore, in this paper, we propose a quality-aware selective fusion network (QSF-Net) to conduct VDT salient object detection, which contains three subnets including the initial feature extraction subnet, the quality-aware region selection subnet, and the region-guided selective fusion subnet. Firstly, except for extracting features, the initial feature extraction subnet can generate a preliminary prediction map from each modality via a shrinkage pyramid architecture, which is equipped with the multi-scale fusion (MSF) module. Then, we design the weakly-supervised quality-aware region selection subnet to generate the quality-aware maps. Concretely, we first find the high-quality and low-quality regions by using the preliminary predictions, which further constitute the pseudo label that can be used to train this subnet. Finally, the region-guided selective fusion subnet purifies the initial features under the guidance of the quality-aware maps, and then fuses the triple-modal features and refines the edge details of prediction maps through the intra-modality and inter-modality attention (IIA) module and the edge refinement (ER) module, respectively. Extensive experiments are performed on VDT-2048 dataset, and the results show that our saliency model consistently outperforms 13 state-of-the-art methods with a large margin. Our code and results are available at https://github.com/Lx-Bao/QSFNet.

Missingness-Pattern-Adaptive Learning With Incomplete Data.

Many real-world problems deal with collections of data with missing values, e.g., RNA sequential analytics, image completion, video processing, etc. Usually, such missing data is a serious impediment to a good learning achievement. Existing methods tend to use a universal model for all incomplete data, resulting in a suboptimal model for each missingness pattern. In this paper, we present a general model for learning with incomplete data. The proposed model can be appropriately adjusted with different missingness patterns, alleviating competitions between data. Our model is based on observable features only, so it does not incur errors from data imputation. We further introduce a low-rank constraint to promote the generalization ability of our model. Analysis of the generalization error justifies our idea theoretically. In additional, a subgradient method is proposed to optimize our model with a proven convergence rate. Experiments on different types of data show that our method compares favorably with typical imputation strategies and other state-of-the-art models for incomplete data. More importantly, our method can be seamlessly incorporated into the neural networks with the best results achieved. The source code is released at https://github.com/YS-GONG/missingness-patterns.

Reformulating Graph Kernels for Self-Supervised Space-Time Correspondence Learning.

Self-supervised space-time correspondence learning utilizing unlabeled videos holds great potential in computer vision. Most existing methods rely on contrastive learning with mining negative samples or adapting reconstruction from the image domain, which requires dense affinity across multiple frames or optical flow constraints. Moreover, video correspondence prediction models need to uncover more inherent properties of the video, such as structural information. In this work, we propose HiGraph+, a sophisticated space-time correspondence framework based on learnable graph kernels. By treating videos as a spatial-temporal graph, the learning objective of HiGraph+ is issued in a self-supervised manner, predicting the unobserved hidden graph via graph kernel methods. First, we learn the structural consistency of sub-graphs in graph-level correspondence learning. Furthermore, we introduce a spatio-temporal hidden graph loss through contrastive learning that facilitates learning temporal coherence across frames of sub-graphs and spatial diversity within the same frame. Therefore, we can predict long-term correspondences and drive the hidden graph to acquire distinct local structural representations. Then, we learn a refined representation across frames on the node-level via a dense graph kernel. The structural and temporal consistency of the graph forms the self-supervision of model training. HiGraph+ achieves excellent performance and demonstrates robustness in benchmark tests involving object, semantic part, keypoint, and instance labeling propagation tasks. Our algorithm implementations have been made publicly available at https://github.com/zyqin19/HiGraph.