Search

Accurate predictions from deep neural networks are crucial for distinguishing clean data and correcting noisy labels in current label noise learning methods. However, the conventional Softmax classifier used in most relevant works is highly sensitive to label noise due to its inherent competition-prompting mechanism, i.e., similar categories are encouraged to compete for limited confidence scores during class activation, especially between the noisy classes and the ground-truth, which can inevitably lead to suboptimal predictions and eventually hamper model performance. To address this inter-class competition problem, we propose a novel Sigmoid-based Sample Selection and Correction method named SigCo for learning with noisy labels. Different from previous works, we develop a Sigmoid-based network in which each Sigmoid classifier independently predicts its respective class, improving the reliability of the selection and correction process through more accurate predictions. Besides, in order to mitigate the negative impact of noisy labels, we design a noise-adaptive learning strategy by imposing stringent class masking constraints on clean samples to enhance the learning of discriminative features, while adopting a loose masking strategy for noisy data to improve the robustness to label noise. Additionally, we introduce a co-training strategy between our Sigmoid-based network and the conventional Softmax-based network to implicitly boost the generalization capability of the model. Extensive experiments on synthetic and real-world benchmarks show that SigCo consistently outperforms state-of-the-art methods. Especially on CIFAR-100N with 80% and 90% symmetric noise ratios, it improves test accuracy by 5.10% and 18.44%, respectively. [ABSTRACT FROM AUTHOR]

Shrub encroachment, characterized by the proliferation of shrubs into grasslands, is a challenge faced by grasslands worldwide that significantly impacts livestock production and ecosystem functions. Rapid and accurate estimation of shrub dominance is important for understanding changes in plant community structures and formulating grassland management policies. However, the limited spatial resolution of commonly used satellite imagery poses a challenge when estimating shrub dominance at the landscape scale. The rapid development of Unoccupied Aerial Vehicles (UAVs) has opened up new opportunities for cross-scale observations of shrub encroachment in grasslands by effectively bridging the scale gap between ground sampling and satellite image pixels while reducing the required groundwork. This study utilized ground reference data, UAV data (RGB, hyperspectral, and LiDAR), and satellite data (Sentinel-1 and Sentinel-2) to estimate shrub and total above-ground biomass (AGB) in temperate grasslands to map the shrub dominance. First, UAV data were applied at the plot scale for the classification of shrub and herbaceous vegetation using the maximum entropy model (MaxEnt), estimation of shrub AGB by employing the vegetation index weighted canopy volume model (CVM VI), and estimation of herbaceous AGB based on the partial least squares regression (PLSR). Second, UAV AGB mapping results were upscaled as samples at the landscape scale and integrated with satellite imagery to establish the shrub and total AGB models using the extreme gradient boosting (XGBoost). Finally, shrub dominance, represented as shrub AGB/total AGB, was mapped across the study area. We found that at the plot scale, the MaxEnt model achieved an overall accuracy of 0.990 for object-based classification. The CVM VI combined with canopy height model and narrow-band vegetation index achieved the highest accuracy for estimating shrub AGB (R2 = 0.821, RMSE = 30.1 g). The PLSR combined with features derived from all UAV data achieved the highest accuracy for estimating herbaceous AGB (R2 = 0.856, RMSE = 9.1 g/m2). At the landscape scale, the XGBoost achieved high accuracy for estimating both the shrub AGB (R2 = 0.719, RMSE = 4.2 g/m2) and total AGB (R2 = 0.961, RMSE = 5.0 g/m2). The high-precision mapping results further facilitate the generation of shrub dominance maps at a landscape scale. This study presents a more accurate and efficient framework for mapping shrub AGB, total AGB, and shrub dominance using multi-scale remote sensing data, which offers new approaches for large-scale grassland AGB mapping and monitoring of shrub encroachment in grasslands. • Multisensory UAV data were evaluated for shrub and herbaceous AGB estimation at the plot scale. • SAR and multispectral satellite data were used to map AGB and shrub dominance at the landscape scale. • Our multi-scale framework outperformed the traditional AGB estimation method. • Assessing shrub dominance is important for understanding shrub encroachment. [ABSTRACT FROM AUTHOR]

Aboveground biomass (AGB) in grasslands directly reflects the net primary productivity, making it a sensitive indicator of grassland resource quality and ecological degradation. Accurately estimating AGB over large regions to reveal long-term AGB evolution trends remains a formidable challenge. In this study, we divided Inner Mongolia Autonomous Region (IMAR) grasslands into three study regions based on their spatial distribution of grassland types. We combined remote sensing data with ground-based sample data collected over the past 19 years from 6114 field plots using the Google Earth Engine platform. We constructed random forest (RF) and traditional regression AGB inversion models for each region and selected the best-performing model through accuracy assessment to estimate IMAR grassland AGB for the period 2000–2022. We also examined the trends in AGB changes and identified the driving forces affecting IMAR grasslands through the application of Theil-Sen estimation, Mann-Kendall trend analysis, and the Geodetector model. The main findings are as follows: (1) Compared with the univariate parametric traditional regression model, the AGB monitoring accuracy of the multivariate non-parametric RF model in the three study regions increased by 5.94%, 5.08% and 19.14%, respectively. (2) The average AGB per unit area of IMAR grasslands from 2000 to 2022 was 731.41 kg/hm2, with alpine meadow having the highest average AGB (1271.70 kg/hm2) and temperate grassland desertification having the lowest (469.06 kg/hm2). IMAR grasslands exhibited an overall increasing trend in AGB over the past 23 years (6.01 kg/hm2•yr), with the increasing trend covering 83.52% of the grassland area and the decreasing trend covering 16.48%. (3) Spatially, IMAR grassland AGB showed a gradual decline from northeast to southwest and exhibited an increasing trend with increasing longitude (45.423 kg/hm2 per degree) and latitude (71.9 kg/hm2 per degree). (4) Meteorological factors were the most significant factors affecting IMAR grassland AGB, with precipitation (five-year average q value of 0.61) being the most prominent. In the western part of IMAR, where precipitation is consistently limited throughout the year, the primary drivers of influence were human activities, with particular emphasis on the number of livestock (with a five-year average q value of 0.44). It is evident that reducing human activity disturbance and pressure in fragile grassland areas or implementing near-natural restoration measures will be beneficial for the sustainable development of grassland ecosystems. The results of this research hold substantial reference importance for the protection and restoration of grasslands, the supervision and administration of grassland resources, as well as the development of policies related to grassland management. [Display omitted] • The aboveground biomass (AGB) of grasslands in IMAR increased over the past 23 years. • IMAR grassland AGB shows an upward trend with increasing longitude and latitude. • Precipitation is the main factor affecting grassland AGB in IMAR. • Human activities have a greater impact on the grassland AGB in western IMAR. [ABSTRACT FROM AUTHOR]

[Display omitted] • The northwest area of the BTSSCP suffered severe dust-affected zones based on the dust flow trajectory. • The ESs in severe dust-affected areas were relatively poor. • ESs and the dust flow trajectory were integrated to identify ecological governance zones. • Governance suggestions were proposed for the ecological management zone combined with the driving forces. Achieving ecological conservation and governance based on ecosystem services (ESs) can balance regional ecological protection and development, which is of great significance for sustainable development. Existing research has mainly focused on the extraction of priority ecological conservation areas, thereby neglecting to consider the classification and management of ecological engineering areas. To improve regional ecological governance, we combine ESs and dust movement trajectory, using the Beijing-Tianjin Sandstorm Source Control Project (BTSSCP) as the research object to generate watersheds' ecological governance zones. Six ESs were assessed to determine ecological zones. Trajectory simulation of dust movement, using the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model and potential source contribution function, was used to determined zones affected by dust. By combining these two aspects, five main ecological governance clusters were identified. We found that: (1) the area with low ESs in the northwest of the BTSSCP suffered from severe dust impacts; (2) slight dust-affected zones were mainly distributed in the eastern of the BTSSCP, and in the area, ESs gradually decreased from south to north; (3) ESs in the region was primarily influenced by natural factors and the influencing factors of ESs vary among different zones. These findings will not only contribute to the overall planning and governance of the BTSSCP, but can offer general guidance for the formulation of targeted ecological protection and restoration measures in other areas. [ABSTRACT FROM AUTHOR]