Machine learning is increasingly used to study vegetation growth, however, more often than not, predicting and simulating functions are prioritized over quantitative estimates of the drivers of vegetation growth such as climate. In this paper, we, for the first time, systematically investigate the model bias in calculating factor importance of climate on vegetation growth, especially when various kinds of machine learning models are considered. We undertake two case studies to simulate research in remote sensing and ground-based scenarios from which the difference in quantitative relationships between climate and vegetation is evaluated across multiple models. We found that model complexity increased the determination coefficient (R2) but reduced the absolute importance of the preselected independent variables. As the fitting accuracy increases, the absolute factor importance of dominant factor and all the other influencing factors decreases simultaneously, and factor importance calculated by different models tended to be more normally distributed across the study region. The reduction in factor importance was accompanied by the increased effect of model selection; e.g. the model that was used to estimate vegetation growth played a larger role in producing the factor importance (shown by variance analysis, remote sensing scenario, F-statistic = 555.2; ground based scenario, F-statistic = 30.8) than the climate variables (variance analysis, remote sensing scenario, F = 460.8; ground based scenario, F = 28.8). Critically, for those machine learning models with highest fitting accuracy, the resultant factor importance of climate factors had smaller difference with that of random factor. In contrast, the relative factor influence among the selected climate factor is more robust and reliable (in variation analysis, model was detected no significant impact on the resultant factor importance). For 5 of 8 models, the dominant factor (temperature) has relative influence over 0.85, ranging from 0.88 to 0.99. According to the relevant result, we suggest testing the stability of factor contribution in future studies, particularly when using machine learning models in ecological research and dealing with numerous factors, before drawing relative conclusions. The balance between simple and accurate models is contested and we believe that our study will contribute to a better understand of the data behind this debate.

Forest growth in the majority of northern China is currently limited by drought and low nitrogen (N) availability. Drought events with increasing intensity have threatened multiple ecosystem services provided by forests. Whether N addition will have a detrimental or beneficial moderation effect on forest resistance and recovery to drought events was unclear. Here, our study focuses on Pinus tabulaeformis, which is the main plantation forest species in northern China. We investigated the role of climate change and N addition in driving multi-year tree growth with an 8-year soil nitrogen fertilization experiment and analyzing 184 tree ring series. A moderate drought event occurred during the experiment, providing an opportunity for us to explore the effects of drought and N addition on tree resistance and recovery. We found that N addition was beneficial for increasing the resistance of middle-aged trees, but had no effect on mature trees. The recovery of trees weakened significantly with increasing N addition, and the reduction in fine root biomass caused by multiyear N addition was a key influencing factor limiting recovery after moderate drought. Our study implies that the combined effect of increasing drought and N deposition might increase the risk of pine forest mortality in northern China.

As more machine learning and deep learning models are applied in studying the quantitative relationship between the climate and terrestrial vegetation growth, the uncertainty of these advanced models requires clarification. Partial dependence plots (PDPs) are one of the most widely used methods to estimate the marginal effect of independent variables on the predicted outcome of a machine learning model, and it is regarded as the main basis for conclusions in relevant research. As more controversies regarding the reliability of the results of the PDPs emerge, the uncertainty of the PDPs remains unclear. In this paper, we experiment with real, remote sensing data to systematically analyze the uncertainty of partial dependence relationships between four climate variables (temperature, rainfall, radiation, and windspeed) and vegetation growth, with one conventional linear model and six machine learning models. We tested the uncertainty of the PDP curves across different machine learning models from three aspects: variation, whole linear trends, and the trait of change points. Results show that the PDP of the dominant climate factor (mean air temperature) and vegetation growth parameter (indicated by the normalized difference vegetation index, NDVI) has the smallest relative variation and the whole linear trend of the PDP was comparatively stable across the different models. The mean relative variation of change points across the partial dependence curves of the non-dominant climate factors (i.e. radiation, windspeed, and rainfall) and vegetation growth ranged from 8.96% to 23.8%, respectively, which was much higher than those of the dominant climate factor and vegetation growth. Lastly, the model used for creating the PDP, rather than the relative importance of these climate factors, determines the fluctuation of the PDP output of these climate variables and vegetation growth. These findings have significant implications for using remote sensing data and machine learning models to investigate the quantitative relationships between the climate and terrestrial vegetation.

Due to the influence of climate change and extensive grazing, a large proportion of steppe grassland has been degraded worldwide. The Chinese government initiated a series of grassland restoration programs to reverse the degradation. However, the limiting factors and the restoration potential remain unknown. Here we present a process-based model to assess the restoration gap (RG) defined as maximum biomass differences between non-degraded and degraded grasslands with different degrees of soil and vegetation degradation. The process-based model Agricultural Production Systems Simulator (APSIM) was evaluated utilizing observation data from both typical and meadow steppes under natural conditions in terms of phenology, dynamics of above-ground biomass and soil water content. Scenario analysis and sensitivity analysis were subsequently performed to address the RG and controlling factors during 1969-2018. The results showed that the calibrated model performed well with r > 0.75 and model efficiency factor EF > 0.5 for all the simulation components. According to our model results, the RG was larger in typical steppe compared to that of meadow steppe and it increased with increasing soil and/or vegetation degradation, to ~60% under extremely degraded scenarios. Both soil and vegetation degradation led to reduced water use efficiency, with an elevated proportion of soil evaporation to evapotranspiration (Es/ET), however, the limiting factor for RG varied. The degradation of soil water holding capacity contributed more to RG regardless of climate conditions for typical steppe in all years and for meadow steppe in dry years. In wet years the importance of vegetation coverage reduction increased for RG in meadow steppe, where the relative importance of vegetation coverage (valued at 62.8%) was 25.6% higher than that of soil degradation. Our results demonstrated the importance of considering climate variations when developing protection and restoration programs for grassland ecosystems.

AbstractThe fate of organic carbon deposited in floodplain sediments is an important control on the magnitude and direction of the carbon flux from anthropogenically accelerated erosion and channelization of the riverine network. Globally, carbon deposition rates and mean residence time (MRT) within different geomorphic settings remains poorly constrained. We sampled soil profiles to 0.8 m depth from two geomorphic zones: active channel belt (ACB) and lowland floodplain, under long‐term pasture adjacent to the river Culm in SW England, UK. We evaluated sedimentation rates and carbon storage using fallout radionuclide 137Cs, particle size and total carbon analyses. Variation in decomposition was assessed via empirical (soil aggregate size, density fractionation combined with natural abundance 13C analysis) and modelling simulation (using the RothC model and catchment implications explored using a floodplain evolution model). Sedimentation and carbon accumulation rates were 5–6 times greater in the ACB than the floodplain. Carbon decomposition rates also varied with geomorphic setting. In floodplain cores, faster decomposition rates were indicated by greater 13C‐enrichment and subsoils dominated by mineral‐associated soil organic carbon. Whereas, in the ACB, carbon was less processed and 13C‐depleted, with light fraction and macroaggregate‐carbon throughout the cores, and RothC modelled decomposition rates were 4‐fold less than lowland floodplain cores. Including the ACB in floodplain carbon MRT calculations increased overall MRT by 10%. The major differences in the balance of sedimentation and decomposition rates between active and inactive floodplains suggests the relative extent of these contrasting zones is critical to the overall carbon balance. Restoration projects could enhance soil carbon storage by maximizing active floodplain areas by increasing river channel complexity.

The extent to which soil erosion is a net source or sink of carbon globally remains unresolved but has the potential to play a key role in determining the magnitude of CO2 emissions from land-use change in rapidly eroding landscapes. The effects of soil erosion on carbon storage in low-input agricultural systems, in acknowledged global soil erosion hotspots in developing countries, are especially poorly understood. Working in one such hotspot, the Indian Himalaya, we measured and modelled field-scale soil budgets, to quantify erosion-induced changes in soil carbon storage. In addition, we used long-term (1-year) incubations of separate and mixed soil horizons to better understand the mechanisms controlling erosion-induced changes in soil carbon cycling. We demonstrate that high rates of soil erosion did not promote a net carbon loss to the atmosphere at the field scale. Furthermore, our experiments showed that rates of decomposition in the organic matter-rich subsoil layers in depositional areas were lower per unit of soil carbon than from other landscape positions; however, these rates could be increased by mixing with topsoils. The results indicate that, the burial of soil carbon, and separation from fresh carbon inputs, led to reduced rates of decomposition offsetting potential carbon losses during soil erosion and transport within the cultivated fields. We conclude that the high rates of erosion experienced in these Himalayan soils do not, in isolation, drive substantial emissions of organic carbon, and there is the potential to promote carbon storage through sustainable agricultural practice.

The area of karst terrain in China covers 3.63×106 km2, with more than 40% in the southwestern region over the Guizhou Plateau. Karst comprises exposed carbonate bedrock over approximately 1.30×106 km2 of this area, which suffers from soil degradation and poor crop yield. This paper aims to gain a better understanding of the environmental controls on crop yield in order to enable more sustainable use of natural resources for food production and development. More precisely, four kinds of artificial neural network were used to analyse and simulate the spatial patterns of crop yield for seven crop species grown in Guizhou Province, exploring the relationships with meteorological, soil, irrigation and fertilization factors. The results of spatial classification showed that most regions of high-level crop yield per area and total crop yield are located in the central-north area of Guizhou. Moreover, the three artificial neural networks used to simulate the spatial patterns of crop yield all demonstrated a good correlation coefficient between simulated and true yield. However, the Back Propagation network had the best performance based on both accuracy and runtime. Among the 13 influencing factors investigated, temperature (16.4%), radiation (15.3%), soil moisture (13.5%), fertilization of N (13.5%) and P (12.4%) had the largest contribution to crop yield spatial distribution. These results suggest that neural networks have potential application in identifying environmental controls on crop yield and in modelling spatial patterns of crop yield, which could enable local stakeholders to realize sustainable development and crop production goals.

Increasing temperature over recent decades is expected to positively impact tree growth in humid regions. However, high stand density could increase the negative effects of warming-induced drought through inter-tree competition. How neighborhood competition impacts tree growth responding to climate change remains unclear. Here, we utilized the Changbai Mountain region in northeastern Asia as our study area. We quantified individual tree growth using tree-ring samples collected from three dominant tree species growing in three forest stand density levels. We estimated the effects of climate warming and forest stand density on growth processes and tested for a species-specific response to climate. Our results demonstrated that overall 25% of Korean pine, but only ~3% of Mongolian oak and ~ 4% of Manchurian ash experienced growth reduction. Increased forest density can also exacerbate growth reduction. We identified a climate turning point in 1984, where warming rapidly increased, and defined two groups, “enhance group” (EG) and “decline group” (DG), according to the individual tree growth trend after 1984. For the EG, climate warming increased temperature sensitivity, but the temperature sensitivity declined with increasing stand density for the whole study period. For the DG, tree growth sensitivity shifted from temperature to precipitation after 1984, driven by increased competition pressure under climate warming. Our study concludes that growth decline from warming-induced drought might be amplified by high forest stand density, was especially pronounced in conifer trees.

In this paper, cross-spectrum analysis was used to verify the agreement of periodicity between the global LAI (leaf area index) and climate factors. The results demonstrated that the LAI of deciduous forests and permanent wetlands have high agreement with temperature, rainfall and radiation over annual cycles. A low agreement between the LAI and seasonal climate variables was observed for some of the temperate and tropical vegetation types including shrublands and evergreen broadleaf forests, possibly due to the diversity of vegetation and human activities. Across all vegetation types, the LAI demonstrated a large time lag following variation in radiation (>1 month), whereas relatively short lag periods were observed between the LAI and annual temperature (around 2 weeks)/rainfall patterns (less than 10 days), suggesting that the impact of radiation on global vegetation growth is relatively slow, which is in accord with the results of previous studies. This work can provide a benchmark of the phenological drivers in global vegetation, from the perspective of periodicity, as well as helping to parameterize and refine the DGVMs (Dynamic Global Vegetation Models) for different vegetation types.

Floodplain soils provide an important link in the land-ocean aquatic continuum. Understanding microbial activity in these soils, which can be many metres deep, is a key component in our understanding of the role of floodplains in the carbon (C) cycle. We sampled the mineral soil profile to 3 m depth from two floodplain sites under long-term pasture adjacent to the river Culm in SW England, UK. Soil chemistry (C, nitrogen (N), phosphorus (P), soil microbial biomass (SMB), moisture content) and soil solution (pH, dissolved organic C (DOC) and N, nitrate, ammonium, water extractable P) were analysed over the 3 m depth in 6 increments: 0.0–0.2, 0.2–0.7, 1.0–1.5, 1.5–2.0, 2.0–2.5, and 2.5–3.0 m. 14C-glucose was added to the soil and the evolution of 14CO2 measured during a 29 d incubation. From soil properties and 14C-glucose mineralisation, three depth groups emerged, with distinct turnover times extrapolated from initial k1 mineralisation rate constants of 2 h (topsoil 0.0–0.2 m), 4 h (subsoil 0.2–0.7 m), and 11 h (deep subsoil 1.0–3.0 m). However, when normalised by SMB, k1 rate constants had no significant differences across all depths. Deep subsoil had a 2 h lag to reach maximal 14CO2 production whereas the topsoil and subsoil (0.2–0.7 m) achieved maximum mineralisation rates immediately. SMB decreased with depth, but only to half of the surface population, with the proportion of SMB-C to total C increasing from 1% in topsoil to 15% in deep subsoil (>1.0 m). The relatively large SMB concentration and rapid mineralisation of 14C-glucose suggests that DOC turnover in deep soil horizons in floodplains is limited by access to biologically available C and not the size of the microbial population.