Next Article in Journal
Erratum: Raney, R.K. Hybrid Dual-Polarization Synthetic Aperture Radar. Remote Sens. 2019, 11, 1521
Previous Article in Journal
Mapping Human Settlements with Higher Accuracy and Less Volunteer Efforts by Combining Crowdsourcing and Deep Learning
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 11
Issue 15
10.3390/rs11151800
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Predicting the Spatial Distribution and Severity of Soil Erosion in the Global Tropics using Satellite Remote Sensing

by
Tor-Gunnar Vågen
*,† and
Leigh Ann Winowiecki
†
World Agroforestry Centre (ICRAF), P.O. Box 30677-00100 Nairobi, Kenya
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Remote Sens. 2019, 11(15), 1800; https://0-doi-org.brum.beds.ac.uk/10.3390/rs11151800
Submission received: 27 May 2019 / Revised: 18 July 2019 / Accepted: 26 July 2019 / Published: 31 July 2019
(This article belongs to the Section Environmental Remote Sensing)
Browse Figures
Versions Notes

Abstract

:

Simple Summary

This research paper presents the first consistent spatial assessments of soil erosion prevalence based on Earth observation data and systematic field surveys to train models that predict soil erosion prevalence across the global tropics. The results of the study have major implications for efforts to identify land degradation hotspots and design interventions to restore degraded lands given the importance of soil erosion as a major process of land degradation globally. The paper also presents analysis of changes in soil erosion prevalence over the 15 year period from 2002 to 2017. The results of the study can also be used to better understand ecosystem health and resilience in the global tropics, for assessments of food security, and the implications of land degradation on climate change.

Abstract

Soil erosion has long been recognized as a major process of land degradation globally, affecting millions of hectares of land in the tropics and resulting in losses in productivity and biodiversity, decreased resilience of both marine and terrestrial ecosystems, and increased vulnerability to climate change. This paper presents an assessment of the extent of soil erosion in the global tropics at a moderate spatial resolution (500 m) based on a combination of systematic field surveys using the Land Degradation Surveillance Framework (LDSF) methodology and Earth observation data from the Moderate Resolution Imaging Spectroradiometer (MODIS) platform. The highest erosion prevalence was observed in wooded grassland, bushland, and shrubland systems in semi-arid areas, while the lowest occurrence was observed in forests. Observed erosion decreased with increasing fractional vegetation cover, but with high rates of erosion even at 50–60% fractional cover. These findings indicate that methods to assess soil erosion need to be able to detect erosion under relatively dense vegetation cover. Model performance was good for prediction of erosion based on MODIS, with high accuracy (~89% for detection) and high overall precision (AUC = 0.97). The spatial predictions from this study will allow for better targeting of interventions to restore degraded land and are also important for assessing the dynamics of land health indicators such as soil organic carbon. Given the importance of soil erosion for land degradation and that the methodology gives robust results that can be rapidly replicated at scale, we would argue that soil erosion should be included as a key indicator in international conventions such as the United Nations Convention to Combat Desertification.

1. Introduction

Soil loss through erosion is widely recognized as one of the most important and widespread forms of land degradation globally [1,2,3] and was reported as being a major threat to soil productivity as early as in the 1930s [4]. While studies have assessed global soil erosivity [5], potential soil loss [6], and soil erosion potential [7], the spatial extent and severity of soil erosion has not yet been consistently quantified. Estimates of both spatial and temporal dynamics of soil erosion are needed to improve progress towards the implementation of measures to reduce soil erosion. Efforts to reduce soil erosion have been largely ineffective to date despite the widespread recognition that it is an increasing problem [8,9,10,11]. This is particularly true in developing countries where population growth, agricultural expansion and mining, coupled with scarcity of land [12] and adverse economic and/or political policies [13] have resulted in there being little if any progress towards reducing land degradation through erosion. Also, up-to-date and spatially explicit estimates of soil erosion prevalence are needed for more effective targeting and scaling of efforts to avoid future land degradation. The importance of soil erosion goes beyond its widely recognized effects on agricultural productivity, with studies showing that the potential impacts of erosion on the global carbon budget, and, hence, climate change, can be significant [14,15]. These factors point to the need for improved methods to spatially assess the prevalence of soil erosion at scales relevant to management interventions, including the tracking of the occurrence and severity of erosion in landscapes over time.
Human activities are currently having profound impacts on the environment with some studies suggesting that human-driven processes of land degradation are threatening global food security [16] by impacting on the adaptive capacity and resilience of ecosystems worldwide, reducing their ability to sustain productivity and to provide other critical ecosystem services. Continued conversion of natural ecosystems into agricultural land is also leading to loss of habitats and biodiversity [17,18], and often to further increases in land degradation and the loss of critical ecosystem functions for regulation of water and nutrient cycling. Land degradation can also lead to increased risks to human health through the emergence and spread of infectious diseases [19], for example in deforested areas or in areas where wetlands are degraded resulting in sharp increases in malaria risk [20].
Although attempts have been made to assess the extent of global land degradation, either through expert opinion such as the Global Assessment of Land Degradation (GLASOD) [21,22], by using proxies such as the Normalized Difference Vegetation Index (NDVI) [23] or rain-use efficiency (RUE) [24], or through the use of empirical models such as the Revised Universal Soil Loss Equation (RUSLE) [25], estimates remain crude at best. In some cases, such assessments (e.g., the use of NDVI as a proxy for land degradation) can potentially be misleading, resulting in the implementation of costly interventions that are not suitable for the restoration of degraded ecosystems. Reference [7] applied the RUSLE empirical model to predict global soil erosion using a combination of factors derived from: (1) the Shuttle Radar Topography Mission (SRTM) and ASTER GDEM digital elevation models (DEMs) for slope length and steepness (LS factor) and (2) the WorldClim database [26] for precipitation and temperature data to estimate erosivity (R factor), the MODIS Vegetation Continuous Fields product (VCF, MOD44B) as a proxy for vegetation (C factor), and World Soil Information (ISRIC) SoilGrids [27] for soil organic carbon to estimate soil erodibility (K factor). They produced a global map of the potential rates of soil erosion at 250 m spatial resolution, although the actual spatial resolution is likely to be lower due to the lower spatial resolution of some of the input parameters used such as WorldClim, which has a spatial resolution of 1 km by 1 km. Also, this study did not include validations against observed rates of soil erosion for the years modeled (2001 and 2012; [7]), which is challenging due to a general lack of field-based monitoring of soil erosion rates.
We present a study where we use Earth observation to detect and map the occurrence of soil erosion across the global tropics, defined here as the part of the Earth found within the boundaries of the parallels of 40° south and 40° north. The main objective of the study was to explore the application of Earth observation, which has been shown in the past to be promising at more local scales [28,29], to provide rapid assessments of soil erosion for spatially distributed monitoring. We combined systematically collected field observations of erosion with Earth observation data to predict the spatial and temporal distribution of soil erosion prevalence at moderate spatial resolution (500 m) using MODIS data for the years 2002, 2007, 2012, and 2017. The spatial assessments of erosion provide estimates of land degradation hotspots and can be combined with other indicators of ecosystem health, including social factors, to better assess and identify drivers of land degradation and target land management interventions to reverse degradation.
We also explored interactions between vegetation cover and soil erosion prevalence, as well as erosion prevalence by geographical regions and Holdridge life zones in the tropics [30]. While the mapping of Holdridge life zone regions globally remain coarse, they do present a useful framework within which processes of, for example, land degradation, soil condition, hydrological functioning, and climate change can be assessed, particularly in the context of ecological resilience and for understanding interactions between inherent properties of soils and landscapes and human influences. This study was part of wider efforts to collect critical datasets on ecosystem health that are consistent and can be applied to assess the state of ecosystems worldwide at multiple spatial scales, moving beyond simple descriptive analytics of land degradation processes such as soil erosion towards diagnostics and predictive analytics to help target interventions that avoid or reverse land degradation.

2. Materials and Methods

This study used a combination of systematically sampled and characterized sites from across a range of different tropical ecosystems and remote sensing or Earth observation. As described in more detail below, field observations of soil erosion were used to train machine learning algorithms (models) to predict the occurrence of soil erosion based on moderate resolution satellite imagery over a period of 15 years from 2002 to 2017. In order to achieve consistent estimates of erosion, the timing (year) of field observations were matched with annual composites of satellite image reflectance data and the predictive models were tested using an independent dataset.

2.1. Field Data Collection

The field data used in this study were collected using the Land Degradation Surveillance Framework (LDSF) [28,29,31] over the period 2005 to 2017 (Table 1). Data collection was conducted as part of a range of different projects coordinated by the authors. Observations of the presence or absence of visible erosion patterns (i.e., mass/gully, rill, and sheet erosion) from 171 one hundred km2 LDSF sentinel sites were included in the study, each site consisting of 16 one-km2 spatially stratified sampling clusters with 10 one-thousand-m2 plots per cluster (Figure 1). Within each plot (N = 26,091), four subplots with an area of 100 m2 were assessed for visible signs of erosion (see also Reference [32]). Erosion prevalence was scored at the plot level by summing up the number of subplots with visible signs of erosion, with 0 being no observed erosion and 4 being erosion observed in all four subplots. We took a cut-off at three subplots or more (>50%) to represent “severe” soil erosion and used this in the modeling of erosion prevalence. Vegetation structure was classified at the plot level as part of the LDSF surveys, using the Land Cover Classification System (LCCS) [33].

2.2. Processing of Remote Sensing Data

Remote sensing data from the Bidirectional Reflection Distribution Function (BRDF)-corrected MODIS (MCD43A4) product was applied in the current study, following Reference [29]. To obtain consistent annual composite reflectance values, we calculated the Soil Adjusted Total Vegetation Index (SATVI) [34] for each daily MCD43A4 image. Short-wave infrared (SWIR) bands have been found to be sensitive to both green and senescent vegetation [35], unlike the NDVI, which uses the near-infrared band instead of SWIR. Annual reflectance composites were created for each year between 2002 and 2017 based on the date of SATVImax in each MODIS pixel.
The pre-processing of the MODIS Earth observation data was conducted in Python, using the Google Earth Engine (GEE) [36] Python API. Processing steps included removal of clouds and water bodies (using the MODIS water mask product (MOD44W)). We also excluded areas with elevation values higher than 4300 m based on elevation data from the Shuttle Radar Topography Mission (SRTM) and excluded hyper-arid areas (deserts) from our analysis based on Tropical Rainfall Monitoring Mission (TRMM) [37] calibrated precipitation data for the period 1998 to 2014. Each annual MODIS composite was matched to the year that each LDSF site was surveyed (Table 2).

2.3. Prediction Model for Mapping of Soil Erosion

Erosion prevalence was modeled in R Statistics [38] using a decision-tree approach known as Random Forests (RF) with 100 decision trees implemented as part of the ranger R library [39] to generate a classification model for soil erosion with MODIS reflectance bands 2 through 7 as independent variables in the model. In other words, only spectral data were used in the prediction model without other covariates meaning that a relatively low amount of input data is needed once the RF model has been trained. Model parameters were optimized or tuned using cross-validation as implemented in the trainControl function of the caret R library [40], followed by a tuning grid using the expand.grid function of the caret R library. The RF algorithm is a machine learning technique where multiple decision trees are constructed using bootstrap sampling of a training dataset and the trees in this ensemble are combined (bagged). This class of models was introduced by Reference [41] and has since been applied across a wide range of disciplines.
We used 70% of the LDSF plots to train the random forest (RF) model (N = 18,261). The remaining 30% of the LDSF plots (N = 7830) were used for model testing. Model performance was assessed by measuring the percentage of correctly classified test instances relative to observed instances, expressed as a confusion matrix [42]. We also calculated the Receiver Operating Characteristic curve (ROC) [43,44], which evaluates the accuracy of a model by considering errors that are either false positives or false negatives. A true positive fraction (TPF) is calculated as T P F c = P { M > c | D = 1 } , where M is the probability of erosion in our case, which we defined as positive if it exceeded a fixed threshold c, while D was the binary outcome (erosion = yes/no). A false positive fraction (FPF) is then the probability that we falsely predict the occurrence of erosion where there is none. This is calculated as F P F c = P { M > c | D = 0 } . The ROC curve is a graphical representation of the TPF against the FPF for all possible values of c (0 to 1). After assessing the various accuracy metrics of the RF model, we applied it to annual composite MODIS imagery to predict probabilities of soil erosion for the global tropics for the years 2002, 2007, 2012, and 2017, respectively.

2.4. Assessment of Soil Erosion by Vegetation Cover, Holdridge Life Zone, Continent, and Sub-Continent

In order to better understand the variations in predicted erosion prevalence we randomly sampled the erosion prevalence map for 2017 using 1 million randomly generated points. We extracted the predicted prevalence of erosion for 2017 for each random point, along with the Holdridge life zone [45], continent, and sub-continent (as defined by the United Nations Statistics Division). The extracted values were then stored in a database and used to assess erosion across Holdridge life zones, continents, and sub-continents.
The Holdridge life zone system has a spatial resolution of one-half degree latitude and longitude and consists of 38 classes. The rationale for including this in our analysis of predicted soil erosion prevalence was that the system has been shown to reflect tropical vegetation zones well through the inclusion of annual precipitation (mm), biotemperature (humidity provinces), and potential evapotranspiration (PET) ratio as barycentric subdivisions. Hence, assessing soil erosion prevalence across these zones is useful in understanding bioclimatic factors that may drive soil erosion in tropical ecosystems.

3. Results

The LDSF data included in this study represented a range of different climate zones and ecosystems, as illustrated by the distribution of LDSF plots across various LCCS vegetation structure classes shown in Table 2 and Figure 2. The most dominant vegetation structure class was cropland (34%), followed by grassland (15%), woodland (12%), and forest and shrubland (10% each; Table 2).
As far as land cover is concerned, forest ecosystems had the highest fractional vegetation cover (i.e., SATVImax) on average as expected. Woodlands, shrublands, croplands, and grasslands showed a wide range of SATVImax values due to the high degree of mixture between senescent and green vegetation, as well as bare plots. Observed erosion prevalence also ranged widely for the sites included in our study, with a relative frequency of severe erosion of about 46% on average for the whole dataset.

3.1. Prediction Model Performance

Parameter tuning for the RF model showed optimal mtry and nodesize values of 1.5 and 5, respectively. Calibration model performance for mapping of erosion based on MODIS was good, with an out-of-bag (OOB) prediction error of 23.8% and an overall accuracy of the calibration model of over 90%. As shown in Figure 3 (Panel 1), the area under the ROC curve (AUC) was 0.99 for the calibration dataset and 0.96 when we tested this model on the 30% of LDSF plots that were held out for model validation or testing. Both the calibration and validation models were able to separate eroded and non-eroded plots well (Figure 3, Panel 2).
Model accuracy for the detection of erosion was about 86%, while accuracy for non-detection was somewhat higher at about 91%, on average (Figure 3, Panel 3). The model also showed good agreement with K >= 0.6, indicating good precision overall. These results show that we are able to predict the prevalence of soil erosion with good accuracy and prediction based on MODIS reflectance data. Calculations of variable importance in the calibration model, indicates that the red (3), NIR (4) and SWIR (6 and 7) bands are the most important variables for prediction of soil erosion from MODIS. The importance of these bands is likely due to the combination of their sensitivity to soil cover (NIR and red in particular), and senescent vegetation and bare ground (SWIR bands and NIR) [46].
Model performance was comparable to that reported by Reference [28] when applying similar modeling approaches to the local mapping of soil erosion based on Landsat in a case study from Ethiopia.

3.2. Mapping Erosion Prevalence in the Global Tropics

Given the high levels of accuracy, we applied the RF model to the MODIS image library of annual reflectance composites for 2002, 2007, 2012, and 2017, producing spatial maps of predicted erosion prevalence at a 500 m resolution for each year covering the global tropics (Figure 4).
Our model estimates for 2017 showed a median of about 47% erosion when deserts were excluded. Erosion hotspots globally include semi-arid ecosystems in Kenya, Ethiopia, Somalia, southern Africa, western Africa, Australia, and Argentina, as well as fragile ecosystems in the Andes and the Tibetan plateau. Figure 5 shows a summary of predicted erosion prevalence rates by continent and subcontinent in the global tropics, based on a set of randomly sampled plots (N = 391,042 after excluding hyper-arid areas). We did not include Oceania and Europe in this analysis since we do not have observations of erosion on these continents, although the map in Figure 4 does extend to these continents. Median erosion prevalence was highest in southern Africa, followed by northern Africa, eastern Africa, and western Africa (Figure 5). However, as Figure 5 shows, there is a lot of variability in erosion prevalence within continents. Below we discuss some examples from contrasting parts of the tropics, including predictions for 2002 to 2017.

4. Discussion

Earth observation offers opportunities for mapping and monitoring of soil erosion in landscapes that are unique in many ways. By utilizing information about the land surface captured across multiple parts of the electromagnetic spectrum, satellite data can be used to derive predictive models of often complex processes and patterns by using field observations to train these models. One such complex process is soil erosion, which can occur across a wide range of topographic, geomorphological, and land cover conditions. A key advantage of Earth observation in this regard is that we are able to derive accurate predictions of the severity and spatial extent of soil erosion using only information from the spectral bands of the sensor, which means that no additional covariates that may be time consuming to collect, expensive, or simply unavailable, are needed. This means that we can achieve results with unprecedented levels of consistency. The use of remote sensing in land surface process models not only provides useful information for numerical modelling but can also reduce the risk of propagating errors [47,48] from the derivation of covariate layers. An example where such errors may propagate is in the use of empirical models for predicting soil erosion such as the RUSLE which relies on factors or proxies derived from layers that may have large uncertainties associated with their estimation.

4.1. Land Cover Effects on Soil Erosion

The role of vegetation cover in controlling rates of soil erosion by protecting soils from the impacts of rain drops (and wind) is widely accepted [16]. This is also reflected in the observations of erosion in this study, with the lowest observed erosion frequencies in forest ecosystems and thickets (Figure 6) where vegetation cover tends to be dense. Reference [49] reported similar results with lower rates of soil erosion under native forest vegetation than conventional agriculture based on compiled data from a range of different studies. We also observed high erosion rates in wooded grasslands, shrublands, and bushlands. Many of these ecosystems can have relatively high vegetation cover, but with high rates of observed erosion. Observed signs of erosion in croplands were somewhat lower than one might expect at about 42% of the surveyed plots. This could be the result of a number of factors such as plowing of fields, which may result in erosion patterns not being detected in cropland, although this was observed in a limited number of plots and was unlikely to have influenced much the prediction results. Also, the median slope for cropland plots was about 7% (4°), which could be another factor resulting in an erosion being observed in a moderate number of plots.
To examine this further, we binned the SATVImax values into 10 categories and assessed these against the probability of erosion in each plot (Figure 7). As we see from Figure 7, the highest rates of erosion were observed where plots had SATVImax values less than 10%, as expected, while there was a high level of variability at between 10% and 50% fractional vegetation cover. Similarly, for higher than 50% fractional cover, erosion prevalence rates decreased, but again with high levels of variability until it reached 70% or higher fractional cover (Figure 7), beyond which erosion rates dropped strongly. This confirms that erosion prevalence can be high even when vegetation cover is relatively dense, such as in areas where invasive species are inhibiting the growth of native grasses or herbs in the understory due to allelopathic effects [50], resulting in low herbaceous cover despite high woody cover. These effects are often exacerbated due to the ability of many invasive species to thrive in areas that are degraded, often outcompeting native species in terms of access to available soil nutrients and/or water [51]. A concrete example of ecosystems where this occurs include areas with high prevalence of invasive woody species such as Prosopis juliflora in the degraded drylands of East Africa.
Given the above, methods used to detect soil erosion from remote sensing need to be able to accurately predict its occurrence under varying levels of vegetation cover as we may observe SATVImax values higher than 50%, but with rates of erosion higher than 60%, for example. Another implication of these findings is that land cover alone is not an adequate indicator or soil erosion, or land degradation. There appear to be thresholds in fractional vegetation cover beyond which erosion reduces drastically, which will have implications for ecological regime shifts that can lead to a loss in ecosystem resilience and determine “flipping point(s)” [52], where an ecosystem enters into an alternative stable state [53] (e.g., through increased soil erosion as a response to vegetation cover loss). The predictions (maps) presented here can help in determining flipping points both in terms of thresholds beyond which soil erosion severity may lead to accelerated land degradation and in terms of determining temporal trends in erosion prevalence. Such thresholds should be studied in more detail in future studies, particularly in terms of the more general patterns presented here as compared to local processes that lead to increased soil erosion. By combining these maps and temporal trends with information on changes in factors such as land cover or land use, one can also assess the drivers of increased soil erosion that may lead to ecosystems flipping. Determining such flipping points is of critical importance for management, particularly in the case of land degradation, as the new (degraded) state may also be highly resilient, making restoration of the system difficult [51,54]. Future studies will need to investigate these relationships further, particularly interactions between woody and herbaceous cover fractions.

4.2. Assessing Changes in Soil Erosion Prevalence over Time

Closer examinations of the predictions presented for 2002, 2007, 2012, and 2017 for East Africa and India show varying patterns and trends. In East Africa 8, Laikipia, Meru, Isiolo, and Samburu counties of Kenya show moderate to high erosion prevalence, but with an increase in areas with high erosion prevalence (>75%, Figure 8, panel below maps) between 2002 and 2017, particularly in dry-lands. The majority of the areas shown in Figure 8 are rangeland systems. Other studies from eastern and southern Africa show that over-grazing and poor rangeland management have resulted in reduced grassland productivity [55,56], often leading to severe soil erosion and further declines in productivity in these ecosystems. Bush encroachment is a problem in some of the most severely eroded areas in Figure 8, exacerbating the situation further by leading to reduced herbaceous cover and increasing erosion. The drivers of soil erosion and land degradation in these systems are often complex, and may include increased population densities, demographic changes, and land tenure relations that restrict movement of pastoralists, and in some areas the introduction of unsustainable agricultural practices. Further, interactions with climate change have been reported in a number of studies, including more aggressive rainfall regimes that are likely to have important impacts on the rates of soil erosion [57,58,59].
Figure 9 shows estimates of erosion prevalence for the Indian state of Karnataka. There appears to be less soil erosion in 2017 relative to 2002, in many parts of the state (see panel below maps in Figure 9). Assessments of actual trends will need to be made, but the findings presented here are consistent with studies conducted in this part of India, including Reference [60] which reported reductions in soil erosion in Karnataka, which were generally consistent with the estimates of our study (Figure 9). Assessments of soil erosion for the Kurnool watershed in 2006/2007 reported that 68% of this watershed was eroded, while in a more recent project report, extensive soil and water conservation, including agroforestry, were reported in Kushtagi taluk [61].

4.2.1. Erosion by Holdridge Life Zones

Estimated rates of erosion by Holdridge life zones showed the highest predicted erosion prevalence values in desert scrub and “tropical thorn woodland” and “very dry forest” (Figure 10). The latter two life zones broadly correspond to what we label shrubland, (drier) wooded grassland, and grassland systems, and were also consistent with our field observations of erosion (Figure 6). These results point to the vulnerability of drier ecosystems in terms of land degradation [62] due to the presence of soil erosion because of high rainfall intensities, as discussed earlier for East Africa (Figure 8). If we consider the increasing human influence on these systems [63] due to the facts of population growth and agricultural expansion, coupled with climate change and increasingly erratic and, in many cases, rising levels of rainfall intensity [59], improved land management to reduce soil erosion is particularly important in these systems. Spatially explicit assessments of land degradation in dryland ecosystems are of critical importance for targeting interventions to restore degraded lands and monitor changes, given their vast spatial extent and the number of people that depend on them for their livelihoods. As expected, the lowest predicted erosion prevalence was found in moist and wet ecosystems, such as in rain forests (Figure 10).
The results of this study show that the Holdridge life zones provide a useful framework for assessing soil erosion risk since they integrate multiple bioclimatic factors, and these can be applied in analyses to better understand both risk as well as constraint envelopes for restoration of degraded ecosystems. Such assessments will be particularly important in the context of future climate scenarios and their potential influence on the different bioclimatic factors.

4.2.2. Limitations of the Presented Approach

While we have shown that the approach to mapping soil erosion in the current study can produce results with high levels of accuracy, there are limitations to the use of Earth observation for assessing these types of processes. A key limitation in our opinion is in assessing drivers of these processes, which will require a combination of both social and ecological data that can be challenging to detect remotely. Another limitation lies in unpacking the often complex biophysical interactions that lead to systems becoming more prone to erosion and land degradation in general, including the adaptive capacity of ecosystems. Finally, the predictions presented here represent estimates at a moderate spatial resolution. This is highly useful for assessments at regional, country or global scales, but has limitations in terms of applications for the restoration of degraded land where information at more detailed spatial scales will be needed. The use of higher resolution Earth observation data such as Landsat and Sentinel 2 will need to be explored.

5. Conclusions

Soil erosion continues to adversely affect millions of hectares of land in the global tropics, resulting in losses in productivity and increased food insecurity, decreased resilience of ecosystems, and increased vulnerability to climate change. Its spatial extent is generally not well understood and interventions to limit erosion have had limited success due to the lack of appropriately targeted interventions, hampering progress towards preventing further land degradation. We used systematic field surveys based on a spatially balanced sampling design, analysis of Earth observation data and ensemble modeling techniques to map soil erosion prevalence at high levels of accuracy across a range of tropical ecosystems. The high model accuracy warranted the use of these models for estimating erosion prevalence in the global tropics based on satellite imagery from the MODIS platform.
The median erosion prevalence in 2017 was estimated at 47% when desert ecosystems were excluded, with hot-spots particularly in semi-arid ecosystems. The highest occurrences of erosion overall were found in southern and northern Africa with 70% and 60% predicted erosion, respectively. Median erosion prevalence in Eastern Africa was estimated at 54%, while for Western Africa, including the Sahel, it was 51% and for southern Asia 50%. The lowest median estimated erosion prevalence by subcontinent was found in Central Asia (29%) and south-eastern Asia (30%). Our predictions showed reductions in erosion in parts of India over the period 2002 to 2017 that can potentially be attributed to extensive soil and water conservation activities. In contrast, dryland systems in Kenya showed an increase in the severity and extent of erosion over the same period. Erosion was found to be high even under relatively high fractional vegetation cover in some cases, which has important implications for methods to assess soil erosion and land degradation in general.
Our study has important implications for management of ecosystem health in landscapes, specifically for the quantification and spatial prediction of soil erosion prevalence and the targeting of interventions to control erosion and reverse land degradation. Soil erosion is currently not featured as part of the indicator sets included in major global initiatives to tackle land degradation, such as the Bonn Challenge and targets set by the Parties of the UN Convention to Combat Desertification (UNCCD) to achieve Land Degradation Neutrality (LDN). Given the accuracy of the results presented in this study for detecting and mapping soil erosion based on Earth observation, soil erosion should be included as a key indicator in these and other international conventions on land degradation and desertification.

Author Contributions

L.A.W. has been leading projects and field surveys that contributed to this paper since 2008, and also provided inputs into the statistical modeling of the results of these surveys and the writing of the paper. T.-G.V. co-designed the LDSF methodology and has been coordinating a number of surveys since 2005. He conducted the analysis of the earth observation data that were used in this study, contributed to the statistical modeling of the data and led the writing of this paper.

Funding

This research was funded by International Fund for Agricultural Development, grant number 20000001302; Bill and Melinda Gates Foundation, grant number 51353; and the CGIAR research programme on Forests, Trees and Agroforestry (FTA).

Acknowledgments

This paper builds on research coordinated by the authors and carried out in a number of projects. We would particularly like to acknowledge the CGIAR research programme on Forests, Trees and Agroforestry (FTA) for supporting the study through the funding of LDSF field surveys as part of the Sentinel Landscape initiative. We also acknowledge the Bill and Melinda Gates Foundation (BMGF) for partial funding as part of the Africa Soil Information Service (AfSIS) project, as well as Wajibu MS and the Northern Rangelands Trust (NRT) for funding data collection in Kenya, the CGIAR Challenge Programme on Water and Food (CPWF) for supporting sampling in Ethiopia and the International Fund for Agricultural Development (IFAD) for funding data collection and analysis.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Bennett, H.H. Soil changes due to erosion. In Proceedings of the Soil Science Society, 1939; pp. 399–401. Available online: https://dl.sciencesocieties.org/publications/sssaj/abstracts/4/C/SS00400C0399 (accessed on 20 March 2019).
  2. Carter, L. Soil Erosion: The Problem Persists Despite the Billions Spent on It. Science 1977, 196, 409–411. [Google Scholar] [CrossRef] [PubMed]
  3. Lal, R. Soil erosion and carbon dynamics. Soil Tillage Res. 2005, 81, 137–142. [Google Scholar] [CrossRef]
  4. Champion, A.M. Soil Erosion in Africa. Geogr. J. 1933, 82, 130–139. [Google Scholar] [CrossRef]
  5. Panagos, P.; Borrelli, P.; Meusburger, K.; Yu, B.; Klik, A.; Lim, K.J.; Yang, J.E.; Ni, J.; Miao, C.; Chattopadhyay, N.; et al. Global rainfall erosivity assessment based on high-temporal resolution rainfall records. Sci. Rep. 2017, 7, 1–12. [Google Scholar] [CrossRef] [Green Version]
  6. Yang, D.; Kanae, S.; Oki, T.; Koike, T.; Musiake, K. Global potential soil erosion with reference to land use and climate changes. Hydrol. Process. 2003, 17, 2913–2928. [Google Scholar] [CrossRef]
  7. Borrelli, P.; Robinson, D.A.; Fleischer, L.R.; Lugato, E.; Ballabio, C.; Alewell, C.; Meusburger, K.; Modugno, S.; Schütt, B.; Ferro, V.; et al. An assessment of the global impact of 21st century land use change on soil erosion. Nat. Commun. 2017, 8, 2013. [Google Scholar] [CrossRef]
  8. Chase, S. Rich Land—Poor Land; McGraw-Hill: New York, NY, USA, 1936; p. 350. [Google Scholar]
  9. Morgan, R. Soil Erosion; Longman: London, UK, 1979; p. 113. [Google Scholar]
  10. Lal, R. Soil degradation by erosion. Land Degrad. Dev. 2001, 12, 519–539. [Google Scholar] [CrossRef]
  11. Bibby, J. R.P.C. Morgan Soil Erosion and Conservation. Longman Group UK Ltd., Harlow, Essex. x 298 pp. Clay Miner. 1987, 22, 246. [Google Scholar] [CrossRef]
  12. Jayne, T.S.; Chamberlin, J.; Headey, D.D. Land pressures, the evolution of farming systems, and development strategies in Africa: A synthesis. Food Policy 2014, 48, 1–17. [Google Scholar] [CrossRef] [Green Version]
  13. Phalan, B.; Bertzky, M.; Butchart, S.H.M.; Donald, P.F.; Scharlemann, J.P.W.; Stattersfield, A.J.; Balmford, A. Crop Expansion and Conservation Priorities in Tropical Countries. PLoS ONE 2013, 8, e51759. [Google Scholar] [CrossRef]
  14. Lal, R. Soil erosion and the global carbon budget. Environ. Int. 2003, 29, 437–450. [Google Scholar] [CrossRef]
  15. Van Oost, K.; Quine, T.A.; Govers, G.; De Gryze, S.; Six, J.; Harden, J.W.; Ritchie, J.C.; McCarty, G.W.; Heckrath, G.; Kosmas, C.; et al. The Impact of Agricultural Soil Erosion on the Global Carbon Cycle. Science 2007, 318, 626–629. [Google Scholar] [CrossRef] [PubMed]
  16. Pimentel, D.; Burgess, M. Soil Erosion Threatens Food Production. Agriculture 2013, 3, 443–463. [Google Scholar] [CrossRef] [Green Version]
  17. Pereira, H.M.; Leadley, P.W.; Proenca, V.; Alkemade, R.; Scharlemann, J.P.W.; Fernandez-Manjarres, J.F.; Araujo, M.B.; Balvanera, P.; Biggs, R.; Cheung, W.W.L.; et al. Scenarios for Global Biodiversity in the 21st Century. Science 2010, 330, 1496–1501. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  18. Brien, T.G.O.; Baillie, J.E.M.; Krueger, L.; Cuke, M.; O’Brien, T.G. The Wildlife Picture Index: Monitoring top trophic levels. Anim. Conserv. 2010, 13, 335–343. [Google Scholar] [CrossRef]
  19. Patz, J.A.; Daszak, P.; Tabor, G.M.; Aguirre, A.A.; Pearl, M.; Epstein, J.; Wolfe, N.D.; Kilpatrick, A.M.; Foufopoulos, J.; Molyneux, D.; et al. Unhealthy Landscapes: Policy Recommendations on Land Use Change and Infectious Disease Emergence. Environ. Health Perspect. 2004, 112, 1092–1098. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  20. Dale, P.E.R.; Knight, J.M. Wetlands and mosquitoes: A review. Wetl. Ecol. Manag. 2008, 16, 255–276. [Google Scholar] [CrossRef]
  21. Oldeman, L. The Global Extent of Soil Degradation; International Soil Reference and Information Centre (ISRIC): Wageningen, The Netherlands, 1994; pp. 19–36. [Google Scholar]
  22. Oldeman, L.R.; Lynden, G.W. Revisiting the GLASOD methodology. In Methods for Assessment of Soil Degradation; International Soil Reference and Information Centre: Wageningen, The Netherlands, 1996. [Google Scholar]
  23. Tucker, C.J. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 1979, 8, 127–150. [Google Scholar] [CrossRef] [Green Version]
  24. Bai, Z.; Dent, D.; Olsson, L.; Schaepman, M. Proxy global assessment of land degradation. Soil Use Manag. 2008, 24, 223–234. [Google Scholar] [CrossRef]
  25. Renard, K.; Foster, G.; Weesies, G.A.; Porter, J.P. RUSLE: Revised universal soil loss equation. J. Soil Water Conserv. 1991, 46, 30–33. [Google Scholar]
  26. Hijmans, R.J.; Cameron, S.E.; Parra, J.L.; Jones, P.G.; Jarvis, A. Very high resolution interpolated climate surfaces for global land areas. Int. J. Climatol. 2005, 25, 1965–1978. [Google Scholar] [CrossRef]
  27. Hengl, T.; Mendes de Jesus, J.; Heuvelink, G.B.M.; Ruiperez Gonzalez, M.; Kilibarda, M.; Blagotić, A.; Shangguan, W.; Wright, M.N.; Geng, X.; Bauer-Marschallinger, B.; et al. SoilGrids250m: Global gridded soil information based on machine learning. PLoS ONE 2017, 12, e0169748. [Google Scholar] [CrossRef] [PubMed]
  28. Vågen, T.-G.; Winowiecki, L.A.; Abegaz, A.; Hadgu, K.M. Landsat-based approaches for mapping of land degradation prevalence and soil functional properties in Ethiopia. Remote Sens. Environ. 2013, 134, 266–275. [Google Scholar] [CrossRef]
  29. Vågen, T.-G.; Winowiecki, L.A.; Tondoh, J.E.; Desta, L.T.; Gumbricht, T. Mapping of soil properties and land degradation risk in Africa using MODIS reflectance. Geoderma 2016, 263, 216–225. [Google Scholar] [CrossRef] [Green Version]
  30. Holdridge, L.R. Life Zone Ecology; Tropical Science Center: San Jose, Costa Rica, 1967; p. 206. [Google Scholar]
  31. Vågen, T.-G.; Winowiecki, L.A.; Tamene Desta, L.; Tondoh, J.E. The Land Degradation Surveillance Framework (LDSF)—Field Guide; World Agroforestry Centre: Nairobi, Kenya, 2013; 14p. [Google Scholar]
  32. Vågen, T.-G.; Winowiecki, L.A.; Twine, W.; Vaughan, K. Spatial Gradients of Ecosystem Health Indicators across a Human-Impacted Semiarid Savanna. J. Environ. Qual. 2018, 47, 746. [Google Scholar] [CrossRef]
  33. Jansen, L.J.M.; Gregorio, A.D. Parametric land cover and land-use classifications as tools for environmental change detection. Agric. Ecosyst. Environ. 2002, 91, 89–100. [Google Scholar] [CrossRef]
  34. Hagen, S.C.; Heilman, P.; Marsett, R.; Torbick, N.; Salas, W.; van Ravensway, J.; Qi, J. Mapping Total Vegetation Cover Across Western Rangelands with Moderate-Resolution Imaging Spectroradiometer Data. Rangel. Ecol. Manag. 2012, 65, 456–467. [Google Scholar] [CrossRef]
  35. Marsett, R.C.; Qi, J.; Heilman, P.; Biedenbender, S.H.; Watson, M.C.; Amer, S.; Weltz, M.; Goodrich, D.; Marsett, R. Remote Sensing for Grassland Management in the Arid Southwest. Rangel. Ecol. Manag. 2006, 59, 530–540. [Google Scholar] [CrossRef]
  36. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens. Environ. 2017, 202, 18–27. [Google Scholar] [CrossRef]
  37. Tropical Rainfall Measuring Mission (TRMM). TRMM (TMPA/3B43) Rainfall Estimate L3 1 Month 0.25 Degree × 0.25 Degree V7. Available online: https://cmr.earthdata.nasa.gov/search/concepts/C1282032631-GES_DISC.html (accessed on 20 March 2019).
  38. R Core Team. R: A Language and Environment for Statistical Computing; R Foundation for Statistical Computing: Vienna, Austria, 2018. [Google Scholar]
  39. Wright, M.N.; Ziegler, A. {ranger}: A Fast Implementation of Random Forests for High Dimensional Data in {C++} and {R}. J. Stat. Softw. 2017, 77, 1–17. [Google Scholar] [CrossRef]
  40. Kuhn, M.; Wing, J.; Weston, S.; Williams, A.; Keefer, C.; Engelhardt, A.; Cooper, T.; Mayer, Z.; Kenkel, B.; R Core Team; et al. Caret: Classification and Regression Training. Available online: https://github.com/topepo/caret/ (accessed on 20 March 2019).
  41. Breiman, L. Bagging Predictors. Mach. Learn. 1996, 24, 123–140. [Google Scholar] [CrossRef]
  42. Provost, F.; Fawcett, T.; Kohavi, R. The case against accuracy estimation for comparing induction algorithms. In Proceedings of the 15th International Conference on Machine Learning (imlc-98), Madison, WI, USA, 24–27 July 1998; p. 9. [Google Scholar]
  43. Hanley, J.-A.; McNeil, B.J. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982, 143, 29–36. [Google Scholar] [CrossRef]
  44. Bradley, A. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recognit. 1997, 30, 1145–1159. [Google Scholar] [CrossRef] [Green Version]
  45. Leemans, R. Global Data Sets Collected and Compiled by the Biosphere Project; International Institute for Applied Systems Analysis (IIASA): Laxenburg, Austria, 1990. [Google Scholar]
  46. Pleniou, M.; Koutsias, N. Sensitivity of spectral reflectance values to different burn and vegetation ratios: A multi-scale approach applied in a fire affected area. ISPRS J. Photogramm. Remote Sens. 2013, 79, 199–210. [Google Scholar] [CrossRef]
  47. Nelson, M.A.; Bishop, T.F.A.; Triantafilis, J.; Odeh, I.O.A. An error budget for different sources of error in digital soil mapping. Eur. J. Soil Sci. 2011, 62, 417–430. [Google Scholar] [CrossRef]
  48. Sexton, J.O.; Noojipady, P.; Anand, A.; Song, X.P.; McMahon, S.; Huang, C.; Feng, M.; Channan, S.; Townshend, J.R. A model for the propagation of uncertainty from continuous estimates of tree cover to categorical forest cover and change. Remote Sens. Environ. 2015, 156, 418–425. [Google Scholar] [CrossRef] [Green Version]
  49. Montgomery, D.R. Soil erosion and agricultural sustainability. Proc. Natl. Acad. Sci. USA 2007, 104, 13268–13272. [Google Scholar] [CrossRef] [Green Version]
  50. Vila, M.; Weiner, J. Are invasive plant species better competitors than native plant species?—Evidence from pair-wise experiments. Oikos 2004, 105, 229–238. [Google Scholar] [CrossRef]
  51. Suding, K.N.; Gross, K.L.; Houseman, G.R. Alternative states and positive feedbacks in restoration ecology. Trends Ecol. Evol. 2004, 19, 46–53. [Google Scholar] [CrossRef]
  52. Muradian, R. Ecological thresholds: A survey. Ecol. Econ. 2001, 38, 7–24. [Google Scholar] [CrossRef]
  53. Cumming, G.S.; Collier, J. Change and Identity in Complex Systems. Ecol. Soc. 2005, 10, 29. [Google Scholar] [CrossRef]
  54. Kinzig, A.P.; Ryan, P.; Etienne, M.; Allison, H.; Elmqvist, T.; Walker, B.H. Resilience and Regime Shifts: Assessing Cascading Effect. Ecol. Soc. 2006, 11, 20. [Google Scholar] [CrossRef]
  55. Angassa, A. The ecological impact of bush encroachment on the yield of grasses in Borana rangeland ecosystem. Afr. J. Ecol. 2005, 43, 14–20. [Google Scholar] [CrossRef]
  56. Rohde, R.F.; Moleele, N.M.; Mphale, M.; Allsopp, N.; Chanda, R.; Hoffman, M.T.; Magole, L.; Young, E. Dynamics of grazing policy and practice: Environmental and social impacts in three communal areas of southern Africa. Environ. Sci. Policy 2006, 9, 302–316. [Google Scholar] [CrossRef]
  57. Imeson, A.C.; Lavee, H. Soil erosion and climate change: The transect approach and the influence of scale. Geomorphology 1998, 23, 219–227. [Google Scholar] [CrossRef]
  58. Meadows, M.E.; Hoffman, T.M. Land Degradation and Climate Change in South Africa. Geogr. J. 2003, 169, 168–177. [Google Scholar] [CrossRef]
  59. Nearing, M.A.; Pruski, F.F.; O’Neal, M.R. Expected Climate Change Impacts on Soil Erosion Rates: A Review. J. Soil Water Conserv. 2004, 59, 43–50. [Google Scholar]
  60. Singh, P.; Behera, H.; Singh, A. Impact and Effectiveness of “Watershed Development Programmes” in India; Centre for Rural Studies, National Insitute of Admininstrative Research: Mussoorie, India, 2010; p. 67. [Google Scholar]
  61. Government of Karnataka. Integrated Watershed Management Programme—Detailed Project Report; Government of Karnataka, Watershed Development Department: Bangalore, India, 2009; p. 776.
  62. Ludwig, J.A.; Tongway, D.J. Monitoring the Condition of Australian Arid Lands: Linked Plant-Soil Indicators. In Ecological Indicators; Springer US: Boston, MA, USA, 1992; pp. 765–772. [Google Scholar]
  63. Reynolds, J.F.; Maestre, F.T.; Kemp, P.R.; Stafford-Smith, D.M.; Lambin, E. Natural and Human Dimensions of Land Degradation in Drylands: Causes and Consequences. In Terrestrial Ecosystems in a Changing World; Canadell, J.G., Pataki, D.E., Pitelka, L.F., Eds.; Global Change: Berlin, Germany, 2007; p. 336. [Google Scholar]
Remotesensing 11 01800 g001 550
Figure 1. Schematic illustration of the Land Degradation Surveillance Framework (LDSF) sampling framework showing the spatial stratification of each site and the randomization of the 16 clusters and 160 plots.
Figure 1. Schematic illustration of the Land Degradation Surveillance Framework (LDSF) sampling framework showing the spatial stratification of each site and the randomization of the 16 clusters and 160 plots.
Remotesensing 11 01800 g001
Remotesensing 11 01800 g002 550
Figure 2. The relative frequency of maximum annual fractional vegetation cover (i.e., SATVImax) vegetation structure class in the dataset. The values were divided into 10 groups.
Figure 2. The relative frequency of maximum annual fractional vegetation cover (i.e., SATVImax) vegetation structure class in the dataset. The values were divided into 10 groups.
Remotesensing 11 01800 g002
Remotesensing 11 01800 g003 550
Figure 3. Model performance metrics for the RF prediction model, based on both calibration (training) and validation (testing) datasets. Panel 1 shows receiver–operator characteristics (ROCs), Panel 2 shows predicted probabilities versus observed erosion (0/1) as simple boxplots for both models’ runs, and Panel 3 shows a confusion matrix for predicted and observed (actual) erosion classes using the test dataset.
Figure 3. Model performance metrics for the RF prediction model, based on both calibration (training) and validation (testing) datasets. Panel 1 shows receiver–operator characteristics (ROCs), Panel 2 shows predicted probabilities versus observed erosion (0/1) as simple boxplots for both models’ runs, and Panel 3 shows a confusion matrix for predicted and observed (actual) erosion classes using the test dataset.
Remotesensing 11 01800 g003
Remotesensing 11 01800 g004 550
Figure 4. Predicted probabilities of soil erosion in the global tropics. The prediction model was developed based on MODIS imagery over the period 2005 to 2017, matching annual composites to the years of LDSF field data collection and fitted to annual composite reflectance data for 2002, 2007, 2012, and 2017. Hyper-arid (i.e., deserts) areas and areas with elevations higher than 4300 m were masked (i.e., white areas on the map).
Figure 4. Predicted probabilities of soil erosion in the global tropics. The prediction model was developed based on MODIS imagery over the period 2005 to 2017, matching annual composites to the years of LDSF field data collection and fitted to annual composite reflectance data for 2002, 2007, 2012, and 2017. Hyper-arid (i.e., deserts) areas and areas with elevations higher than 4300 m were masked (i.e., white areas on the map).
Remotesensing 11 01800 g004
Remotesensing 11 01800 g005 550
Figure 5. Predicted probabilities of soil erosion by continent and subcontinent (as defined by the United Nations Statistics Division) for 2017, showing the distribution of predicted erosion (%) based on 391,000 random points. The subcontinents were arranged according to median erosion prevalence. Southern Europe and Oceania were not included in the figure given the lack of validation data for these continents.
Figure 5. Predicted probabilities of soil erosion by continent and subcontinent (as defined by the United Nations Statistics Division) for 2017, showing the distribution of predicted erosion (%) based on 391,000 random points. The subcontinents were arranged according to median erosion prevalence. Southern Europe and Oceania were not included in the figure given the lack of validation data for these continents.
Remotesensing 11 01800 g005
Remotesensing 11 01800 g006 550
Figure 6. Relative frequency of observed erosions in LDSF field plots by vegetation structure class. The vegetation structure classes were arranged by frequency of observed erosions. Note that we did not include wetlands and restricted vegetation types. The crossbars show the mean +/− standard errors for the frequency of eroded plots in each vegetation structure class.
Figure 6. Relative frequency of observed erosions in LDSF field plots by vegetation structure class. The vegetation structure classes were arranged by frequency of observed erosions. Note that we did not include wetlands and restricted vegetation types. The crossbars show the mean +/− standard errors for the frequency of eroded plots in each vegetation structure class.
Remotesensing 11 01800 g006
Remotesensing 11 01800 g007 550
Figure 7. Violin plot with box-plot overlaid, showing the probability of observed erosions relative to fractional vegetation cover (SATVImax), divided into 10 categories. The probability of erosion is based on a generalized linear mixed-effects model with LDSF sentinel sites and clusters within sites as random effects.
Figure 7. Violin plot with box-plot overlaid, showing the probability of observed erosions relative to fractional vegetation cover (SATVImax), divided into 10 categories. The probability of erosion is based on a generalized linear mixed-effects model with LDSF sentinel sites and clusters within sites as random effects.
Remotesensing 11 01800 g007
Remotesensing 11 01800 g008 550
Figure 8. Predicted probabilities of soil erosion for Laikipia, Meru, Isiolo, and Samburu counties in central and northern Kenya. The region shown is outlined in gray relative to a map of Kenya below the legend on the right.
Figure 8. Predicted probabilities of soil erosion for Laikipia, Meru, Isiolo, and Samburu counties in central and northern Kenya. The region shown is outlined in gray relative to a map of Kenya below the legend on the right.
Remotesensing 11 01800 g008
Remotesensing 11 01800 g009 550
Figure 9. Predicted probabilities of soil erosion for Karnataka, India for 2002, 2007, 2012, and 2017. The location of Karnataka is shown outlined in gray on the map of India below the legend.
Figure 9. Predicted probabilities of soil erosion for Karnataka, India for 2002, 2007, 2012, and 2017. The location of Karnataka is shown outlined in gray on the map of India below the legend.
Remotesensing 11 01800 g009
Remotesensing 11 01800 g010 550
Figure 10. Predicted probabilities of soil erosion using Holdridge life zones in the tropics (excluding tropical and subtropical deserts), ranked from the lowest to highest median erosion in 2017.
Figure 10. Predicted probabilities of soil erosion using Holdridge life zones in the tropics (excluding tropical and subtropical deserts), ranked from the lowest to highest median erosion in 2017.
Remotesensing 11 01800 g010
Table
Table 1. Summary of the data included in the current study, showing the number of LDSF sites and countries by year.
Table 1. Summary of the data included in the current study, showing the number of LDSF sites and countries by year.
YearNumber of LDSF SitesNumber of Countries
200522
200663
200761
200853
200964
2010278
2011269
20122814
20132312
2014159
20151712
201696
2017148
Table
Table 2. Summary of LDSF plots (%) by vegetation structure class.
Table 2. Summary of LDSF plots (%) by vegetation structure class.
Vegetation Structure ClassLDSF Plots (%)
Cropland34
Grassland15
Woodland12
Forest10
Shrubland10
Wooded grassland8
Bushland6
Other4
Thicket1

Share and Cite

MDPI and ACS Style

Vågen, T.-G.; Winowiecki, L.A. Predicting the Spatial Distribution and Severity of Soil Erosion in the Global Tropics using Satellite Remote Sensing. Remote Sens. 2019, 11, 1800. https://0-doi-org.brum.beds.ac.uk/10.3390/rs11151800

AMA Style

Vågen T-G, Winowiecki LA. Predicting the Spatial Distribution and Severity of Soil Erosion in the Global Tropics using Satellite Remote Sensing. Remote Sensing. 2019; 11(15):1800. https://0-doi-org.brum.beds.ac.uk/10.3390/rs11151800

Chicago/Turabian Style

Vågen, Tor-Gunnar, and Leigh Ann Winowiecki. 2019. "Predicting the Spatial Distribution and Severity of Soil Erosion in the Global Tropics using Satellite Remote Sensing" Remote Sensing 11, no. 15: 1800. https://0-doi-org.brum.beds.ac.uk/10.3390/rs11151800

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop