Spatial Prediction Models Research Articles

When context matters: Spatial prediction models of environmental conditions can identify target areas for wild bee habitat management interventions

Prioritizing where to implement management interventions is critical because managers have limited budgets and the effect of habitat enhancement depends on site-specific environmental conditions. Field experiments can identify the conditions where habitat enhancement is most effective, but are typically of limited extent and thus not sufficient for producing spatial predictions that can guide management efforts. We tested if we could produce spatial predictions maps – showing where management interventions to enhance bee habitat would be most successful – by combining spatial predictions of plant community composition (i.e., environmental conditions) obtained from field surveys with a field experiment, in which we quantified the effect of three types of management interventions on bee species richness. Using information from digital maps, we predicted plant species composition within power line clearings across southeast Norway. The intervention type, which involved cutting and removal of the woody vegetation, resulted in the largest increase in bee species richness, but the enhanced bee species richness was limited to clearings with forb-dominated vegetation. Importantly, the estimated effects on bee species richness did not differ between models using the predicted, versus the empirically observed, plant species composition as predictor, making it possible to produce spatial predictions of the increase in bee richness from implementing different management interventions. Synthesis and applications: Combining field surveys with data from field experiments can be used to produce high-resolution maps showing where wild bee habitat enhancement is likely to have the greatest effect. Such maps can inform decisions about where to allocate costly management interventions.

High spatial resolution health risk assessment of road traffic noise on ischemic heart disease deaths for Melbourne, Australia in 2011

TPS 781: Health effects of noise, Exhibition Hall, Ground floor, August 27, 2019, 3:00 PM - 4:30 PM Introduction: Road traffic noise pollution increases risk of mortality from ischemic heart disease (IHD). Noise pollution is highly localized, so high resolution mapping of the exposure surface is a key concern. For health impact assessments optimally the data on exposure and health outcomes would both be available at small-area level. However, in Australia IHD deaths data are only easily accessed at low resolution, aggregated to local government areas (LGA, mean ~130,000 people). Our method utilizes low-resolution health data with high spatial resolution noise mapping for producing health risk maps at the mesh-block level (MB, ~75 people). Methods: We used a noise contour map for Melbourne, Australia from 2011 based on a spatial predictive model to estimate population exposures at the MB level. We used the non-linear exposure–response function for traffic noise and IHD as recommended by the World Health Organization to calculate attributable fractions of deaths for MBs. LGA-level IHD rates were then downscaled to MBs using publicly available data to estimate attributable numbers. Results: Average noise exposures for MBs was 35 dB and peaked at 72 dB. Approximately 5% of the population was exposed to traffic noise above the threshold of 55 dB. Attributable rates of IHD deaths due to noise were generally very low but in some MBs was as high as 5-10 per 100,000 and in extreme cases can represent more than 8% of the total expected burden of IHD death. Conclusion: Road traffic noise is a public health concern because it increases risk for IHD death. We present a method that does not require high resolution health data and can be used with high resolution exposure data to provide estimates of health risk. This method could be applied in a wide range of other environmental health settings.

Spatial probability modelling of forest productivity indicator in Italy

• Geostatistics and GIS connected topo-climatic factors and forest productivity. • Spatial patterns showed the environmental control on forest productivity. • Probability maps differentiated high and low wood productivity rates. The prediction of forest productivity is essential for sustainable forest management, particularly in countries, like Italy, where forest is an important part of many protected areas. A spatial predictive probability model for forest productivity rates in Italy was developed over the period 1961–1990, based on 135 annually-resolved records of site productivity and auxiliary variables measured at 219 stations. Our analysis shows that the probability of finding high (>7.3 m 3 ha −1 yr −1 ) and low (<5.8 m 3 ha −1 yr −1 ) productivity rates changes across different regions of Italy. The generated spatial patterns contribute to a better understanding of the factors structuring the distribution of forest productivity in Italy because they reflect the dependence of temperature and water availability conditions on the latitudinal and altitudinal location of the study areas. We observed that the temperature control dominates forest productivity at high elevations and latitudes, whereas low-elevation sites in central and southern Italy are more sensitive to water availability. The proposed spatial probability modelling should be further assessed for its possible incorporation into forest management plans.

Importance of spatial predictor variable selection in machine learning applications – Moving from data reproduction to spatial prediction

Machine learning algorithms find frequent application in spatial prediction of biotic and abiotic environmental variables. However, the characteristics of spatial data, especially spatial autocorrelation, are widely ignored. We hypothesize that this is problematic and results in models that can reproduce training data but are unable to make spatial predictions beyond the locations of the training samples. We assume that not only spatial validation strategies but also spatial variable selection is essential for reliable spatial predictions.We introduce two case studies that use remote sensing to predict land cover and the leaf area index for the “Marburg Open Forest”, an open research and education site of Marburg University, Germany. We use the machine learning algorithm Random Forests to train models using non-spatial and spatial cross-validation strategies to understand how spatial variable selection affects the predictions.Our findings confirm that spatial cross-validation is essential in preventing overoptimistic model performance. We further show that highly autocorrelated predictors (such as geolocation variables, e.g. latitude, longitude) can lead to considerable overfitting and result in models that can reproduce the training data but fail in making spatial predictions. The problem becomes apparent in the visual assessment of the spatial predictions that show clear artefacts that can be traced back to a misinterpretation of the spatially autocorrelated predictors by the algorithm. Spatial variable selection could automatically detect and remove such variables that lead to overfitting, resulting in reliable spatial prediction patterns and improved statistical spatial model performance.We conclude that in addition to spatial validation, a spatial variable selection must be considered in spatial prediction models of ecological data to produce reliable results.

Spatiotemporal Monitoring of Coastal Urbanization Dynamics: Case Study of Algiers’ Bay, Algeria

During the last three decades, the Mediterranean coasts have been under strong anthropogenic pressure inducing ecological and biological degradations such as pollution, destruction of marine ecosystems, and loss of coastal dunes. In Algeria, the same findings have been observed in the bay of Algiers due to the massive urbanization as well as the absence of the quantitative analysis tools of this phenomenon. In this context, the assigned objective of this study is the monitoring of the urban growth in Algiers’ bay between 1985 and 2015 and the prediction of the urban growth by 2030 using remote sensing data, cellular automata modeling, and the landscape metrics analysis. Concerning the prediction step, two scenarios are considered: the “tendencial” which presents the unmanaged urban growth of Algiers City and the active implementation of the Algiers Plan by 2029. To achieve this goal, a retrospective map representing land cover has been produced at different times; this later has allowed us to quantify the urban regions and to implement the urbanization growth model. On the other hand, this study has highlighted the contribution of the spatial metrics on multi-scale level (patch/core, class level and the landscape level) to monitor coastal urbanization in a geometric/fractal view through its multi-temporal evolution, as well as in the two predictive landscapes. This allowed through the spatial metrics indicator and predictive modeling to observe a tendency to massive urbanization of the bay at the expense of remaining natural areas, but also to provide and exhibit some geographical aspects such as the layout of urban cores and the linearity of urbanization parallel to the coast, in order to propose some significant measures for policy-makers to avoid the alarmist environmental degradation of the bay.

Forest Fire Susceptibility Modeling Using a Convolutional Neural Network for Yunnan Province of China

Forest fires have caused considerable losses to ecologies, societies, and economies worldwide. To minimize these losses and reduce forest fires, modeling and predicting the occurrence of forest fires are meaningful because they can support forest fire prevention and management. In recent years, the convolutional neural network (CNN) has become an important state-of-the-art deep learning algorithm, and its implementation has enriched many fields. Therefore, we proposed a spatial prediction model for forest fire susceptibility using a CNN. Past forest fire locations in Yunnan Province, China, from 2002 to 2010, and a set of 14 forest fire influencing factors were mapped using a geographic information system. Oversampling was applied to eliminate the class imbalance, and proportional stratified sampling was used to construct the training/validation sample libraries. A CNN architecture that is suitable for the prediction of forest fire susceptibility was designed and hyperparameters were optimized to improve the prediction accuracy. Then, the test dataset was fed into the trained model to construct the spatial prediction map of forest fire susceptibility in Yunnan Province. Finally, the prediction performance of the proposed model was assessed using several statistical measures—Wilcoxon signed-rank test, receiver operating characteristic curve, and area under the curve (AUC). The results confirmed the higher accuracy of the proposed CNN model (AUC 0.86) than those of the random forests, support vector machine, multilayer perceptron neural network, and kernel logistic regression benchmark classifiers. The CNN has stronger fitting and classification abilities and can make full use of neighborhood information, which is a promising alternative for the spatial prediction of forest fire susceptibility. This research extends the application of CNN to the prediction of forest fire susceptibility.

Macro-level accident modeling in Novi Sad: A spatial regression approach

In this study, a macroscopic analysis was conducted in order to identify the factors which have an effect on traffic accidents in traffic analysis zones. The factors that impact accidents vary according to the characteristics of the observed area, which in turn leads to a discrepancy between research and practice. The total number of accidents was observed in this paper, along with the number of motorized and non-motorized mode accidents within a three-year period in the city of Novi Sad. The models used for this analysis were spatial predictive models comprised of the classical predictive space model, spatial lag model and spatial error model. The spatial lag model showed the best performances concerning the total number of accidents and number of motorized mode accidents, whereas the spatial error model was prominent within the number of non-motorized mode accidents. The results found that increasing Daily Vehicle-Kilometers Traveled, parking spaces, 5-legged intersections and signalized intersections increased all types of accidents. The other demographic, traffic, road and environment characteristics showed that they had a different effect on the observed types of accidents. The results of this research can be benefitial to reserachers who deal with traffic engineering, space planning as well as making decisions with the aim of preparing countermeasures necessary for road safety improvement in the analysed area.

A novel voting ensemble model for spatial prediction of landslides using GIS

ABSTRACTInherent complex topography and drastic weather patterns together have concocted various natural disasters worldwide. In difficult terrains such as those prevalent in the North-Eastern regions of India, coupled with the factors such as population explosion and improper land use, lead them to witness some of the world’s most drastic landslides with an astonishing frequency, reckoning landslide susceptibility assessment crucial in such regions. This paper focuses on exploring a promising machine learning ensemble technique of Majority-based voting which has seldom been employed for landslide susceptibility assessment. The ensemble comprises Logistic Regression (LR), Gradient Boosted Decision Trees (GBDT) and Voting Feature Interval (VFI) to prepare landslide susceptibility zonation maps for the Brahmaputra valley region (Assam & Nagaland) and its close vicinity. In the first stage of the study, a landslide inventory for the area comprising 436 landslide locations was prepared in geographic information system (GIS), substantiated by news reports and remote sensing data. In the second stage, 16 landslide causative thematic maps including Elevation, Slope, Slope Aspect, General Curvature, Plan Curvature, Profile Curvature, Surface Roughness, Topographic Wetness Index (TWI), Stream Power Index (SPI), Slope Length, Normalized Difference Vegetation Index (NDVI), Land Use/Land Cover (LULC), Distance from Roads, Rivers, Faults and Railways were prepared. In the third stage, the landslide inventory was annexed with the causative factor maps to obtain a dataset comprising coordinates of the locations and the values of aforementioned causative factors on the corresponding coordinates. The proposed model was then trained and tested on the prepared dataset (70%:30% split). Finally, the efficiency of the new model was tested using the area under receiver operating characteristic curve (AUC of ROC). The validation results demonstrate the mettle of the proposed majority-based voting ensemble LR-GBDT-VFI (AUC: 0.98) against the conventional techniques such as Decision Trees, Support Vector Machines, Random Forest, etc. Altogether, the study offers an approach with wide scope across the field of landslide hazard assessment.

Oak Savannas in Western New York State, Circa 1795: Synthesizing Predictive Spatial Models and Historical Accounts to Understand Environmental and Native American Influences

This study models and maps oak savanna distribution ca. 1795 and determines how environmental conditions and Native American land use (NALU) shaped its distribution. Historical research has analyzed early landscape accounts to assess how Native Americans modified forests throughout the eastern United States. Predictive spatial modeling has sought to quantify anthropogenic and environmental drivers of forest conditions and to predict locations where NALU changed forest composition. Yet, studies have not rigorously synthesized these two methods. This research focused on oak savannas in western New York State (27,617 km2). We trained models of oak savanna distribution from historic vegetation data in relation to environmental predictors and NALU proxies. We then mapped historical accounts of oak savannas and NALU at European-American arrival and compared them to model predictions. Results suggest that 2 to 17 percent (depending on modeling technique) of the study area contained oak savanna, with a favored estimate of 3 to 6 percent. Synthesis of models and accounts suggests that oak savannas were attributable to NALU and dry environmental conditions but that NALU (specifically burning) was present at most oak savanna locations. Models of oak savanna distribution that considered proximity to Native American settlement had higher predictive performance and better predicted locations of historical oak savanna accounts, including those with descriptions of Native American burning. This study suggests that former oak savannas in the study area can be largely attributed to NALU. Furthermore, this study’s methodology and results contribute to a larger body of geographical literature on savanna landscapes. Key Words: anthropogenic burning, forests, modeling, Native Americans, oak savanna.

Testing Application of Geographical Information Systems, Forensic Geomorphology and Electrical Resistivity Tomography to Investigate Clandestine Grave Sites in Colombia, South America.

Colombian forensic investigators required assistance locating clandestine burials of missing persons related to human right atrocities from 14years ago. Geoscientific search methods were trialled, including a predictive spatial statistical model, using various input and database information, to select the most likely grave locations in difficult mountainous terrain. Groundwork using forensic geomorphology, near-surface geophysics (ERT) and subsequent probing identified suspect burial positions. One site was in mountainous terrain and the other in former school grounds, both difficult to access and in poor weather conditions. In the mountainous area, a negative resistivity anomaly area was identified and intrusively investigated, found to be a buried rock. In school grounds, after MESP and intelligence were used to identify a burial site, surface depressions were identified, and ERT datasets collected over the highest priority depression; intrusive investigations discovered a hand-dug pit containing animal bones. This approach is suggested for Latin American searches.

Improved particle swarm optimization LSSVM spatial location trajectory data prediction model in health care monitoring system

With the real-time recording of the health care system, real-time monitoring of user trajectory data can effectively protect the health and health of the body. In order to accurately establish the spatial position trajectory data prediction model, a space trajectory data prediction model based on variable space chaotic particle swarm optimization (CPSO)-optimized least squares support vector machine (LSSVM) is proposed. Firstly, aiming at the problem that particle swarm optimization (PSO) is easy to fall into local optimum, the global optimization performance of PSO algorithm is improved by using variable space chaos search strategy and particle out-of-bounds mirror processing strategy. Numerical simulation experiments verify the superiority of improved particle swarm optimization algorithm. Based on the CPSO algorithm, the hyper parameters of the CPSO-optimized least squares support vector machine (CPSO-LSSVM) are proposed to improve the prediction accuracy of the model. Finally, using GeoLife and T-Drive dataset as the research object, the CPSO-LSSVM spatial position prediction model is established by using the trajectory data of the dataset. The simulation experiment verifies that the CPSO-LSSVM positional space prediction model has higher prediction accuracy and more strong generalization ability to accurately and effectively predict spatial location.

Accounting for covariate distributions in slope-unit-based landslide susceptibility models. A case study in the alpine environment

Thousands or even million of pixels can be contained in a single Slope Unit. Hence, each covariate used in spatial predictive models is characterized by a distribution of values for each Slope Unit.Here, we model the whole covariates' distribution within Slope Units for landslide susceptibility purposes. This is done by finely dissecting each covariate into quantiles and then modeling the susceptibility via a LASSO penalized Binary Logistic Regression. We choose a LASSO penalization because the common Stepwise procedure is not selective enough to shrink a large number of covariates to an interpretable subset (which we also demonstrate here). LASSO mostly selects 6 covariates out of 372 to explain the spatial distribution of shallow landslides in the Upper Badia valley, Italian Alps. This allows us to verify that the selection does not include any quantile close to the median hence, nor to the mean. The latter is the common representation of the covariates' distribution within Slope Units, which we also test and report in the supplements. Overall, we suggest to always investigate the whole distribution because the mean may not be the most informative nor the most performing way to generate Slope-Unit-based susceptibility models. In this general context, we generate our landslide inventory by combining semi-automated (OBIA) and manual mapping procedures. Our inventory, quantile covariates' representation and LASSO penalization produce excellent performances and interpretable relations between covariates and landslides.

Spatial predictive modelling essential to assess the environmental impacts of herbicides

The precise prediction of adsorption coefficient (Kd) of herbicides retention in soil requires a careful and robust assessment of alternative statistical methods for predictive modelling. In this work, Kd was modelled as function of soil variables from a regional soil survey using various frameworks: Ordinary and Partial Least Squares regression, Random Forests, Generalized Boosted regression (GB), and Bayesian regression with INLA (INLA). Each approach is applied with and without spatial coordinates included in the covariates for the mean structure. Further, the residuals from the mean structure are either assumed independent or assumed spatially correlated and kriged. For model validation, measurements of pointwise and global predictive ability were assessed. All methods showed good performance (prediction error < 20%). GB without spatial coordinates in the mean structure, nor residuals kriged (i.e. the raw GB predictions based on the covariates) gave a small pointwise classification rate, but INLA with spatial constraints yielded the best fit (the smallest mean squared prediction errors) which resulted suitable for both, process understanding and mapping. The modelling has been illustrated for mapping glyphosate retention, with aluminum oxides, pH, sand, and clay percentages identified as master variables. Results may be extended to other herbicides and dynamic parameters using georeferenced data.

Walk this way! An IoT-based urban routing system for smart cities

Future smart cities are expected to change radically the way people live, interact and move in urban environments. This will be possible thanks to the massive amount of data that will be generated by ubiquitously deployed sensor devices through the Internet of Things paradigm. Indeed, solutions able to improve the quality of urban mobility for citizens are of particular interests. As a matter of fact, they are a key objective for many municipal administrations as well as one of the priority themes of the European Commission. In this context, this work proposes an advanced smart urban routing service named SURF, which is specifically thought for pedestrians and cyclists willing to move inside a city. The system allows to retrieve the best route between a source location and a destination according to user-defined objective function (e.g., selecting the route with the best air quality or with the lowest average temperature). This is possible through the interaction with a federation of IoT testbeds, deployed worldwide. This paper comments on the implementation and the evaluation of the proposed system, focusing on both the backend (data retrieval and spatio/temporal data interpolation and forecasting operations) and the front-end (graphical user interface). We assess the performance of several spatial interpolation and temporal prediction models, to understand their relationship with the particular sensor measurements (air pollution, temperature, sound pressure level, etc.). We show through experiments that for what concerns spatial interpolation, Universal Kriging is generally able to perform well across all sensor measurements and can be selected as a generic interpolation strategy. As for temporal prediction, experiments highlight a tradeoff between model accuracy and look-ahead capability. We note that short and mid-term prediction methods show satisfactory performance across all sensor measurements. Finally, subjective and objective experiments demonstrate the positive impact of IoT-based solutions for smart routing on urban citizens.

Bayesian data fusion for combining maps of predicted soil classes: A case study using legacy soil profiles and DEM covariates in Iran

This paper tackles the issue of spatially predicting soil classes by combining at best soil information coming directly from legacy soil profiles with soil information indirectly obtained from spatial covariates. Based on Multinomial Logistic Regression (MLR) and Bayesian Maximum Entropy (BME) models, we first show that prediction models easily lead to very different soil maps while having at the same time quite comparable global performances. By relying afterwards on a Bayesian data fusion (BDF) approach, we emphasize the benefit of combining the output of these two prediction models in order to get a single final map that combines the major spatial features of the MLR and BME maps while at the same time improving the quality of the predictions.The advocated methodology is illustrated with the mapping of World Reference Base (WRB) soil classes over a 10,480 km2 area located in Iran. A set of 390 soil profiles allowed us to assign the WRB soil classes at these locations. In parallel, a set of potentially related covariates were computed from a 90 m resolution digital elevation model. Using MLR and BME models, predictions were obtained separately at the nodes of a 90 m resolution grid. Even if the performances of the MLR and BME models compare well, it is shown that a BDF procedure that combines both results yields improved performances, with spatial features that are a balanced combination of those found separately on the MLR and BME maps.These results emphasize the benefit of data fusion in order to improve the quality of the final map. Though the study was conducted here using MLR and BME models for predicting WRB soil classes, we believe this methodology and the corresponding findings are relevant when it comes to handle the results of spatial prediction models that are making use of distinct information sources in other soil science mapping contexts.

Groundwater potential assessment of the Sero plain using bivariate models of the frequency ratio, Shannon entropy and evidential belief function

The goal of the present research is to evaluate three bivariate models of the frequency ratio, Shannon entropy (SE) and evidential belief function in the spatial prediction of groundwater at the Sero plain located in west Azerbaijan, Iran. In the first phase, well locations with groundwater yields $${>}11\hbox { m}^{3}$$ /hr were identified (75 well locations). Ten groundwater conditioning factors affecting the occurrence of groundwater, namely, altitude, slope degree, curvature, slope aspect, rainfall, soil, land-use, geology and distance from the fault and the river, were selected for modelling. Finally, the groundwater potential map results were drawn from three implemented models and they were validated using testing data by area under the receiver operating characteristic curve (AUC). The AUCs of these models were 0.84, 81 and 85%, respectively. The results of the current study demonstrated that these models could be successfully employed for spatial prediction modelling. Moreover, the results of the SE model demonstrated that the most and the least important factors in groundwater occurrences in the area under study were altitude, curvature and rainfall, respectively. The results of this study are helpful for the Regional Water Authority of Urmia and the decision makers to comprehensively assess the groundwater exploration development and environmental management in future planning.

A Critical Review of Spatial Predictive Modeling Process in Environmental Sciences with Reproducible Examples in R

Spatial predictive methods are increasingly being used to generate predictions across various disciplines in environmental sciences. Accuracy of the predictions is critical as they form the basis for environmental management and conservation. Therefore, improving the accuracy by selecting an appropriate method and then developing the most accurate predictive model(s) is essential. However, it is challenging to select an appropriate method and find the most accurate predictive model for a given dataset due to many aspects and multiple factors involved in the modeling process. Many previous studies considered only a portion of these aspects and factors, often leading to sub-optimal or even misleading predictive models. This study evaluates a spatial predictive modeling process, and identifies nine major components for spatial predictive modeling. Each of these nine components is then reviewed, and guidelines for selecting and applying relevant components and developing accurate predictive models are provided. Finally, reproducible examples using spm, an R package, are provided to demonstrate how to select and develop predictive models using machine learning, geostatistics, and their hybrid methods according to predictive accuracy for spatial predictive modeling; reproducible examples are also provided to generate and visualize spatial predictions in environmental sciences.

Assessing potential environmental influences on killer whale (Orcinus orca) distribution patterns in the Bremer Canyon, south-west Australia

ABSTRACTSubmarine canyons play an important role in the regional distribution, abundance and dispersal of marine biota and are increasingly being recognised as geomorphic features of high conservation significance along Australia’s continental margin. Certain canyons have been described as foraging ‘hotspots’ attributable to the high abundance of apex cetacean species aggregating in these areas. Anecdotal evidence of large seasonal aggregations of killer whales in the Bremer Canyon, south-west Australia, has attracted significant research attention in the last decade. To identify important environmental drivers influencing aggregation patterns, a predictive spatial habitat model using the Maxent model was developed based on presence-only whale sighting data. In addition, remotely sensed sea surface temperature and chlorophyll-a concentrations were assessed to investigate the spatio-temporal variation in sea surface conditions. Habitat preference was predicted in areas between canyon heads, with the most influential predictor variables being depth and distance from the continental shelf break. Analysis of remote-sensing data highlighted low localised variability in surface waters and illustrated the seasonal trends of the Leeuwin Current. This study demonstrates the influence of bathymetry and submarine geomorphology on enhanced cetacean abundance and highlights the need for recognition of this potential foraging area in marine reserve planning.

Developing an Optimal Spatial Predictive Model for Seabed Sand Content Using Machine Learning, Geostatistics, and Their Hybrid Methods

Seabed sediment predictions at regional and national scales in Australia are mainly based on bathymetry-related variables due to the lack of backscatter-derived data. In this study, we applied random forests (RFs), hybrid methods of RF and geostatistics, and generalized boosted regression modelling (GBM), to seabed sand content point data and acoustic multibeam data and their derived variables, to develop an accurate model to predict seabed sand content at a local scale. We also addressed relevant issues with variable selection. It was found that: (1) backscatter-related variables are more important than bathymetry-related variables for sand predictive modelling; (2) the inclusion of highly correlated predictors can improve predictive accuracy; (3) the rank orders of averaged variable importance (AVI) and accuracy contribution change with input predictors for RF and are not necessarily matched; (4) a knowledge-informed AVI method (KIAVI2) is recommended for RF; (5) the hybrid methods and their averaging can significantly improve predictive accuracy and are recommended; (6) relationships between sand and predictors are non-linear; and (7) variable selection methods for GBM need further study. Accuracy-improved predictions of sand content are generated at high resolution, which provide important baseline information for environmental management and conservation.

Multi-source data integration for soil mapping using deep learning

Abstract. With the advances of new proximal soil sensing technologies, soil properties can be inferred by a variety of sensors, each having its distinct level of accuracy. This measurement error affects subsequent modelling and therefore must be integrated when calibrating a spatial prediction model. This paper introduces a deep learning model for contextual digital soil mapping (DSM) using uncertain measurements of the soil property. The deep learning model, called the convolutional neural network (CNN), has the advantage that it uses as input a local representation of environmental covariates to leverage the spatial information contained in the vicinity of a location. Spatial non-linear relationships between measured soil properties and neighbouring covariate pixel values are found by optimizing an objective function, which can be weighted with respect to a measurement error of soil observations. In addition, a single model can be trained to predict a soil property at different soil depths. This method is tested in mapping top- and subsoil organic carbon using laboratory-analysed and spectroscopically inferred measurements. Results show that the CNN significantly increased prediction accuracy as indicated by the coefficient of determination and concordance correlation coefficient, when compared to a conventional DSM technique. Deeper soil layer prediction error decreased, while preserving the interrelation between soil property and depths. The tests conducted suggest that the CNN benefits from using local contextual information up to 260 to 360 m. We conclude that the CNN is a flexible, effective and promising model to predict soil properties at multiple depths while accounting for contextual covariate information and measurement error.