Search

• Soil reflectance composites were generated with Sentinel-2 L2A data. • A multi-channel CNN learning framework for soil organic carbon prediction is proposed. • A dual input deep learning architecture improves the predictions. • The framework produced results that can be interpreted. The Copernicus Sentinel-2 multispectral imagery data may be aggregated to extract large-scale, bare soil, reflectance composites, which enable soil mapping applications. In this paper, this approach was tested in the German federal state of Bavaria, to provide estimations for soil organic carbon (SOC). Different temporal ranges were considered for the generation of the composites, including multi-annual and seasonal ranges. A novel multi-channel convolutional neural network (CNN) is proposed. By leveraging the advantages of deep learning techniques, it utilizes complementary information from different spectral pre-treatment techniques. The SOC predictions indicated little dissimilarity amongst the different composites, with the best performance attained for the six-year composite containing only spring months (RMSE = 12.03 g C · kg−1, R2 = 0.64, RPIQ = 0.89). It has been demonstrated that these outcomes outperform other well-known machine learning techniques. An ablation analysis was accordingly performed to evaluate the interplay of the CNN's different components to disentangle the advantages of each aspect of the proposed framework. Finally, a DUal inPut deep LearnIng architecture, named DUPLICITE, is proposed, which concatenates deep spectral features derived from the CNN mentioned earlier, as well as topographical and environmental covariates through an artificial neural network (ANN) to exploit their complementarity. The proposed approach was demonstrated to provide an improvement in the overall prediction performance (RMSE = 11.60 gC · kg−1, R2 = 0.67, RPIQ = 0.92). [ABSTRACT FROM AUTHOR]

Digital soil mapping (DSM) techniques are widely employed to generate soil maps. Soil properties are typically predicted individually, while ignoring the interrelation between them. Models for predicting multiple properties exist, but they are computationally demanding and often fail to provide accurate description of the associated uncertainty. In this paper a convolutional neural network (CNN) model is described to predict several soil properties with quantified uncertainty. CNN has the advantage that it incorporates spatial contextual information of environmental covariates surrounding an observation. A single CNN model can be trained to predict multiple soil properties simultaneously. I further propose a two-step approach to estimate the uncertainty of the prediction for mapping using a neural network model. The methodology is tested mapping six soil properties on the French metropolitan territory using measurements from the LUCAS dataset and a large set of environmental covariates portraying the factors of soil formation. Results indicate that the multivariate CNN model produces accurate maps as shown by the coefficient of determination and concordance correlation coefficient, compared to a conventional machine learning technique. For this country extent mapping, the maps predicted by CNN have a detailed pattern with significant spatial variation. Evaluation of the uncertainty maps using the median of the standardized squared prediction error and accuracy plots suggests that the uncertainty was accurately quantified, albeit slightly underestimated. The tests conducted using different window size of input covariates to predict the soil properties indicate that CNN benefits from using local contextual information in a radius of 4.5 km. I conclude that CNN is an effective model to predict several soil properties and that the associated uncertainty can be accurately quantified with the proposed approach. • A convolutional neural network is used for soil mapping. • Soil texture, organic carbon, pH and nitrogen are predicted using a single model. • The model is trained on topsoil LUCAS data over France. • A two-step method to quantify the uncertainty of the prediction is tested. • The method for uncertainty quantification is applicable to any neural network model. [ABSTRACT FROM AUTHOR]

Abstract: Time Domain Reflectometry (TDR) has become an established and reliable electromagnetic method for quantification of the soil water content (SWC). The existing calibration models used to convert the dielectric permittivity to SWC cannot be assumed to be universal, mainly due to the fact that the dielectric permittivity of the soils depends upon various soil specific properties. This paper focuses on development of a methodology using Artificial Neural Network (ANN) for enhanced determination of SWC based on dielectric permittivity measurement via TDR. Within this study, the dielectric permittivity obtained with the TDR measurement, the bulk dry density, the specific gravity and the fines content were selected as the input parameters for an ANN model. The performance of the ANN model has been compared with some of the existing calibration models. The average RMSE of the ANN-based approach was 0.009cm3 cm−3 and 0.007cm3 cm−3 with 8 and 11 nodes, respectively, while the same value was in a range of 0.019 to 0.033cm3 cm−3 for the existing calibration models and 0.015 to 0.018cm3 cm−3 for the best-fitted calibration models when all soil classes considered together. The results showed that the SWC predicted by the ANN model as proposed in this study is more accurate than that of the existing calibration models for sandy soils. [Copyright &y& Elsevier]

Abstract: The aims of this paper are: to compare different data mining algorithms for modelling soil visible–near infrared (vis–NIR: 350–2500nm) diffuse reflectance spectra and to assess the interpretability of the results. We compared multiple linear regression (MLR), partial least squares regression (PLSR), multivariate adaptive regression splines (MARS), support vector machines (SVM), random forests (RF), boosted trees (BT) and artificial neural networks (ANN) to estimate soil organic carbon (SOC), clay content (CC) and pH measured in water (pH). The comparisons were also performed using a selected set of wavelet coefficients from a discrete wavelet transform (DWT). Feature selection techniques to reduce model complexity and to interpret and evaluate the models were tested. The dataset consists of 1104 samples from Australia. Comparisons were made in terms of the root mean square error (RMSE), the corresponding R 2 and the Akaike Information Criterion (AIC). Ten-fold-leave-group out cross validation was used to optimise and validate the models. Predictions of the three soil properties by SVM using all vis–NIR wavelengths produced the smallest RMSE values, followed by MARS and PLSR. RF and especially BT were out-performed by all other approaches. For all techniques, implementing them on a reduced number of wavelet coefficients, between 72 and 137 coefficients, produced better results. Feature selection (FS) using the variable importance for projection (FSVIP) returned 29–31 selected features, while FSMARS returned between 11 and 14 features. DWT–ANN produced the smallest RMSE of all techniques tested followed by FSVIP–ANN and FSMARS–ANN. However, both the FSVIP–ANN and FSMARS–ANN models used a smaller number of features for the predictions than DWT–ANN. This is reflected in their AIC, which suggests that, when both the accuracy and parsimony of the model are taken into consideration, the best SOC model was the FSMARS–ANN, and the best CC and pH models were those from FSVIP–ANN. Analysis of the selected bands shows that: (i) SOC is related to wavelengths indicating C―O, C═O, and N―H compounds, (ii) CC is related to wavelengths indicating minerals, and (iii) pH is related to wavelengths indicating both minerals and organic material. Thus, the results are sensible and can be used for comparison to other soils. A systematic comparison like the one presented here is important as the nature of the target function has a strong influence on the performance of the different algorithms. [Copyright &y& Elsevier]

Abstract: The selection of calibration method is one of the main factors influencing the measurement accuracy with visible (vis) and near infrared (NIR) spectroscopy. This paper compared the performance of three calibration methods, namely, principal component regression (PCR), partial least squares regression (PLSR) and back propagation neural network (BPNN) analyses for the accuracy of measurement of selected soil properties, namely, organic carbon (OC) and extractable forms of potassium (K), sodium (Na), magnesium (Mg) and phosphorous (P). A total of 168 soil samples collected from Belgium and Northern France were used as the data set for the calibration–validation procedure. Optical scanning was carried out on fresh soil samples with a fibre-type, vis–NIR (LabSpec®Pro Near Infrared Analyzer, Analytical Spectral Devices, Inc, USA) with a measurement range of 350–2500nm. The entire data set was split randomly into 3 replicates of 90% and 10% for the cross-validation set and prediction set, respectively. The input of BPNN was the first 5 principal components (PCs) resulted from the principal component analysis (PCA) and the optimal number of latent variables (LVs) obtained from PLSR. Both the leave-one-out cross validation and prediction for the three replicates showed that all BPNN-LV models outperformed PCR, PLSR and BPNN-PCs models. Furthermore, BPNN-PCs and PLSR provided, respectively, better performance than PCR. The best predictions were obtained with BPNN-LVs modelling for OC (R pre 2 =0.84 and residual prediction deviation (RPD)=2.54) and Mg (R pre 2 =0.82 and RPD=2.54), which were classified as excellent model predictions. The prediction of K, P and Na was classified as good (R pre 2 =0.68–0.74 and RPD=1.77–1.94), where quantitative predictions were considered possible. It is recommended to adopt BPNN-LVs modelling technique for higher accuracy measurement of the selected soil properties with vis–NIR spectroscopy, in comparison with PCR, PLS and BPNN-PCs modelling techniques. [Copyright &y& Elsevier]

Abstract: There is a need for accurate, quantitative soil information for natural resource planning and management. This information shapes the way decisions are made as to how soil resources are assessed and managed. This paper proposes a novel method for whole-soil profile predictions (to 1m) across user-defined study areas where limited soil information exists. Using the Edgeroi district in north-western NSW as the test site, we combined equal-area spline depth functions with digital soil mapping techniques to predict the vertical and lateral variations of carbon storage and available water capacity (AWC) across the 1500km2 area. Neural network models were constructed for both soil attributes to model their relationship with a suite of environmental factors derived from a digital elevation model, radiometric data and Landsat imagery. Subsequent fits of the models resulted in an R 2 of 44% for both carbon and AWC. For validation at selected model depths, R 2 values ranged between 20 and 27% for carbon prediction (RMSE: 0.30–0.52 log (kg/m3)) and between 8 and 29% for AWC prediction (RMSE: 0.01m/m). Visually, reconstruction of splines at selected validation data points indicated an average fit with raw data values. In order to improve upon our model and validation results there is a need to address some of the structural and metrical uncertainties identified in this study. Nevertheless, the resulting geo-database of quantitative soil information describing its spatial and vertical variations is an example of what can be generated with this proposed methodology. We also demonstrate the functionality of this geo-database in terms of data enquiry for user-defined queries. [Copyright &y& Elsevier]

In this paper we assess the utility of mapping regolith properties as continuously varying fields using environmental correlation over large spatial areas. The assessment is based on a comparison of results from spatially global and local analysis methods.The two methods used are a feed-forward, back propagation neural network, applied globally, and moving window regression, applied locally. These methods are applied to five regolith properties from a field site at Weipa, Queensland, Australia. The properties considered are the proportions of oxides of aluminum, iron, silica and titanium present in samples, as well as depth to the base of the bauxite layer. These are inferred using a set of surface measurable features, consisting of Landsat data, geomorphometric indices, and distances from streams and swamps.The moving window regression results show a much stronger relationship than do those from the spatially global neural networks. The implication is that the scale of the analysis required for environmental correlation is of the order of hundreds of metres, and that spatially global analyses may incur an automatic reduction in accuracy by not modelling geographically local relationships. In this case, this effect is up to 45% error at a tolerance near half of a standard deviation. [Copyright &y& Elsevier]

[Display omitted] • The autoencoder framework is applied in VNIR-SWIR soil spectroscopy. • It maps the input data into a new compressed latent space in an unsupervised way. • This compressed information leads to enhanced accuracy of prediction. • A multi-input autoencoder utilizing multiple spectral pre-treatments is also presented. • The proposed methodologies outperform their counterparts. In the past two decades, research efforts have focused on using near infrared diffuse reflectance spectroscopy (350–2500 nm) as a rapid and cost-effective method for soil analysis. Given the multi-faceted importance of the soil ecosystem, and considering the increasing pressures exerted upon it due to climate change, degradation and urbanization, the advantages of soil spectroscopy are significant. Large soil spectral libraries have been developed to this end throughout the world. The soil matrix is however complex, posing a challenge in the determination of key soil properties from the spectra. To tackle this challenge, two methodologies are generally used: a) the use of spectral pre-processing techniques to transfer the spectra into a new space in which the association between spectrum and soil property is supposed to be more clear, and b) the use of more sophisticated machine learning models (e.g. deep learning). In this paper, we propose a novel methodology using stacked autoencoders to transform the initially recorded spectra in a new compressed (i.e. latent) space which can help the chemometric models enhance the accuracy of prediction. This is an unsupervised learning approach which only depends on the input data (i.e. the spectra). Following the significant results obtained in the literature using combinations of different spectra pre-processing techniques and the simultaneous prediction of multiple soil properties, the proposed methodology is extended to facilitate these approaches. We demonstrate this capacity by applying it in the mineral samples of the LUCAS 2009 topsoil database, and simultaneously predicting eight properties (the particle size distribution, pH, CEC, organic carbon, calcium carbonate, and total nitrogen) using an artificial neural network. Compared to standard pre-processing techniques and other transformations such as the principal components space, the proposed methodology using only one spectral source as input decreases the RMSE on average by 8.4% and by 3.5%, respectively. With respect to the current state-of-the-art and in particular a multi-input convolutional neural network which was recently proposed and outperformed the compared methodologies, the results of our multi-input methodology exhibit an average RMSE decrease of 9.9%. The interpretability aspect of the transformed feature space and the compressed spectra was also examined to identify how the compressed information encodes the input data and enables better associations between input and output. [ABSTRACT FROM AUTHOR]

• A novel 1-D CNN learning framework for soil spectral analysis is proposed. • Multiple spectral sources are combined to maximize the spectral information. • All soil properties are simultaneously predicted to utilize the cross-correlations. • A local adaptive error-correction mechanism enhances the prediction performance. • The framework produced interpretable and accurate results in the LUCAS SSL. The use of visible near-infrared and shortwave-infrared (VNIR-SWIR) diffuse reflectance spectroscopy for the estimation of soil properties is increasingly maturing with large-scale soil spectral libraries (SSLs) of laboratory spectra developed across the globe. Such an SSL is the publicly available LUCAS topsoil database with approximately 20,000 soil samples encompassing 23 countries of the European Union. A wide variety of machine learning tools have been applied to the LUCAS SSL to predict some of the soil samples' physicochemical properties with different degrees of accuracy. In this paper, we developed and examined the use of a novel one-dimensional convolutional neural network (CNN) to simultaneously predict ten physicochemical properties of the LUCAS SSL. Leveraging on the use of multiple-input channels it uses as model inputs the absorbance spectra along with some pre-processed spectra developed using standard techniques. Moreover, it exploits the use of local spectral neighborhoods to perform an adaptive error-correction mechanism. This novel localized multi-channel 1-D CNN was applied to all the available physicochemical properties of the LUCAS SSL and was statistically compared with the current state-of-the-art where it was shown to statistically outperform its counterparts, as well as with other CNNs where it exhibited the best performance. In particular, for the mineral soil samples, the RMSE for the Clay content was 4.80% ( R 2 0.86), for soil organic carbon the RMSE was 10.96 g kg−1 ( R 2 0.86), while for total nitrogen the RMSE was 0.66 g kg−1 ( R 2 0.83). [ABSTRACT FROM AUTHOR]