Deep Learning Architectures for Soil Property Prediction

Veres, Matthew; Lacey, Griffin; Taylor, Graham W.

doi:10.1109/crv.2015.15

Cited by 35 publications

(15 citation statements)

References 18 publications

(29 reference statements)

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…The architecture of the optimal 1D-CNN is shown in Table 2. Compared to other published 1D-CNNs 21,23 , our optimised model is very simple. Interestingly, such a simple model achieved the best performance in our study.…”

Section: Resultsmentioning

confidence: 99%

“…To find the set of hyperparameters that generates optimal model performance, one needs to employ hyperparameter tuning, or optimisation 27 (HPO). None of the CNNs for soil spectroscopic modelling reported in the literature [21][22][23][24]26 have undertaken a thorough hyperparameter search, often relying on only manual tuning. Not using HPO may fail to fully exploit the capability of CNNs, and the lack of information about the tuning process can also affect the reliability of the models.…”

Section: Scientific Reportsmentioning

confidence: 99%

“…They are well-known for extracting useful local correlations in the data 20 . To utilise this feature for soil property prediction with spectra, researchers have developed various CNNs with one-dimensional (1D-CNNs) [21][22][23] and twodimensional structures (2D-CNNs) [24][25][26] . The 1D-CNNs process one-dimensional arrays, e.g.…”

mentioning

confidence: 99%

See 2 more Smart Citations

Automated spectroscopic modelling with optimised convolutional neural networks

Shen

Rossel

2021

Sci Rep

View full text Add to dashboard Cite

Convolutional neural networks (CNN) for spectroscopic modelling are currently tuned manually, and the effects of their hyperparameters are not analysed. These can result in sub-optimal models. Here, we propose an approach to tune one-dimensional CNN (1D-CNNs) automatically. It consists of a parametric representation of 1D-CNNs and an optimisation of hyperparameters to maximise a model’s performance. We used a large European soil spectroscopic database to demonstrate our approach for estimating soil organic carbon (SOC) contents. To assess the optimisation, we compared it to random search, and to understand the effects of the hyperparameters, we calculated their importance using functional Analysis of Variance. Compared to random search, the optimisation produced better final results and showed faster convergence. The optimal model produced the most accurate estimates of SOC with $$\hbox {RMSE} = 9.67 \pm 0.51$$ RMSE = 9.67 ± 0.51 (s.d.) and $${R}^2 = 0.89 \pm 0.013$$ R 2 = 0.89 ± 0.013 (s.d.). The hyperparameters associated with model training and architecture critically affected the model’s performance, while those related to the spectral preprocessing had little effect. The optimisation searched through a complex hyperparameter space and returned an optimal 1D-CNN. Our approach simplified the development of 1D-CNNs for spectroscopic modelling by automatically selecting hyperparameters and preprocessing methods. Hyperparameter importance analysis shed light on the tuning process and increased the model’s reliability.

show abstract

Section: Resultsmentioning

confidence: 99%

Section: Scientific Reportsmentioning

confidence: 99%

See 1 more Smart Citation

Automated spectroscopic modelling with optimised convolutional neural networks

Shen

Rossel

2021

Sci Rep

View full text Add to dashboard Cite

show abstract

“…Occasionally, in discussions, the suggestion is made that the very popular “machine” or “deep” learning techniques ( LeCun et al, 2015 ; Willcock et al, 2018 ) could perhaps provide a way to upscale microscale modeling of soils to the macroscopic scale ( Veres et al, 2015 ). Machine learning explores the study and construction of algorithms that can learn from data and make data-driven predictions.…”

Section: Upscaling How?mentioning

confidence: 99%

Emergent Properties of Microbial Activity in Heterogeneous Soil Microenvironments: Different Research Approaches Are Slowly Converging, Yet Major Challenges Remain

et al. 2018

View full text Add to dashboard Cite

Over the last 60 years, soil microbiologists have accumulated a wealth of experimental data showing that the bulk, macroscopic parameters (e.g., granulometry, pH, soil organic matter, and biomass contents) commonly used to characterize soils provide insufficient information to describe quantitatively the activity of soil microorganisms and some of its outcomes, like the emission of greenhouse gasses. Clearly, new, more appropriate macroscopic parameters are needed, which reflect better the spatial heterogeneity of soils at the microscale (i.e., the pore scale) that is commensurate with the habitat of many microorganisms. For a long time, spectroscopic and microscopic tools were lacking to quantify processes at that scale, but major technological advances over the last 15 years have made suitable equipment available to researchers. In this context, the objective of the present article is to review progress achieved to date in the significant research program that has ensued. This program can be rationalized as a sequence of steps, namely the quantification and modeling of the physical-, (bio)chemical-, and microbiological properties of soils, the integration of these different perspectives into a unified theory, its upscaling to the macroscopic scale, and, eventually, the development of new approaches to measure macroscopic soil characteristics. At this stage, significant progress has been achieved on the physical front, and to a lesser extent on the (bio)chemical one as well, both in terms of experiments and modeling. With regard to the microbial aspects, although a lot of work has been devoted to the modeling of bacterial and fungal activity in soils at the pore scale, the appropriateness of model assumptions cannot be readily assessed because of the scarcity of relevant experimental data. For significant progress to be made, it is crucial to make sure that research on the microbial components of soil systems does not keep lagging behind the work on the physical and (bio)chemical characteristics. Concerning the subsequent steps in the program, very little integration of the various disciplinary perspectives has occurred so far, and, as a result, researchers have not yet been able to tackle the scaling up to the macroscopic level. Many challenges, some of them daunting, remain on the path ahead. Fortunately, a number of these challenges may be resolved by brand new measuring equipment that will become commercially available in the very near future.

show abstract

“…The sources of approximation models based on determination of the NIR absorbance connected with the vector dimensionality reduction are being searched for in statistical methods (regression with PCA transformation, stepwise regression, partial least-square regression PLSRa dominating approach in the spectral response analysis) and in machine learning methods (MultiLayer Perceptrons-MLP, Radial Basis Functions-RBF, Support Vector Machines -SVM models, stack autoencoders, convolution networks in regression applications, random trees, including the so-called random forests). In addition to raw data (readings of reflectance or absorbance vectors in the NIR spectrum), the input data include transformed vectors like those subjected to PCA reduction, first and second derivatives of spectral response vectors and data filtered by autoen-STANIS£AW GRUSZCZYÑSKI* Spectral response of soil samples in the near-infrared range coders (Fuentes et al 2012;Qiu et al 2014;Shi et al 2015;Veres et al 2015;Zhang et al 2016;Conforti et al 2018;Mohamed et al 2018). It is not possible to indicate objectively the best approach.…”

Section: Introductionmentioning

confidence: 99%

Prediction of soil properties with machine learning models based on the spectral response of soil samples in the near infrared range

Gruszczyński

2019

Soil Science Annual

View full text Add to dashboard Cite

One of the basic methods for soil analysis time and cost reduction is using soil sample spectral response in laboratory conditions. The problem with this method lies in determining the relationship between the shape of the soil spectral response and soil physical or chemical properties. The LUCAS soil database collected by the EU’s ESDAC research centre is good material to analyse the relationship between the soil properties and the near infrared (NIR) spectral response. The modelling described in the paper is based on these data. The analysis of the impact of soil properties configuration on absorbance levels in various NIR spectrum ranges was conducted using the stepwise regression models with the properties, properties squared and products of properties being explanatory variables. The analysis of partial correlation of soil properties values with absorbance values and absorbance derivative in the entire spectral range was conducted in order to evaluate the impact of the absorbance transformation (the first derivative of absorbance vector) on the change of significance of relationship with properties values. The Multi Layer Perceptron (MLP) models were used to estimate the absorbance relationship with single soil features. Soil property modelling based on the selection and transformation algorithm of raw values and first and second absorbance derivatives was also conducted along with the suitability evaluation of such models in building digital soil maps. The absorbance is affected by a limited number of tested soil features like pH, texture, content of carbonates, SOC, N, and CEC; P and K contents have, in case of this research, a negligible impact. The NIR methodology can be suitable in conditions of limited soil variation and particularly in development of thematic soil maps.

show abstract

Deep Learning Architectures for Soil Property Prediction

Cited by 35 publications

References 18 publications

Automated spectroscopic modelling with optimised convolutional neural networks

Automated spectroscopic modelling with optimised convolutional neural networks

Emergent Properties of Microbial Activity in Heterogeneous Soil Microenvironments: Different Research Approaches Are Slowly Converging, Yet Major Challenges Remain

Prediction of soil properties with machine learning models based on the spectral response of soil samples in the near infrared range

Contact Info

Product

Resources

About