`resemble`

Memory-Based Learning in Spectral Chemometrics*Last update: 30.08.2022*

Think Globally, Fit Locally (Saul and Roweis, 2003)

The `resemble`

package provides high-performing
functionality for data-driven modeling (including local modeling),
nearest-neighbor search and orthogonal projections in spectral data.

A new vignette for `resemble`

explaining its core
functionality is available at:
https://CRAN.R-project.org/package=prospectr/vignettes/vignette.html

The core functionality of the package can be summarized into the following functions:

** mbl**: implements memory-based learning
(MBL) for modeling and predicting continuous response variables. For
example, it can be used to reproduce the famous LOCAL algorithm proposed
by Shenk et al. (1997). In general, this function allows you to easily
customize your own MBL regression-prediction method.

** dissimilarity**: Computes dissimilarity
matrices based on various methods (e.g. Euclidean, Mahalanobis, cosine,
correlation, moving correlation, Spectral information divergence,
principal components dissimilarity and partial least squares
dissimilarity).

** ortho_projection**: A function for
dimensionality reduction using either principal component analysis or
partial least squares (a.k.a projection to latent structures).

** search_neighbors**: A function to
efficiently retrieve from a reference set the k-nearest neighbors of
another given data set.

During the recent lockdown we invested some of our free time to come
up with a new version of our package. This new `resemble`

2.0
comes with MAJOR improvements and new functions! For these improvements
major changes were required. The most evident changes are in the
function and argument names. These have been now adapted to properly
follow the tydiverse style
guide. A number of changes have been implemented for the sake of
computational efficiency. These changes are documented in
`inst\changes.md`

.

New interesing functions and fucntionality are also available, for
example, the `mbl()`

function now allows sample spiking,
where a set of reference observations can be forced to be included in
the neighborhhoods of each sample to be predicted. The
`serach_neighbors()`

function efficiently retrieves from a
refence set the k-nearest neighbors of another given data set. The
`dissimilarity()`

function computes dissimilarity matrices
based on various metrics.

If you want to install the package and try its functionality, it is
very simple, just type the following line in your `R`

console:

`install.packages('resemble')`

If you do not have the following packages installed, it might be good to update/install them first

```
install.packages('Rcpp')
install.packages('RcppArmadillo')
install.packages('foreach')
install.packages('iterators')
```

**Note**: Apart from these packages we stronly recommend
to download and install Rtools https://cran.r-project.org/bin/windows/Rtools/).
This is important for obtaining the proper C++ toolchain that might be
needed for `resemble`

.

Then, install `resemble`

You can also install the development version of `resemble`

directly from github using `devtools`

:

`devtools::install_github("l-ramirez-lopez/resemble")`

NOTE: in some MAC Os it is still recommended to install
`gfortran`

and `clang`

from here. Even for R
>= 4.0. For more info, check this issue.

After installing `resemble`

you should be also able to run
the following lines:

```
library(resemble)
library(tidyr)
library(prospectr)
data(NIRsoil)
# Proprocess the data
NIRsoil <- NIRsoil[NIRsoil$CEC %>% complete.cases(),]
wavs <- as.numeric(colnames(NIRsoil$spc))
NIRsoil$spc_p <- NIRsoil$spc %>%
standardNormalVariate() %>%
resample(wavs, seq(min(wavs), max(wavs), by = 11)) %>%
savitzkyGolay(p = 1, w = 5, m = 1)
# split into calibration/training and test
train_x <- NIRsoil$spc_p[as.logical(NIRsoil$train), ]
train_y <- NIRsoil$CEC[as.logical(NIRsoil$train)]
test_x <- NIRsoil$spc_p[!as.logical(NIRsoil$train), ]
test_y <- NIRsoil$CEC[!as.logical(NIRsoil$train)]
# Use MBL as in Ramirez-Lopez et al. (2013)
sbl <- mbl(
Xr = train_x, Yr = train_y, Xu = test_x,
k = seq(50, 130, by = 20),
method = local_fit_gpr(),
control = mbl_control(validation_type = "NNv")
)
sbl
plot(sbl)
get_predictions(sbl)
```

Figure 1. Standard plot of the results of the
** mbl** function.

`resemble`

implements functions dedicated to non-linear modelling of complex
visible and infrared spectral data based on memory-based learning (MBL,
*a.k.a* instance-based learning or local modelling in the
chemometrics literature). The package also includes functions for:
computing and evaluate spectral dissimilarity matrices, projecting the
spectra onto low dimensional orthogonal variables, spectral neighbor
search, etc.

To expand a bit more the explanation on the `mbl`

function, let’s define first the basic input data:

**Reference (training) set**: Dataset with*n*reference samples (e.g. spectral library) to be used in the calibration of spectral models. Xr represents the matrix of samples (containing the spectral predictor variables) and Yr represents a response variable corresponding to Xr.**Prediction set**: Data set with*m*samples where the response variable (Yu) is unknown. However it can be predicted by applying a spectral model (calibrated by using Xr and Yr) on the spectra of these samples (Xu).

To predict each value in Yu, the `mbl`

function takes each
sample in Xu and searches in Xr for its *k*-nearest neighbours
(most spectrally similar samples). Then a (local) model is calibrated
with these (reference) neighbours and it immediately predicts the
correspondent value in Yu from Xu. In the function, the
*k*-nearest neighbour search is performed by computing spectral
dissimilarity matrices between observations. The `mbl`

function offers the following regression options for calibrating the
(local) models:

** 'gpr'**: Gaussian process with linear
kernel.

** 'pls'**: Partial least squares.

** 'wapls'**: Weighted average partial least
squares (Shenk et al., 1997).

Figure 2 illustrates the basic steps in MBL for a set of five observations.

Figure 2. Example of the main steps in memory-based learning for predicting a response variable in five different observations based on set of p-dimesnional variables.

Simply type and you will get the info you need:

`citation(package = "resemble")`

`resemble`

2022.03: Ng et al., 2022 uses MBL (implemented in

`resemble`

) to asses the feasibility of quantifying a number of soil properties from IR spectra. They also show that MBL achieved better accuracy than Cubist regression.2022.02: Li et al., 2022 show how useful the combination of MBL and spiking (implemented in

`resemble`

) can be to accurately predict soil properties from NIR data in China.2021.12: Yu et al., 2022 uses MBL with External Parameter Orthogonalization to predict soil properties in in the field.

2021.10: In this paper we use MBL to predict soil properties in Africa.

2020.08: Charlotte Rivard shows how to use MBL in IR spectroscopy here.

2020.04: Tsakiridis et al. (2020), used the optimal principal components dissimilarity method implemented in

`resemble`

in combination with convolutional neural networks for simultaneous prediction of soil properties from vis-NIR spectra.2019-04: Tziolas et al. (2019), used

`resemble`

to investigate on improved MBL methods for quantitative predictions of soil properties using NIR spectroscopy and geographical information.2019.03,08: Tsakiridis et al. (2019a) and Tsakiridis et al. (2019b), compared several machine learning methods for predictive soil spectroscopy and show that MBL

`resemble`

offers highly competive results.2020.01: Sanderman et al., (2020) used

`resemble`

for the prediction of soil health indicatorsin the United States.2019-03: I published a scientific paper were we used memory-based learning (MBL) for digital soil mapping. Here we use MBL to remove local calibration outliers rather than using this approach to overcome the typical complexity of large spectral datasets. (Ramirez‐Lopez, L., Wadoux, A. C., Franceschini, M. H. D., Terra, F. S., Marques, K. P. P., Sayão, V. M., & Demattê, J. A. M. (2019). Robust soil mapping at the farm scale with vis–NIR spectroscopy. European Journal of Soil Science. 70, 378–393).

2019-01: In this scientific paper we use

`resemble`

to model MIR spectra from a continental soil spectral library in United States. (Dangal, S.R., Sanderman, J., Wills, S. and Ramirez-Lopez, L., 2019. Accurate and Precise Prediction of Soil Properties from a Large Mid-Infrared Spectral Library. Soil Systems, 3(1), p.11).2019-03: Jaconi et al. (2019) implemented a memory-based learning algorithm (using

`resemble`

) to conduct accurate NIR predictions of soil texture at National scale in Germany. (Jaconi, A., Vos, C. and Don, A., 2019. Near infrared spectroscopy as an easy and precise method to estimate soil texture. Geoderma, 337, pp.906-913).2018-12: Chen, et al. (2018) implemented a memory-based learning algorithm (using

`resemble`

) to improve the accuracy of NIR predictions of soil organic matter in China. (Hong, Y., Chen, S., Liu, Y., Zhang, Y., Yu, L., Chen, Y., Liu, Y., Cheng, H. and Liu, Y. 2019. Combination of fractional order derivative and memory-based learning algorithm to improve the estimation accuracy of soil organic matter by visible and near-infrared spectroscopy. Catena, 174, pp.104-116).2018-11: In this recent scientific paper the authors used

`resemble`

to predict the chemoical composition of Common Beans in Spain. (Rivera, A., Plans, M., Sabaté, J., Casañas, F., Casals, J., Rull, A., & Simó, J. (2018). The Spanish core collection of common beans (Phaseolus vulgaris L.): an important source of variability for breeding chemical composition. Frontiers in Plant Science, 9).2018-07: Another use-case of

`resemble`

is presented by Gholizadeh et al.(2018) for a soil science application in Czech Republic. (Gholizadeh, A., Saberioon, M., Carmon, N., Boruvka, L. and Ben-Dor, E., 2018. Examining the Performance of PARACUDA-II Data-Mining Engine versus Selected Techniques to Model Soil Carbon from Reflectance Spectra. Remote Sensing, 10(8), p.1172).2018-01: Dotto, et al. (2018) have implemented memory-based learning with

`resemble`

to accurately predict soil organic Carbon at a region in Brazil. (Dotto, A. C., Dalmolin, R. S. D., ten Caten, A., & Grunwald, S. (2018). A systematic study on the application of scatter-corrective and spectral-derivative preprocessing for multivariate prediction of soil organic carbon by Vis-NIR spectra. Geoderma, 314, 262-274).2017-11: Here the authors predicted brix values in differet food products using memory-based learning implemented with

`resemble`

. (Kopf, M., Gruna, R., Längle, T. and Beyerer, J., 2017, March. Evaluation and comparison of different approaches to multi-product brix calibration in near-infrared spectroscopy. In OCM 2017-Optical Characterization of Materials-conference proceedings (p. 129). KIT Scientific Publishing).2016-05: In this scientific paper the authors sucesfully used

`resemble`

to predict soil organic carbon content at national scale in France. (Clairotte, M., Grinand, C., Kouakoua, E., Thébault, A., Saby, N. P., Bernoux, M., & Barthès, B. G. (2016). National calibration of soil organic carbon concentration using diffuse infrared reflectance spectroscopy. Geoderma, 276, 41-52).2016-04: This paper shows some interesting results on applying memory-based learning to predict soil properties.

2016-04: In some recent entries of this blog, the author shows some exmaples on the use

`resemble`

2016-02: As promised,

`resemble 1.2 (alma-de-coco)`

is now available on CRAN.2016-01: The version 1.2 (alma-de-coco) has been submitted to CRAN and is available from the github repository!

2015-11: A pre-release of the version 1.2.0 (1.2.0.9000 alma-de-coco) is now available!

`resemble`

is now faster! Some critical functions (e.g. pls and gaussian process regressions were re-written in C++ using`Rcpp`

. This time the new version will be available at CRAN very soon!.2015-11 Well, the version 1.1.3 was never released on CRAN since we decided to carry out major improvements in terms of computational performance.

2014-10: A pre-release of the version 1.1.3 of the package is already available at this website. We hope it will be available at CRAN very soon!

2014-06: Check this video where a renowned NIR scientist talks about local calibrations.

2014-04: A short note on the resemble and prospectr packages was published in this newsletter. There we provide some examples on representative subset selection and on how to reproduce the LOCAL and spectrum-based learner algorithms. In those examples the dataset of the Chemometric challenge of ‘Chimiométrie 2006’ (included in the prospectr package) is used.

2014-03: The package released on CRAN!

- Check our
other project called
`prospectr`

. - Check this presentation in which we used the resemble package to predict soil attributes from large scale soil spectral libraries. https://www.fao.org/fileadmin/user_upload/GSP/docs/Spectroscopy_dec13/SSW2013_f.pdf

You can send an e-mail to the package maintainer (ramirez.lopez.leo@gmail.com) or create an issue on github.

Lobsey, C. R., Viscarra Rossel, R. A., Roudier, P., & Hedley, C. B. 2017. rs-local data-mines information from spectral libraries to improve local calibrations. European Journal of Soil Science, 68(6), 840-852.

Ramirez-Lopez, L., Behrens, T., Schmidt, K., Stevens, A., Dematte, J.A.M., Scholten, T. 2013. The spectrum-based learner: A new local approach for modeling soil vis-NIR spectra of complex data sets. Geoderma 195-196, 268-279.

Saul, L. K., & Roweis, S. T. 2003. Think globally, fit locally: unsupervised learning of low dimensional manifolds. Journal of machine learning research, 4(Jun), 119-155.

Shenk, J., Westerhaus, M., and Berzaghi, P. 1997. Investigation of a LOCAL calibration procedure for near infrared instruments. Journal of Near Infrared Spectroscopy, 5, 223-232.