Package 'soilFlux'

Description

Add texture classification column to a data frame

Usage

add_texture(
  df,
  sand_col = "sand_total",
  silt_col = "silt",
  clay_col = "clay",
  out_col = "Texture"
)

Arguments

Value

The input data frame with an additional out_col column.

Examples

df <- data.frame(sand_total = c(70, 20), silt = c(15, 50), clay = c(15, 30))
add_texture(df)

Description

Scales df[, scaler$cols] using the precomputed min/range. Returns a numeric matrix with the same column order as scaler$cols.

Usage

apply_minmax(df, scaler)

Arguments

Value

A numeric matrix scaled to approximately [0, 1] per column. Columns correspond to scaler$cols.

Examples

df_train <- data.frame(sand = c(20, 40, 60), clay = c(30, 20, 10))
sc       <- fit_minmax(df_train, c("sand", "clay"))
df_new   <- data.frame(sand = c(50, 25), clay = c(15, 28))
apply_minmax(df_new, sc)

Description

Generates four sets of collocation points (random soil-property vectors and associated pF values) used to evaluate the physics constraints during training.

Usage

build_residual_sets(
  df_raw,
  x_inputs,
  S1 = 1500L,
  S2 = 500L,
  S3 = 500L,
  S4 = 1500L,
  pF_lin_min = 5,
  pF_lin_max = 7.6,
  pF0_pos = 6.2,
  pF1_neg = 7.6,
  pF_sat_min = -2,
  pF_sat_max = -0.3,
  seed = 123L
)

Arguments

Value

A named list with four data frames: set1, set2, set3, set4. Each data frame has one row per collocation point, with columns corresponding to x_inputs (sampled uniformly within training range) and a pF column.

Examples

## Not run: 
df <- data.frame(
  clay       = c(20, 30, 10),
  silt       = c(30, 40, 20),
  sand_total = c(50, 30, 70),
  Depth_num  = c(15, 30, 60)
)
sets <- build_residual_sets(df, c("clay", "silt", "sand_total", "Depth_num"),
                            S1 = 50L, S2 = 20L, S3 = 20L, S4 = 50L)

## End(Not run)

Description

Constructs a Keras model implementing the monotone-integral architecture of Norouzi et al. (2025). The returned list contains two models that share weights:

• theta_model — full prediction model (pF query + covariates → theta).

• param_model — extracts the saturated water content (theta_s) from covariates only.

Usage

build_swrc_model(n_covariates, hidden = c(128L, 64L), dropout = 0.1, K = 64L)

Arguments

Value

A named list:

theta_model

Keras model: inputs ⁠[Xseq_knots, pf_norm]⁠, output shape ⁠[N, 2]⁠ (theta_hat, theta_s).

param_model

Keras model: input Xseq_knots, output shape ⁠[N, 1]⁠ (theta_s only).

K

The K value used.

dk

The knot spacing 1 / (K - 1).

knot_grid

Numeric vector of knot positions.

Examples

## Not run: 
mod <- build_swrc_model(n_covariates = 9L)

## End(Not run)

Description

Returns the USDA texture class name for each row based on the sand, silt, and clay fractions. Inputs are expected in per-cent (0–100) and must sum to approximately 100.

Usage

classify_texture(sand, silt, clay, tol = 1)

Arguments

Value

Character vector of USDA texture class names. Returns NA for rows where values are missing or do not sum to approximately 100.

Examples

classify_texture(sand = c(70, 20, 10, 40),
                silt = c(15, 50, 30, 40),
                clay = c(15, 30, 60, 20))

Description

Functions to prepare soil data for input to the CNN1D monotone-integral model, including 3-D sequence array construction and observation matrix creation.

Description

A convenience wrapper that calls predict_swrc() and swrc_metrics().

Usage

evaluate_swrc(object, newdata, obs_col = "theta_n")

Arguments

Value

A tibble with columns R2, RMSE, MAE, n.

Description

Computes per-column minimum and range from df[, cols]. Constant columns (range == 0) are assigned a range of 1 to avoid division by zero.

Usage

fit_minmax(df, cols)

Arguments

Value

A list with elements:

min

Named numeric vector of per-column minima.

rng

Named numeric vector of per-column ranges.

cols

The character vector cols (stored for later use).

Examples

df <- data.frame(sand = c(20, 40, 60), clay = c(30, 20, 10))
sc <- fit_minmax(df, c("sand", "clay"))
sc

Description

The main user-facing function for training. Given prepared training and (optionally) validation data, it builds the model, creates physics residual sets, runs the training loop with early stopping, and returns a fitted object for prediction and evaluation.

Usage

fit_swrc(
  train_df,
  x_inputs,
  val_df = NULL,
  hidden = c(128L, 64L),
  dropout = 0.1,
  lr = 0.001,
  epochs = 80L,
  batch_size = 256L,
  patience = 5L,
  K = 64L,
  lambdas = norouzi_lambdas("norouzi"),
  S1 = 1500L,
  S2 = 500L,
  S3 = 500L,
  S4 = 1500L,
  pF_lin_min = 5,
  pF_lin_max = 7.6,
  pF0_pos = 6.2,
  pF1_neg = 7.6,
  pF_sat_min = -2,
  pF_sat_max = -0.3,
  wet_split_cm = 4.2,
  w_wet = 1,
  w_dry = 1,
  pf_left = -2,
  pf_right = 7.6,
  seed = 123L,
  verbose = TRUE
)

Arguments

Value

An S3 object of class swrc_fit, a named list containing:

theta_model

The fitted Keras model.

param_model

The theta_s extractor model.

x_inputs

Covariate names used.

scaler

Fitted min-max scaler.

K

Number of knot points.

dk

Knot spacing.

knot_grid

Knot positions in [0, 1].

pf_left,pf_right

pF domain boundaries.

theta_factor

Unit multiplier for theta.

best_epoch

Epoch at which validation loss was minimised.

lambdas

Loss weights used during training.

history

Data frame of per-epoch training/validation losses.

Examples

## Not run: 
if (reticulate::py_module_available("tensorflow")) {
  df  <- prepare_swrc_data(swrc_example, depth_col = "depth")
  fit <- fit_swrc(df,
                  x_inputs = c("clay", "silt", "bd_gcm3", "soc", "Depth_num"),
                  epochs = 2L, verbose = FALSE)
}

## End(Not run)

Description

If the median raw value is > 10 it is assumed to be in kg/m3 and is divided by 100 to convert to g/cm3.

Usage

fix_bd_units(bd_raw)

Arguments

Value

Numeric vector in g/cm3.

Examples

fix_bd_units(c(1.2, 1.45, 1.3))   # already g/cm3
fix_bd_units(c(120, 145, 130))    # kg/m3 -> g/cm3

Description

Convert pF to matric head (cm)

Usage

head_from_pf(pf)

Arguments

Value

Numeric vector of matric head values in cm.

Examples

head_from_pf(c(0, 1, 2, 3, 4.2))

Description

Convenience wrapper: converts head to pF then normalises.

Usage

head_normalize(h_cm, pf_left = -2, pf_right = 7.6)

Arguments

Value

Numeric vector in ⁠[0, 1]⁠.

Examples

head_normalize(c(1, 10, 100, 15850))

Description

Converts scaled values back to original units.

Usage

invert_minmax(X_scaled, scaler)

Arguments

Value

Numeric matrix in the original (unscaled) units.

Examples

df <- data.frame(sand = c(20, 40, 60), clay = c(30, 20, 10))
sc <- fit_minmax(df, c("sand", "clay"))
Xs <- apply_minmax(df, sc)
invert_minmax(Xs, sc)

Description

Functions to persist a fitted swrc_fit object to disk and reload it for later use (prediction, spatial mapping, etc.).

Description

Reconstructs the CNN1D Keras model from the saved weights and metadata and returns a swrc_fit-compatible list that can be passed to predict_swrc(), predict_swrc_dense(), etc.

Usage

load_swrc_model(dir = "./models/swrc", name = "swrc_model")

Arguments

Value

A swrc_fit object (without history or param_model).

Examples

## Not run: 
if (reticulate::py_module_available("tensorflow")) {
  df  <- prepare_swrc_data(swrc_example, depth_col = "depth")
  fit <- fit_swrc(df,
                  x_inputs = c("clay", "silt", "bd_gcm3", "soc", "Depth_num"),
                  epochs = 2L, verbose = FALSE)
  save_swrc_model(fit, dir = tempdir(), name = "model_test")
  fit2 <- load_swrc_model(tempdir(), "model_test")
}

## End(Not run)

Description

Functions to compute R², RMSE, and MAE for soil water content predictions.

Description

Build the physics-informed 1-D CNN with a monotone integral output layer, as described in Norouzi et al. (2025).

Details

Architecture

The model takes two inputs:

1. Xseq_knots: a 3-D tensor of shape ⁠[N, K, p+1]⁠ — for each observation, p scaled covariates are broadcast across K knot positions, and the knot positions themselves form the last channel.

2. pf_norm: a 2-D tensor of shape ⁠[N, 1]⁠ — the query pF value normalised to ⁠[0, 1]⁠.

The output satisfies:

$\hat{\theta}(pF) = \theta_s - \int_0^{pF} \text{softplus}(s(t))\,dt$

where $s(t)$ is a 1-channel convolutional output. Monotone decrease is guaranteed by construction because the integrand is always positive.

References

Norouzi, A. M., et al. (2025). Physics-Informed Neural Networks for Estimating a Continuous Form of the Soil Water Retention Curve. Journal of Hydrology.

Description

Table 1 of Norouzi et al. (2025) defines six loss-weight hyperparameters. This function returns them as a named list that can be passed to fit_swrc() and compute_physics_loss().

Usage

norouzi_lambdas(config = c("norouzi", "smooth"))

Arguments

Value

A named list:

lambda_wet

Weight for wet-end data loss (lambda1 = 1).

lambda_dry

Weight for dry-end data loss (lambda2 = 10).

lambda3

S1 dry-end linearity (lambda3 = 1 or 10).

lambda4

S2 non-negativity at pF0 (lambda4 = 1000).

lambda5

S3 non-positivity at pF1 (lambda5 = 1000).

lambda6

S4 saturated-plateau flatness (lambda6 = 1).

Examples

norouzi_lambdas()
norouzi_lambdas("smooth")

Description

Accepts strings of the form "0-5", "5-15", "100", etc. and returns the numeric midpoint and a human-readable label (e.g. "0-5 cm").

Usage

parse_depth(s)

Arguments

Value

A named list with elements:

mid

Numeric midpoint in cm.

label

Character label, e.g. "0-5 cm".

Examples

parse_depth("0-5")
parse_depth("100-200")
parse_depth("30")

Description

Applies parse_depth() row-wise and appends Depth_num and Depth_label columns.

Usage

parse_depth_column(df, depth_col = "depth")

Arguments

Value

The input data frame with two extra columns: Depth_num (numeric midpoint) and Depth_label (factor, ordered by depth).

Examples

df <- data.frame(depth = c("0-5", "5-15", "15-30"), x = 1:3)
parse_depth_column(df, "depth")

Description

Convert matric head (cm) to pF

Usage

pf_from_head(h_cm)

Arguments

Value

Numeric vector of pF values ($\log_{10}(h)$).

Examples

pf_from_head(c(1, 10, 100, 1000, 15850))

Description

Maps the pF domain ⁠[pf_left, pf_right]⁠ linearly to ⁠[0, 1]⁠. Values outside the domain are clipped.

Usage

pf_normalize(pf, pf_left = -2, pf_right = 7.6)

Arguments

Value

Numeric vector in ⁠[0, 1]⁠.

Examples

pf_normalize(c(-2, 0, 4, 7.6))

Description

Functions for defining, building, and evaluating the four physics-based residual constraints of Norouzi et al. (2025).

References

Norouzi, A. M., et al. (2025). Physics-Informed Neural Networks for Estimating a Continuous Form of the Soil Water Retention Curve. Journal of Hydrology.

Description

Creates a scatter plot of predicted vs. observed theta with a 1:1 line and optional regression line, optionally faceted by a grouping variable.

Usage

plot_pred_obs(
  df,
  obs_col = "theta_n",
  pred_col = "theta_predicted",
  group_col = NULL,
  show_lm = TRUE,
  show_stats = TRUE,
  ncol = 5L,
  base_size = 12,
  point_alpha = 0.25,
  title = NULL
)

Arguments

Value

A ggplot object.

Examples

## Not run: 
df_plot <- data.frame(
  theta_n         = c(0.30, 0.25, 0.20, 0.15, 0.10),
  theta_predicted = c(0.28, 0.26, 0.22, 0.14, 0.11)
)
plot_pred_obs(df_plot)

## End(Not run)

Description

Creates a ggplot2 figure showing continuous SWRC predictions (lines) optionally overlaid with observed data points.

Usage

plot_swrc(
  pred_curves,
  obs_points = NULL,
  curve_col = "theta",
  obs_col = "theta_n",
  group_col = "PEDON_ID",
  facet_row = NULL,
  facet_col = NULL,
  x_limits = NULL,
  y_limits = c(-0.25, 7.75),
  line_colour = "steelblue4",
  point_colour = "black",
  base_size = 12,
  title = NULL
)

Arguments

Value

A ggplot object.

Examples

## Not run: 
pred_curves <- data.frame(
  PEDON_ID = rep(c("P1", "P2"), each = 5),
  pF       = rep(c(0, 1, 2, 3, 4), 2),
  theta    = c(0.42, 0.36, 0.28, 0.18, 0.08,
               0.38, 0.32, 0.25, 0.16, 0.07)
)
plot_swrc(pred_curves, group_col = "PEDON_ID")

## End(Not run)

Description

Creates a bar chart comparing R², RMSE, and MAE across multiple models or configurations.

Usage

plot_swrc_metrics(
  metrics_df,
  model_col = "model",
  palette = "Blues",
  base_size = 12,
  title = NULL
)

Arguments

Value

A ggplot object.

Examples

## Not run: 
m1 <- swrc_metrics(c(0.30, 0.25, 0.20), c(0.28, 0.26, 0.22)) |>
  dplyr::mutate(model = "Model 1")
m2 <- swrc_metrics(c(0.30, 0.25, 0.20), c(0.29, 0.24, 0.21)) |>
  dplyr::mutate(model = "Model 2")
plot_swrc_metrics(dplyr::bind_rows(m1, m2))

## End(Not run)

Description

Plots the per-epoch training and (optionally) validation loss from the history slot of a swrc_fit object. A log(1 + x) y-axis is used so that both the (larger) composite training loss and the (small) validation MSE remain clearly visible on the same panel.

Usage

plot_training_history(
  fit,
  loss_col = "loss",
  val_col = "val_mse",
  log_scale = TRUE,
  base_size = 12
)

Arguments

Value

A ggplot object.

Description

Functions for visualising soil water retention curves, predicted vs. observed scatter plots, and model performance metrics.

Description

Functions for generating soil water content predictions from a fitted swrc_fit object, both at specific pF points and as dense continuous curves.

Value

See predict_swrc() and predict_swrc_dense() for return value details.

Description

Given a fitted swrc_fit object and a new data frame of soil properties, returns predicted volumetric water content at each supplied pF (or matric head) value.

Usage

predict_swrc(object, newdata, pf = NULL, heads = NULL, ...)

Arguments

Value

A numeric vector of predicted theta values (m3/m3), one per row in newdata.

Examples

## Not run: 
if (reticulate::py_module_available("tensorflow")) {
  df   <- prepare_swrc_data(swrc_example, depth_col = "depth")
  fit  <- fit_swrc(df,
                   x_inputs = c("clay", "silt", "bd_gcm3", "soc", "Depth_num"),
                   epochs = 2L, verbose = FALSE)
  pred <- predict_swrc(fit, newdata = df)
}

## End(Not run)

Description

For each unique (PEDON_ID × depth) profile in newdata, predicts theta across a dense grid of pF values and returns a tidy long-format tibble.

Usage

predict_swrc_dense(
  object,
  newdata,
  n_points = 1000L,
  pf_range = NULL,
  id_cols = c("PEDON_ID", "Depth_num", "Depth_label", "Texture")
)

Arguments

Value

A tibble with columns: all id_cols present in newdata, pF, matric_head, and theta (predicted volumetric water content in m3/m3).

Examples

## Not run: 
if (reticulate::py_module_available("tensorflow")) {
  df    <- prepare_swrc_data(swrc_example, depth_col = "depth")
  fit   <- fit_swrc(df,
                    x_inputs = c("clay", "silt", "bd_gcm3", "soc", "Depth_num"),
                    epochs = 2L, verbose = FALSE)
  dense <- predict_swrc_dense(fit, newdata = df, n_points = 50)
}

## End(Not run)

Description

Uses the param_model (which maps covariate inputs to theta_s) to extract the modelled saturated water content for each row of newdata.

Usage

predict_theta_s(object, newdata)

Arguments

Value

Numeric vector of theta_s values (m3/m3).

Description

Dispatches to predict_swrc().

Usage

## S3 method for class 'swrc_fit'
predict(object, newdata, pf = NULL, heads = NULL, ...)

Arguments

Value

A numeric vector of predicted volumetric water content values (m3/m3), one per row in newdata.

Description

A convenience wrapper that:

1. Renames columns to standard names.

2. Parses the depth column.

3. Fixes bulk-density units.

4. Detects and normalises volumetric water content units.

5. Computes per-profile maximum theta.

6. Drops rows with missing key variables.

Usage

prepare_swrc_data(
  df,
  x_cols = NULL,
  depth_col = "depth",
  fix_bd = TRUE,
  fix_theta = TRUE
)

Arguments

Value

A tibble with standardised columns plus Depth_num, Depth_label, bd_gcm3, theta_n (normalised WC), and theta_max_n (per-profile maximum theta).

Examples

## Not run: 
df <- data.frame(
  ID           = rep("P01", 3),
  sand_pct     = c(50, 50, 50),
  silt         = c(30, 30, 30),
  clay         = c(20, 20, 20),
  soc          = c(1.2, 1.2, 1.2),
  bd_gcc       = c(1.3, 1.3, 1.3),
  matric_head  = c(10, 100, 1000),
  theta        = c(0.40, 0.30, 0.20),
  depth        = c("0-30", "0-30", "0-30")
)
df_prep <- prepare_swrc_data(df,
  x_cols = c(PEDON_ID = "ID", sand = "sand_pct", bd = "bd_gcc",
             water_content = "theta"))

## End(Not run)

Description

Print method for swrc_fit

Usage

## S3 method for class 'swrc_fit'
print(x, ...)

Arguments

Value

Invisibly returns x (called for its side effect of printing a summary of the fitted model to the console).

Description

Saves the Keras model weights as an HDF5 file and the R metadata (scalers, hyperparameters, etc.) as an .rds file inside dir.

Usage

save_swrc_model(fit, dir, name = "swrc_model")

Arguments

Value

Invisibly returns a named list with paths to the two files:

weights_path

Path to the .weights.h5 file.

meta_path

Path to the .rds metadata file.

Examples

## Not run: 
if (reticulate::py_module_available("tensorflow")) {
  df  <- prepare_swrc_data(swrc_example, depth_col = "depth")
  fit <- fit_swrc(df,
                  x_inputs = c("clay", "silt", "bd_gcm3", "soc", "Depth_num"),
                  epochs = 2L, verbose = FALSE)
  save_swrc_model(fit, dir = tempdir(), name = "model_test")
}

## End(Not run)

Description

Fit and apply a column-wise min-max scaler to a data frame or matrix, mapping each feature to [0, 1].

Description

Summary method for swrc_fit

Usage

## S3 method for class 'swrc_fit'
summary(object, ...)

Arguments

Value

Invisibly returns object (called for its side effect of printing a detailed summary including covariates, training parameters, and loss weights to the console).

Description

A synthetic but physically realistic soil characterisation dataset generated to illustrate the functions in soilFlux. The data mimics the structure of the Florida Soil Characterization Database (FSCD) used in Rodrigues et al. / Norouzi et al. (2025), with van Genuchten curves used to produce internally consistent water-content observations.

Usage

swrc_example

Format

A data frame with 4 800 rows and 14 columns:

PEDON_ID

Character. Unique soil profile identifier (e.g. "P0001").

sand_total

Numeric. Total sand content (%).

silt

Numeric. Silt content (%).

clay

Numeric. Clay content (%).

soc

Numeric. Soil organic carbon (%).

bd

Numeric. Bulk density (g/cm3).

sand_vf

Numeric. Very fine sand fraction (%).

sand_f

Numeric. Fine sand fraction (%).

sand_m

Numeric. Medium sand fraction (%).

sand_c

Numeric. Coarse sand fraction (%).

matric_head

Numeric. Matric head (cm H2O, positive).

water_content

Numeric. Volumetric water content (m3/m3). Approximately 1% of values are NA to simulate missing measurements.

depth

Character. Depth interval (e.g. "0-5", "5-15", etc.).

Texture

Character. USDA texture class (one of: Sand, Loamy Sand, Sandy Loam, Loam, Silt Loam, Clay Loam, Clay).

Details

The dataset contains 120 unique profiles (PEDON_ID) across five depth intervals (0–5, 5–15, 15–30, 30–60, 60–100 cm) and eight matric-head points per depth (pF approximately 0, 1, 1.5, 2, 2.5, 3, 4.2, 7). Profiles are evenly distributed across seven USDA texture classes.

Water-content observations were generated with the van Genuchten (1980) equation using parameters that vary realistically with texture and depth, then small Gaussian noise was added.

Source

Synthetic dataset generated by data-raw/create_example_data.R.

References

van Genuchten, M. T. (1980). A closed-form equation for predicting the hydraulic conductivity of unsaturated soils. Soil Science Society of America Journal, 44(5), 892–898.

Examples

data("swrc_example")
head(swrc_example)
table(swrc_example$Texture[!duplicated(swrc_example$PEDON_ID)])

# Prepare for modelling
df <- prepare_swrc_data(swrc_example, depth_col = "depth")

Description

Returns R², RMSE, and MAE between observed and predicted volumetric water content (or any continuous response).

Usage

swrc_metrics(observed, predicted, na.rm = TRUE)

Arguments

Value

A tibble::tibble() with one row and columns R2, RMSE, MAE, n (number of non-missing pairs).

Examples

obs  <- c(0.30, 0.25, 0.20, 0.15, 0.10)
pred <- c(0.28, 0.26, 0.22, 0.14, 0.11)
swrc_metrics(obs, pred)

Description

Applies swrc_metrics() within each level of one or more grouping variables, returning a tidy data frame.

Usage

swrc_metrics_by_group(df, obs_col, pred_col, group_col, na.rm = TRUE)

Arguments

Value

A tibble with one row per group and columns: grouping variables, R2, RMSE, MAE, n.

Examples

df <- data.frame(
  obs  = c(0.30, 0.25, 0.20, 0.15, 0.10, 0.35, 0.28, 0.18),
  pred = c(0.28, 0.26, 0.22, 0.14, 0.11, 0.33, 0.27, 0.19),
  texture = c("Clay","Clay","Clay","Clay","Clay","Sand","Sand","Sand")
)
swrc_metrics_by_group(df, "obs", "pred", "texture")

Description

Check whether a model directory contains a valid saved model

Usage

swrc_model_exists(dir = "./models/swrc", name = "swrc_model")

Arguments

Value

Logical scalar.

Description

Classify soil samples into USDA texture classes from sand, silt, and clay percentages, and create texture triangle plots.

Description

Creates a USDA texture triangle with filled texture-class regions, USDA class boundary lines, class name labels, and optional data points — all using a pure ggplot2 ternary-to-Cartesian projection (Clay at top, Sand at bottom-left, Silt at bottom-right). No external ternary-plot package is required.

Usage

texture_triangle(
  df,
  sand_col = "sand_total",
  silt_col = "silt",
  clay_col = "clay",
  color_col = NULL,
  title = "Soil Texture Triangle",
  point_size = 3,
  alpha = 0.85
)

Arguments

Details

The background is always coloured by USDA texture class (12-class USDA-inspired palette). Data points are overlaid and can be mapped to a separate variable via color_col.

Value

A ggplot object.

Examples

df <- data.frame(sand_total = c(70, 20, 10, 40),
                 silt       = c(15, 50, 30, 40),
                 clay       = c(15, 30, 60, 20),
                 Texture    = c("Sandy Loam", "Silt Loam", "Clay", "Loam"))
texture_triangle(df, color_col = "Texture")

Description

Returns 100 if the maximum value suggests percentage units (> 1.5), otherwise returns 1 (m3/m3 assumed).

Usage

theta_unit_factor(theta_vec)

Arguments

Value

Numeric scalar: 100 (percentage) or 1 (m3/m3).

Examples

theta_unit_factor(c(0.1, 0.35, 0.5))  # returns 1
theta_unit_factor(c(10, 35, 50))       # returns 100

Description

High-level and low-level functions for training the physics-informed CNN1D model, including the eager-mode train step and the main fitting loop with early stopping.

Description

Internal and exported helpers for pF conversion, depth parsing, and unit detection.