Fractional Cover Modeling with Planet Tanager and Multivariate Random Forests

Introduction

The purpose of this document is to demonstrate the application of multivariate random forests (MRFs) for predicting fractional cover using Planet Tanager hyperspectral imagery. Changes in fractional cover correspond to vegetation phenology, as well as drought, wildfire, disease, and other ecosystem disturbances. Fractional cover is also useful for mapping crop tillage practices. Our previous work has found that MRFs have robust capabilities for accurately mapping fractional cover, and we’re proposing to use Tanager scenes in combination with field-measured fractional cover to validate MRF fractional cover mapping.

MRFs are random forests that are trained to predict multiple response variables simultaneously (Ishwaran et al. 2008). Despite their limited use to date in the field of remote sensing, MRFs come with all of the same benefits of their much more popular univariate counterparts (e.g., handling of large, correlated predictor sets, ability to leverage non-linear predictor-response relationships, etc.) (Breiman 2001). In the context of fractional cover mapping, they have the added benefit of an innate sum-to-one constraint for predictions. In building trees that minimize loss across multiple response variables, MRFs leverage the complementary nature of predictor-response relationships (i.e., if one fraction increases, necessarily the other(s) must decrease). In other words, the covariance structure of the response variables helps the forest learn meaningful patterns among the predictor variables, enhancing predictive ability.

Synthetic Mixtures

We will train and test our MRF model using 10,000 synthetic spectral mixtures that are feature random mixtures of green vegetation (GV), non-photosynthetic vegetation (NPV), and soil endmembers, derived from field and lab spectra (Dennison et al. 2023). Each mixture has a known fraction of GV, NPV, and soil, along with spectral reflectance at 159 bands (10nm band spacing) between 440 and 2400nm (440 - 1320nm; 1490 - 1770nm; 2000 - 2400nm). The spectra were darkened to approximate a real-world sample of reflectance, and noise was added to simulate satellite sensor signal-to-noise ratios.

Here is a snapshot of these data:

library(data.table)
library(knitr)
library(Ternary)
library(randomForestSRC)
library(terra)
library(rhdf5)

# suppress terra progress bars
terraOptions(progress = 0)

# define directory structure
main_dir <- "S:/ursa/campbell/Other/planet_tanager"

# read in spectral mixtures data
df <- file.path(main_dir, "synthetic_mixtures.csv") |>
  fread() |> 
  as.data.frame()

# print sample of the data
kable(head(df), digits = 3)

Table 1: First few rows of the synthetic mixture reflectance data used to train and test a MRF fractional cover prediction model.

gv	npv	soil	nm_440	nm_450	nm_460	nm_470	nm_480	nm_490	nm_500	nm_510	nm_520	nm_530	nm_540	nm_550	nm_560	nm_570	nm_580	nm_590	nm_600	nm_610	nm_620	nm_630	nm_640	nm_650	nm_660	nm_670	nm_680	nm_690	nm_700	nm_710	nm_720	nm_730	nm_740	nm_750	nm_760	nm_770	nm_780	nm_790	nm_800	nm_810	nm_820	nm_830	nm_840	nm_850	nm_860	nm_870	nm_880	nm_890	nm_900	nm_910	nm_920	nm_930	nm_940	nm_950	nm_960	nm_970	nm_980	nm_990	nm_1000	nm_1010	nm_1020	nm_1030	nm_1040	nm_1050	nm_1060	nm_1070	nm_1080	nm_1090	nm_1100	nm_1110	nm_1120	nm_1130	nm_1140	nm_1150	nm_1160	nm_1170	nm_1180	nm_1190	nm_1200	nm_1210	nm_1220	nm_1230	nm_1240	nm_1250	nm_1260	nm_1270	nm_1280	nm_1290	nm_1300	nm_1310	nm_1320	nm_1490	nm_1500	nm_1510	nm_1520	nm_1530	nm_1540	nm_1550	nm_1560	nm_1570	nm_1580	nm_1590	nm_1600	nm_1610	nm_1620	nm_1630	nm_1640	nm_1650	nm_1660	nm_1670	nm_1680	nm_1690	nm_1700	nm_1710	nm_1720	nm_1730	nm_1740	nm_1750	nm_1760	nm_1770	nm_2000	nm_2010	nm_2020	nm_2030	nm_2040	nm_2050	nm_2060	nm_2070	nm_2080	nm_2090	nm_2100	nm_2110	nm_2120	nm_2130	nm_2140	nm_2150	nm_2160	nm_2170	nm_2180	nm_2190	nm_2200	nm_2210	nm_2220	nm_2230	nm_2240	nm_2250	nm_2260	nm_2270	nm_2280	nm_2290	nm_2300	nm_2310	nm_2320	nm_2330	nm_2340	nm_2350	nm_2360	nm_2370	nm_2380	nm_2390	nm_2400
0.366	0.533	0.102	0.035	0.037	0.038	0.040	0.041	0.043	0.045	0.048	0.053	0.058	0.062	0.065	0.067	0.067	0.067	0.068	0.070	0.071	0.073	0.075	0.077	0.078	0.081	0.082	0.085	0.089	0.102	0.122	0.150	0.184	0.212	0.235	0.248	0.254	0.259	0.263	0.265	0.268	0.270	0.272	0.272	0.274	0.276	0.277	0.280	0.280	0.281	0.281	0.281	0.281	0.277	0.271	0.265	0.264	0.266	0.268	0.272	0.277	0.281	0.284	0.291	0.295	0.296	0.301	0.300	0.300	0.304	0.301	0.298	0.292	0.282	0.273	0.263	0.262	0.264	0.265	0.264	0.269	0.270	0.276	0.277	0.279	0.281	0.284	0.283	0.280	0.281	0.279	0.271	0.150	0.154	0.160	0.165	0.170	0.174	0.179	0.184	0.188	0.191	0.196	0.199	0.204	0.208	0.208	0.212	0.215	0.214	0.215	0.215	0.213	0.213	0.211	0.209	0.207	0.205	0.201	0.199	0.198	0.114	0.117	0.119	0.120	0.119	0.119	0.118	0.118	0.117	0.119	0.120	0.121	0.123	0.124	0.125	0.127	0.129	0.130	0.132	0.133	0.134	0.135	0.135	0.134	0.130	0.127	0.123	0.120	0.119	0.119	0.115	0.112	0.111	0.111	0.108	0.106	0.107	0.105	0.103	0.102	0.101
0.065	0.327	0.608	0.077	0.079	0.080	0.082	0.083	0.086	0.090	0.093	0.098	0.103	0.108	0.114	0.120	0.126	0.132	0.138	0.143	0.147	0.151	0.154	0.159	0.161	0.165	0.169	0.173	0.177	0.180	0.186	0.191	0.198	0.205	0.212	0.217	0.221	0.225	0.227	0.229	0.233	0.234	0.235	0.238	0.240	0.241	0.244	0.246	0.246	0.249	0.250	0.250	0.252	0.254	0.255	0.257	0.257	0.259	0.261	0.263	0.267	0.267	0.269	0.272	0.274	0.275	0.278	0.279	0.280	0.282	0.284	0.286	0.287	0.286	0.286	0.286	0.288	0.289	0.290	0.293	0.294	0.295	0.300	0.299	0.299	0.304	0.303	0.305	0.306	0.307	0.307	0.308	0.288	0.292	0.294	0.296	0.299	0.301	0.303	0.304	0.303	0.304	0.307	0.308	0.310	0.313	0.316	0.316	0.313	0.317	0.315	0.318	0.318	0.315	0.318	0.315	0.315	0.311	0.313	0.315	0.314	0.281	0.281	0.283	0.282	0.278	0.274	0.275	0.273	0.273	0.272	0.270	0.271	0.274	0.273	0.273	0.274	0.270	0.268	0.264	0.260	0.257	0.257	0.262	0.266	0.264	0.261	0.259	0.257	0.257	0.253	0.252	0.250	0.246	0.244	0.241	0.240	0.237	0.236	0.236	0.235	0.233
0.311	0.110	0.579	0.058	0.058	0.059	0.060	0.061	0.063	0.067	0.071	0.077	0.084	0.090	0.096	0.102	0.106	0.110	0.114	0.119	0.122	0.125	0.128	0.131	0.135	0.138	0.141	0.144	0.149	0.156	0.167	0.181	0.200	0.220	0.237	0.245	0.250	0.254	0.257	0.258	0.261	0.263	0.265	0.266	0.267	0.269	0.271	0.272	0.272	0.273	0.273	0.275	0.275	0.273	0.271	0.271	0.273	0.274	0.275	0.277	0.281	0.286	0.287	0.289	0.293	0.294	0.297	0.297	0.300	0.300	0.299	0.300	0.300	0.293	0.290	0.288	0.288	0.289	0.287	0.289	0.289	0.292	0.294	0.297	0.298	0.298	0.299	0.301	0.299	0.299	0.298	0.294	0.238	0.239	0.242	0.244	0.246	0.251	0.255	0.253	0.256	0.259	0.260	0.262	0.265	0.268	0.269	0.271	0.272	0.273	0.273	0.272	0.271	0.272	0.270	0.270	0.269	0.267	0.266	0.266	0.264	0.221	0.225	0.227	0.227	0.228	0.230	0.230	0.230	0.231	0.231	0.231	0.232	0.234	0.234	0.232	0.233	0.232	0.231	0.228	0.228	0.225	0.226	0.229	0.232	0.232	0.232	0.226	0.228	0.226	0.227	0.225	0.222	0.218	0.217	0.215	0.214	0.214	0.210	0.208	0.208	0.205
0.104	0.790	0.106	0.055	0.057	0.059	0.061	0.063	0.066	0.069	0.072	0.074	0.078	0.081	0.083	0.086	0.088	0.091	0.093	0.095	0.098	0.099	0.103	0.105	0.109	0.111	0.114	0.117	0.121	0.126	0.132	0.139	0.147	0.157	0.167	0.176	0.180	0.185	0.188	0.192	0.196	0.200	0.203	0.205	0.207	0.211	0.214	0.217	0.221	0.222	0.225	0.228	0.230	0.230	0.232	0.232	0.235	0.237	0.241	0.242	0.244	0.246	0.248	0.252	0.255	0.258	0.262	0.262	0.265	0.265	0.268	0.267	0.270	0.268	0.269	0.268	0.270	0.273	0.274	0.274	0.276	0.280	0.280	0.284	0.288	0.290	0.291	0.295	0.298	0.298	0.300	0.299	0.254	0.259	0.262	0.264	0.268	0.273	0.275	0.278	0.280	0.283	0.287	0.289	0.294	0.295	0.300	0.301	0.303	0.301	0.299	0.302	0.300	0.299	0.296	0.296	0.295	0.297	0.298	0.297	0.296	0.260	0.260	0.256	0.250	0.244	0.240	0.234	0.229	0.228	0.225	0.224	0.225	0.226	0.228	0.230	0.232	0.234	0.237	0.236	0.240	0.241	0.243	0.243	0.241	0.234	0.226	0.221	0.213	0.210	0.210	0.208	0.201	0.201	0.199	0.197	0.199	0.202	0.202	0.200	0.202	0.203
0.157	0.318	0.525	0.094	0.096	0.099	0.102	0.105	0.109	0.114	0.121	0.129	0.138	0.146	0.154	0.160	0.166	0.171	0.175	0.180	0.184	0.187	0.192	0.198	0.201	0.205	0.209	0.214	0.220	0.228	0.240	0.253	0.267	0.282	0.293	0.298	0.305	0.311	0.312	0.318	0.321	0.323	0.324	0.328	0.328	0.331	0.332	0.336	0.336	0.338	0.341	0.344	0.346	0.347	0.348	0.348	0.350	0.351	0.353	0.355	0.358	0.359	0.366	0.367	0.372	0.371	0.375	0.379	0.381	0.381	0.380	0.384	0.380	0.380	0.378	0.376	0.381	0.385	0.380	0.383	0.390	0.392	0.398	0.399	0.400	0.401	0.403	0.404	0.405	0.407	0.408	0.404	0.362	0.366	0.371	0.372	0.375	0.379	0.380	0.383	0.388	0.388	0.389	0.391	0.395	0.393	0.394	0.398	0.396	0.395	0.397	0.396	0.398	0.396	0.393	0.392	0.389	0.391	0.390	0.393	0.390	0.336	0.341	0.344	0.346	0.342	0.345	0.345	0.345	0.345	0.345	0.345	0.346	0.348	0.348	0.349	0.348	0.346	0.347	0.345	0.340	0.335	0.335	0.335	0.334	0.334	0.331	0.326	0.323	0.319	0.316	0.313	0.308	0.305	0.304	0.299	0.294	0.298	0.300	0.300	0.298	0.296
0.317	0.335	0.348	0.063	0.065	0.065	0.066	0.068	0.070	0.073	0.078	0.084	0.091	0.097	0.102	0.107	0.111	0.113	0.116	0.119	0.122	0.124	0.127	0.129	0.133	0.135	0.137	0.141	0.147	0.158	0.176	0.199	0.226	0.248	0.264	0.274	0.282	0.285	0.291	0.294	0.296	0.300	0.303	0.305	0.308	0.310	0.315	0.315	0.317	0.321	0.322	0.324	0.325	0.327	0.324	0.323	0.323	0.323	0.326	0.328	0.334	0.338	0.345	0.345	0.350	0.353	0.356	0.357	0.357	0.357	0.352	0.350	0.348	0.342	0.333	0.332	0.331	0.330	0.330	0.330	0.336	0.337	0.342	0.344	0.349	0.350	0.350	0.355	0.354	0.353	0.345	0.345	0.224	0.230	0.234	0.239	0.244	0.248	0.252	0.256	0.257	0.259	0.263	0.268	0.273	0.274	0.278	0.280	0.282	0.281	0.281	0.281	0.280	0.278	0.276	0.271	0.270	0.272	0.271	0.267	0.266	0.199	0.199	0.204	0.205	0.202	0.203	0.201	0.201	0.200	0.199	0.201	0.202	0.204	0.206	0.207	0.207	0.207	0.209	0.208	0.208	0.205	0.207	0.210	0.212	0.209	0.204	0.201	0.198	0.195	0.194	0.191	0.188	0.185	0.185	0.183	0.182	0.181	0.181	0.179	0.179	0.177

Below is a ternary plot featuring a sample of 100 mixtures. Three randomly selected mixtures are highlighted, as their spectra will be plotted out next.

# create empty ternary plot
par(mar = rep(0,4))
TernaryPlot(
  alab = "GV",
  blab = "NPV",
  clab = "Soil",
  lab.col = c("green", "red", "blue")
)

# get sample for plotting
set.seed(57)
df_tern <- df[sample(nrow(df), 100),]

# get rgb colors
cols <- rgb(
  red = df_tern$npv,
  green = df_tern$gv,
  blue = df_tern$soil
)

# grab a few random mixtures
rand_mixs <- sample(nrow(df_tern), 3)

# get fraction columns
frcols <- c("gv", "npv", "soil")

# add sample of points
TernaryPoints(
  df_tern[,frcols],
  pch = 16,
  col = adjustcolor(cols, alpha.f = 0.75),
  cex = 1
)

# add random mixture points
TernaryPoints(
  df_tern[rand_mixs, frcols], 
  pch = 21, 
  bg = cols[rand_mixs], 
  cex = 3
)
TernaryText(
  df_tern[rand_mixs, frcols], 
  labels = c(1,2,3),
  col = "white",
  font = 2
)

Figure 1: Ternary plot with a sample of synthetic mixtures from our training data. Points are colored according to their fractional cover in RGB space, where NPV fraction is displayed through the Red channel, GV Green, and Soil blue. The full reflectance spectra of points 1, 2, and 3 are shown in the next figure.

Below is a plot containing the reflectance spectra of the three points shown in the previous ternary plot.

# get reflectance columns
ref_cols <- grep("^nm", colnames(df), value = T)

# get wavelengths
wvls <- ref_cols |>
  sub("nm\\_", "", x = _) |>
  as.numeric()
wvls_all <- seq(min(wvls), max(wvls), 10)

# create reflectance matrix
ref_mat <- matrix(
  NA,
  nrow = nrow(df_tern), 
  ncol = length(wvls_all)
)
colnames(ref_mat) <- wvls_all
for (i in seq_along(wvls_all)){
  wvl <- wvls_all[i]
  if (paste0("nm_", wvl) %in% colnames(df_tern)){
    ref_mat[,i] <- df_tern[,paste0("nm_", wvl)]
  }
}

# create empty plot
par(mar = c(5,5,1,10), las = 1)
plot(
  x = range(wvls_all),
  y = range(ref_mat[rand_mixs,], na.rm = T),
  type = "n",
  xlab = "Wavelength (nm)",
  ylab = "Reflectance"
)

# add random mixtures
lines(
  x = wvls_all,
  y = ref_mat[rand_mixs[1],],
  lwd = 3,
  col = cols[rand_mixs[1]]
)
lines(
  x = wvls_all,
  y = ref_mat[rand_mixs[2],],
  lwd = 3,
  col = cols[rand_mixs[2]]
)
lines(
  x = wvls_all,
  y = ref_mat[rand_mixs[3],],
  lwd = 3,
  col = cols[rand_mixs[3]]
)

# get fractions of sample mixtures
fracs_1 <- paste0(
  "Mixture 1\nGV: ", 
  formatC(df_tern$gv[rand_mixs[1]], digits = 2, format = "f"), "\nNPV: ",
  formatC(df_tern$npv[rand_mixs[1]], digits = 2, format = "f"), "\nSoil: ",
  formatC(df_tern$soil[rand_mixs[1]], digits = 2, format = "f")
)
fracs_2 <- paste0(
  "Mixture 2\n(GV: ", 
  formatC(df_tern$gv[rand_mixs[2]], digits = 2, format = "f"), "\nNPV: ",
  formatC(df_tern$npv[rand_mixs[2]], digits = 2, format = "f"), "\nSoil: ",
  formatC(df_tern$soil[rand_mixs[2]], digits = 2, format = "f")
)
fracs_3 <- paste0(
  "Mixture 3\n(GV: ", 
  formatC(df_tern$gv[rand_mixs[3]], digits = 2, format = "f"), "\nNPV: ",
  formatC(df_tern$npv[rand_mixs[3]], digits = 2, format = "f"), "\nSoil: ",
  formatC(df_tern$soil[rand_mixs[3]], digits = 2, format = "f")
)

# add legend
par(xpd = NA)
legend(
  x = par("usr")[2] + 0.05 * diff(par("usr")[1:2]),
  y = mean(par("usr")[3:4]),
  legend = c(fracs_1, fracs_2, fracs_3),
  lwd = 5,
  col = cols[rand_mixs],
  bty = "n",
  xjust = 0,
  yjust = 0.5,
  y.intersp = 2
)
par(xpd = F)

Figure 2: Spectral reflectance of the three example synthetic mixtures shown in Figure 1. Lines are colored according to their fractional cover in RGB space, where NPV fraction is displayed through the Red channel, GV Green, and Soil blue.

Train Random Forest Model

Next, we will train a MRF model using the randomForestSRC package (Ishwaran et al. 2008). To do this, we will first randomly split our 10,000 mixtures into training (80%) and test (20%) datasets. A model will be trained with the training data and used to predict fractional cover with the test data. The scatterplots below show error for the test data. GV cover is distinct and is modeled with the highest accuracy, while NPV and soil are more difficult to separate. Past work has demonstrated that imaging spectrometer data, like Tanager, improves accuracy of NPV-soil separation due to being able to resolve lignocellulose absorption.

# split into training and test
train_frac <- 0.8
samp <- sample(nrow(df), nrow(df) * train_frac)
df_train <- df[samp,]
df_test <- df[-samp,]

# train multivariate random forest
mrf <- rfsrc(
  Multivar(gv, npv, soil) ~ .,
  data = df_train
)

# apply to test data
pred <- predict(mrf, df_test)

# compile test predictions and observations
df_pred_obs <- data.frame(
  gv_obs = df_test$gv,
  npv_obs = df_test$npv,
  soil_obs = df_test$soil,
  gv_pred = pred$regrOutput$gv$predicted,
  npv_pred = pred$regrOutput$npv$predicted,
  soil_pred = pred$regrOutput$soil$predicted
)

# create plotting function
plot_fun <- function(frac){
  plot(
    x = c(0,1), 
    y = c(0,1), 
    type = "n", 
    xaxt = "n", 
    yaxt = "n", 
    xlab = NA,
    ylab = NA
  )
  x <- df_pred_obs[,paste0(frac, "_obs")]
  y <- df_pred_obs[,paste0(frac, "_pred")]
  if (frac == "gv"){
    col_pt <- adjustcolor("green", 0.75)
    col_ln <- "darkgreen"
  } else if (frac == "npv"){
    col_pt <- adjustcolor("red", 0.75)
    col_ln <- "darkred"
  } else {
    col_pt <- adjustcolor("blue", 0.75)
    col_ln <- "darkblue"
  }
  grid()
  box()
  abline(0,1)
  points(x, y, pch = 16, col = col_pt)
  mod <- lm(y ~ x)
  abline(mod, lwd = 3, col = col_ln)
  r2 <- 1 - sum((x-y)^2)/sum((x-mean(x))^2)
  r2 <- formatC(r2, digits = 2, format = "f")
  r2 <- bquote(R^2==.(r2))
  rmse <- sqrt(mean((x-y)^2))
  rmse <- formatC(rmse, digits = 2, format = "f")
  rmse <- bquote(RMSE==.(rmse))
  legend(
    "topleft",
    legend = c(r2, rmse),
    text.col = col_ln,
    text.font = 2,
    bty = "n",
    x.intersp = 0
  )
}

# plot predictions and observations
par(mfrow = c(1,3), mar = rep(0,4), oma = c(5,5,2,1), las = 1)
plot_fun("gv")
mtext("GV", side = 3, font = 2, col = "darkgreen", line = 0.5)
axis(1)
axis(2)
plot_fun("npv")
mtext("NPV", side = 3, font = 2, col = "darkred", line = 0.5)
axis(1)
plot_fun("soil")
mtext("Soil", side = 3, font = 2, col = "darkblue", line = 0.5)
axis(1)
mtext("Observed Fraction", 1, 3, outer = T, font = 2)
mtext("Predicted Fraction", 2, 3, outer = T, font = 2, las = 0)

Figure 3: MRF-predicted versus observed fractional cover for GV, NPV, and Soil.

Explore Tanager Scene

With a MRF model now trained, we can now move towards image-based prediction. We acquired a Planet Tanager scene that was captured in Central Utah in 2025 during the Monroe Canyon Fire. We downloaded the orthorectified surface reflectance product (ortho_sr_hdf5).

Below, we will load the scene’s imagery and metadata, and map it out using a natural color composite. The smoke plume of this fire – one of Utah’s largest in recent history – can be easily seen in the image.

# get hdf file
h5_file <- file.path(main_dir, "20250724_190927_83_4001_ortho_sr_hdf5.h5")

# open as SpatRasterDataset
h5_sds <- sds(h5_file)

# find surface reflectance bands
sr_id <- grep(
  "surface_reflectance$",
  names(h5_sds),
  ignore.case = TRUE
)

# create SpatRaster of surface reflectance
sr <- h5_sds[sr_id]

# read hdf attributes
sr_h5_path <- "/HDFEOS/GRIDS/HYP/Data Fields/surface_reflectance"
attrs <- h5readAttributes(h5_file, sr_h5_path)

# get useful attributes
wavelength_nm <- as.numeric(attrs$wavelengths)
fwhm_nm <- as.numeric(attrs$fwhm)
good_band <- as.logical(attrs$good_wavelengths)

# combine
band_metadata <- data.frame(
  layer = seq_len(nlyr(sr)),
  band = sprintf("B%03d", 1:nlyr(sr)),
  wavelength_nm = wavelength_nm,
  fwhm_nm = fwhm_nm,
  good_wavelength = good_band
)

# add band names to imagery
names(sr) <- band_metadata$band

# get rgb image
r_band <- which.min((abs(650 - band_metadata$wavelength_nm)))
g_band <- which.min((abs(550 - band_metadata$wavelength_nm)))
b_band <- which.min((abs(450 - band_metadata$wavelength_nm)))
sr_rgb <- sr[[c(r_band, g_band, b_band)]]
sr_rgb <- ifel(sr_rgb < 0, 0, sr_rgb)
sr_rgb <- ifel(sr_rgb > 0.25, 0.25, sr_rgb)

# plot it
plotRGB(sr_rgb, scale = 0.25)

Figure 4: Natural color composite of Planet Tanager scene captured in Central Utah during the Monroe Canyon Fire.

Interpolation

Given that the wavelengths of Tanager data differ from our 10nm-resolution synthetic mixture data, we will first interpolate each image pixel’s spectrum to match the band centers of the synthetic mixture data.

# create spline interpolated image
sr_interp <- app(
  sr,
  fun = function(img) {
    valid <- (
      is.finite(img) &
      img != -9999 &
      band_metadata$good_wavelength &
      is.finite(band_metadata$wavelength_nm)
    )
    if (sum(valid) < 2) {
      return(rep(NA_real_, length(wvls)))
    }
    approx(
      x = band_metadata$wavelength_nm[valid],
      y = img[valid],
      xout = wvls,
      rule = 1,
      ties = "ordered"
    )$y
  }
)

# set names to match synthetic mixtures
names(sr_interp) <- paste0("nm_", wvls)

Predict Fractional Cover

We will now use the MRF model to predict fractional cover with the interpolated Tanager scene. Below you will see the resulting three-band predicted map, where predicted NPV fraction is being displayed through the red channel, GV through green, and Soil through blue. Some very interesting patterns emerge, including:

• A very useful capture of the Monroe Canyon Fire burn scar, predicted as high soil fraction
• The ecological gradient from soil/NPV-dominated rangeland at lower elevations to the GV-dominated forests at higher elevations
• Older fire scars showing up as NPV-dominated late-summer senesced grass/shrub cover
• Riparian and agricultural areas featuring very high GV fraction
• Issues with clouds being predicted as high soil fraction (see next section) and cloud shadows being predicted as high GV fraction

# define prediction function
pred_fun <- function(model, data, ...) {
  pred <- stats::predict(
    model,
    newdata = data,
    ...
  )
  data.frame(
    npv  = pred$regrOutput$npv$predicted,
    gv   = pred$regrOutput$gv$predicted,
    soil = pred$regrOutput$soil$predicted
  )
}

# predict
pred_img <- terra::predict(
  sr_interp,
  mrf,
  fun = pred_fun,
  na.rm = TRUE
)

# plot predictions
plotRGB(pred_img, scale = 1)

Figure 5: MRF-predicted fractional cover displayed as a three-channel (RGB) image, where NPV fraction is displayed through the red channel, GV through green, and Soil through blue.

Masking

We will use the ortho_sr_hdf5 beta_cloud_mask layer to mask out the clouds (and some larger sections of smoke plume).

# find beta cloud mask
cloud_id <- grep(
  "beta_cloud_mask$",
  names(h5_sds),
  ignore.case = TRUE
)

# extract as a SpatRaster
cloud_mask <- h5_sds[cloud_id]
names(cloud_mask) <- "beta_cloud_mask"

# set cloudy areas to NA
pred_img <- ifel(cloud_mask == 1, NA, pred_img)

# plot predictions
plotRGB(pred_img, scale = 1)

Figure 6: MRF-predicted fractional cover displayed as a three-channel (RGB) image, as in Figure 5, but with clouds masked.

Conclusions

MRF provides a potential machine learning method for global mapping of fractional cover. We have shown that this method can be applied to Planet Tanager imagery in a manner that should be highly transferable across ecosystems while innately enforcing a sum-to-one constraint. Tanager data and field measurement of fractional cover can be used to validate MRF-estimated fractions, demonstrate scalability, and prove applicability across key ecosystems.

References

Breiman, Leo. 2001. “Random Forests.” Machine Learning 45: 5–32. https://doi.org/10.1023/A:1010933404324.

Dennison, Philip E, Brian T Lamb, Michael J Campbell, et al. 2023. “Modeling Global Indices for Estimating Non-Photosynthetic Vegetation Cover.” Remote Sensing of Environment 295: 113715. https://doi.org/10.1016/j.rse.2023.113715.

Ishwaran, Hemant, Udaya B. Kogalur, Eugene H. Blackstone, and Michael S. Lauer. 2008. “Random Survival Forests.” The Annals of Applied Statistics 2 (3): 841–60. https://doi.org/10.1214/08-AOAS169.