Mercurial > repos > ecology > srs_diversity_maps
view prosail-master/R/Lib_PROSAIL_HybridInversion.R @ 0:9adccd3da70c draft default tip
planemo upload for repository https://github.com/Marie59/Sentinel_2A/srs_tools commit b32737c1642aa02cc672534e42c5cb4abe0cd3e7
author | ecology |
---|---|
date | Mon, 09 Jan 2023 13:37:37 +0000 |
parents | |
children |
line wrap: on
line source
# ============================================================================= = # prosail # Lib_PROSAIL_HybridInversion.R # ============================================================================= = # PROGRAMMERS: # Jean-Baptiste FERET <jb.feret@teledetection.fr> # Florian de BOISSIEU <fdeboiss@gmail.com> # Copyright 2019 / 11 Jean-Baptiste FERET # ============================================================================= = # This Library includes functions dedicated to PROSAIL inversion using hybrid # approach based on SVM regression # ============================================================================= = #" This function applies SVR model on raster data in order to estimate #" vegetation biophysical properties #" #" @param raster_path character. path for a raster file #" @param hybridmodel list. hybrid models produced from train_prosail_inversion #" each element of the list corresponds to a set of hybrid models for a vegetation parameter #" @param pathout character. path for directory where results are written #" @param selectedbands list. list of spectral bands to be selected from raster (identified by name of vegetation parameter) #" @param bandname character. spectral bands corresponding to the raster #" @param maskraster character. path for binary mask defining ON (1) and OFF (0) pixels in the raster #" @param multiplyingfactor numeric. multiplying factor used to write reflectance in the raster #" --> PROSAIL simulates reflectance between 0 and 1, and raster data expected in the same range #" #" @return None #" @importFrom progress progress_bar #" @importFrom stars read_stars #" @importFrom raster raster brick blockSize readStart readStop getValues writeStart writeStop writeValues #" @import rgdal #" @export apply_prosail_inversion <- function(raster_path, hybridmodel, pathout, selectedbands, bandname, maskraster = FALSE, multiplyingfactor = 10000) { # explain which biophysical variables will be computed bpvar <- names(hybridmodel) print("The following biophysical variables will be computed") print(bpvar) # get image dimensions if (attr(rgdal::GDALinfo(raster_path, returnStats = FALSE), "driver") == "ENVI") { hdr <- read_envi_header(get_hdr_name(raster_path)) dimsraster <- list("rows" = hdr$lines, "cols" = hdr$samples, "bands" = hdr$bands) } else { dimsraster <- dim(read_stars(raster_path)) dimsraster <- list("rows" = as.numeric(dimsraster[2]), "cols" = as.numeric(dimsraster[1]), "bands" = as.numeric(dimsraster[3])) } # Produce a map for each biophysical property for (parm in bpvar) { print(paste("Computing", parm, sep = " ")) # read by chunk to avoid memory problem blk <- blockSize(brick(raster_path)) # reflectance file r_in <- readStart(brick(raster_path)) # mask file r_inmask <- FALSE if (maskraster == FALSE) { selectpixels <- "ALL" } else if (!maskraster == FALSE) { if (file.exists(maskraster)) { r_inmask <- readStart(raster(maskraster)) } else if (!file.exists(maskraster)) { message("WARNING: Mask file does not exist:") print(maskraster) message("Processing all image") selectpixels <- "ALL" } } # initiate progress bar pgbarlength <- length(hybridmodel[[parm]]) * blk$n pb <- progress_bar$new( format = "Hybrid inversion on raster [:bar] :percent in :elapsedfull, estimated time remaining :eta", total = pgbarlength, clear = FALSE, width = 100) # output files bpvarpath <- file.path(pathout, paste(basename(raster_path), parm, sep = "_")) bpvarsdpath <- file.path(pathout, paste(basename(raster_path), parm, "STD", sep = "_")) r_outmean <- writeStart(raster(raster_path), filename = bpvarpath, format = "ENVI", overwrite = TRUE) r_outsd <- writeStart(raster(raster_path), filename = bpvarsdpath, format = "ENVI", overwrite = TRUE) selbands <- match(selectedbands[[parm]], bandname) # loop over blocks for (i in seq_along(blk$row)) { # read values for block # format is a matrix with rows the cells values and columns the layers blockval <- getValues(r_in, row = blk$row[i], nrows = blk$nrows[i]) fulllength <- dim(blockval)[1] if (typeof(r_inmask) == "logical") { blockval <- blockval[, selbands] # automatically filter pixels corresponding to negative values selectpixels <- which(blockval[, 1] > 0) blockval <- blockval[selectpixels, ] } else if (typeof(r_inmask) == "S4") { maskval <- getValues(r_inmask, row = blk$row[i], nrows = blk$nrows[i]) selectpixels <- which(maskval == 1) blockval <- blockval[selectpixels, selbands] } mean_estimatefull <- NA * vector(length = fulllength) std_estimatefull <- NA * vector(length = fulllength) if (length(selectpixels) > 0) { blockval <- blockval / multiplyingfactor modelsvr_estimate <- list() for (modind in 1:seq_along(hybridmodel[[parm]])) { pb$tick() modelsvr_estimate[[modind]] <- predict(hybridmodel[[parm]][[modind]], blockval) } modelsvr_estimate <- do.call(cbind, modelsvr_estimate) # final estimated value = mean parm value for all models mean_estimate <- rowMeans(modelsvr_estimate) # "uncertainty" = STD value for all models std_estimate <- rowSds(modelsvr_estimate) mean_estimatefull[selectpixels] <- mean_estimate std_estimatefull[selectpixels] <- std_estimate } else { for (modind in 1:seq_along(hybridmodel[[parm]])) { pb$tick() } } r_outmean <- writeValues(r_outmean, mean_estimatefull, blk$row[i], format = "ENVI", overwrite = TRUE) r_outsd <- writeValues(r_outsd, std_estimatefull, blk$row[i], format = "ENVI", overwrite = TRUE) } # close files r_in <- readStop(r_in) if (typeof(r_inmask) == "S4") { r_inmask <- readStop(r_inmask) } r_outmean <- writeStop(r_outmean) r_outsd <- writeStop(r_outsd) # write biophysical variable name in headers hdr <- read_envi_header(get_hdr_name(bpvarpath)) hdr$`band names` <- paste("{", parm, "}", sep = "") write_envi_header(hdr, get_hdr_name(bpvarpath)) } print("processing completed") return(invisible()) } #" get hdr name from image file name, assuming it is BIL format #" #" @param impath path of the image #" #" @return corresponding hdr #" @importFrom tools file_ext file_path_sans_ext #" @export get_hdr_name <- function(impath) { if (tools::file_ext(impath) == "") { impathhdr <- paste(impath, ".hdr", sep = "") } else if (tools::file_ext(impath) == "bil") { impathhdr <- gsub(".bil", ".hdr", impath) } else if (tools::file_ext(impath) == "zip") { impathhdr <- gsub(".zip", ".hdr", impath) } else { impathhdr <- paste(tools::file_path_sans_ext(impath), ".hdr", sep = "") } if (!file.exists(impathhdr)) { message("WARNING : COULD NOT FIND hdr FILE") print(impathhdr) message("Process may stop") } return(impathhdr) } #" This function applies the regression models trained with prosail_hybrid_train #" #" @param regressionmodels list. List of regression models produced by prosail_hybrid_train #" @param refl numeric. LUT of bidirectional reflectances factors used for training #" #" @return hybridres list. Estimated values corresponding to refl. Includes #" - meanestimate = mean value for the ensemble regression model #" - stdestimate = std value for the ensemble regression model #" @importFrom stats predict #" @importFrom matrixStats rowSds #" @importFrom progress progress_bar #" @export prosail_hybrid_apply <- function(regressionmodels, refl) { # make sure refl is right dimensions refl <- t(refl) nbfeatures <- regressionmodels[[1]]$dim if (!ncol(refl) == nbfeatures && nrow(refl) == nbfeatures) { refl <- t(refl) } nbensemble <- length(regressionmodels) estimatedval <- list() pb <- progress_bar$new( format = "Applying SVR models [:bar] :percent in :elapsed", total = nbensemble, clear = FALSE, width = 100) for (i in 1:nbensemble) { pb$tick() estimatedval[[i]] <- predict(regressionmodels[[i]], refl) } estimatedval <- do.call(cbind, estimatedval) meanestimate <- rowMeans(estimatedval) stdestimate <- rowSds(estimatedval) hybridres <- list("meanestimate" = meanestimate, "stdestimate" = stdestimate) return(hybridres) } #" This function trains a suppot vector regression for a set of variables based on spectral data #" #" @param brf_lut numeric. LUT of bidirectional reflectances factors used for training #" @param inputvar numeric. biophysical parameter corresponding to the reflectance #" @param figplot Boolean. Set to TRUE if you want a scatterplot #" @param nbensemble numeric. Number of individual subsets should be generated from brf_lut #" @param withreplacement Boolean. should subsets be generated with or without replacement? #" #" @return modelssvr list. regression models trained for the retrieval of inputvar based on brf_lut #" @importFrom liquidSVM svmRegression #" @importFrom stats predict #" @importFrom progress progress_bar #" @importFrom graphics par #" @importFrom expandFunctions reset.warnings #" @importFrom stringr str_split #" @importFrom simsalapar tryCatch.W.E #" @import dplyr #" @import ggplot2 # @" @import caret #" @export prosail_hybrid_train <- function(brf_lut, inputvar, figplot = FALSE, nbensemble = 20, withreplacement = FALSE) { x <- y <- ymean <- ystdmin <- ystdmax <- NULL # split the LUT into nbensemble subsets nbsamples <- length(inputvar) if (dim(brf_lut)[2] == nbsamples) { brf_lut <- t(brf_lut) } # if subsets are generated from brf_lut with replacement if (withreplacement == TRUE) { subsets <- list() samples_per_run <- round(nbsamples / nbensemble) for (run in (1:nbensemble)) { subsets[[run]] <- sample(seq(1, nbsamples), samples_per_run, replace = TRUE) } # if subsets are generated from brf_lut without replacement } else if (withreplacement == FALSE) { subsets <- split(sample(seq(1, nbsamples, by = 1)), seq(1, nbensemble, by = 1)) } # run training for each subset modelssvr <- list() predictedyall <- list() tunedmodelyall <- list() pb <- progress_bar$new( format = "Training SVR on subsets [:bar] :percent in :elapsed", total = nbensemble, clear = FALSE, width = 100) for (i in 1:nbensemble) { pb$tick() Sys.sleep(1 / 100) trainingset <- list() trainingset$X <- brf_lut[subsets[i][[1]], ] trainingset$Y <- inputvar[subsets[i][[1]]] # liquidSVM r1 <- tryCatch.W.E(tunedmodel <- liquidSVM::svmRegression(trainingset$X, trainingset$Y)) if (!is.null(r1$warning)) { msg <- r1$warning$message valgamma <- str_split(string = msg, pattern = "gamma=")[[1]][2] vallambda <- str_split(string = msg, pattern = "lambda=")[[1]][2] if (!is.na(as.numeric(valgamma))) { message("Adjusting Gamma accordingly") valgamma <- as.numeric(valgamma) tunedmodel <- liquidSVM::svmRegression(trainingset$X, trainingset$Y, min_gamma = valgamma) } if (!is.na(as.numeric(vallambda))) { message("Adjusting Lambda accordingly") vallambda <- as.numeric(vallambda) tunedmodel <- liquidSVM::svmRegression(trainingset$X, trainingset$Y, min_lambda = vallambda) } } modelssvr[[i]] <- tunedmodel } # if scatterplots needed if (figplot == TRUE) { # predict for full brf_lut for (i in 1:nbensemble) { tunedmodely <- stats::predict(modelssvr[[i]], brf_lut) tunedmodelyall <- cbind(tunedmodelyall, matrix(tunedmodely, ncol = 1)) } # plot prediction df <- data.frame(x = rep(1:nbsamples, nbensemble), y = as.numeric(matrix(tunedmodelyall, ncol = 1))) df_summary <- df %>% dplyr::group_by(x) %>% summarize(ymin = min(y), ystdmin = mean(y) - sd(y), ymax = max(y), ystdmax = mean(y) + sd(y), ymean = mean(y)) par(mar = rep(.1, 4)) p <- ggplot(df_summary, aes(x = inputvar, y = ymean)) + geom_point(size = 2) + geom_errorbar(aes(ymin = ystdmin, ymax = ystdmax)) meanpredict <- rowMeans(matrix(as.numeric(tunedmodelyall), ncol = nbensemble)) print(p) } return(modelssvr) } #" Reads ENVI hdr file #" #" @param hdrpath Path of the hdr file #" #" @return list of the content of the hdr file #" @export read_envi_header <- function(hdrpath) { if (!grepl(".hdr$", hdrpath)) { stop("File extension should be .hdr") } hdr <- readLines(hdrpath) ## check ENVI at beginning of file if (!grepl("ENVI", hdr[1])) { stop("Not an ENVI header (ENVI keyword missing)") } else { hdr <- hdr [-1] } ## remove curly braces and put multi-line key-value-pairs into one line hdr <- gsub("\\{([^}]*)\\}", "\\1", hdr) l <- grep("\\{", hdr) r <- grep("\\}", hdr) if (length(l) != length(r)) { stop("Error matching curly braces in header (differing numbers).") } if (any(r <= l)) { stop("Mismatch of curly braces in header.") } hdr[l] <- sub("\\{", "", hdr[l]) hdr[r] <- sub("\\}", "", hdr[r]) for (i in rev(seq_along(l))) { hdr <- c( hdr [seq_len(l [i] - 1)], paste(hdr [l [i]:r [i]], collapse = "\n"), hdr [-seq_len(r [i])] ) } ## split key = value constructs into list with keys as names hdr <- sapply(hdr, split_line, " = ", USE.NAMES = FALSE) names(hdr) <- tolower(names(hdr)) ## process numeric values tmp <- names(hdr) %in% c( "samples", "lines", "bands", "header offset", "data type", "byte order", "default bands", "data ignore value", "wavelength", "fwhm", "data gain values" ) hdr [tmp] <- lapply(hdr [tmp], function(x) { as.numeric(unlist(strsplit(x, ","))) }) return(hdr) } #" ENVI functions #" #" based on https: / / github.com / cran / hyperSpec / blob / master / R / read.ENVI.R #" added wavelength, fwhm, ... to header reading #" Title #" #" @param x character. #" @param separator character #" @param trim_blank boolean. #" #" @return list. #" @export split_line <- function(x, separator, trim_blank = TRUE) { tmp <- regexpr(separator, x) key <- substr(x, 1, tmp - 1) value <- substr(x, tmp + 1, nchar(x)) if (trim_blank) { blank_pattern <- "^[[:blank:]]*([^[:blank:]] + .*[^[:blank:]] + )[[:blank:]]*$" key <- sub(blank_pattern, "\\1", key) value <- sub(blank_pattern, "\\1", value) } value <- as.list(value) names(value) <- key return(value) } #" This function performs full training for hybrid invrsion using SVR with #" values for default parameters #" #" @param minval list. minimum value for input parameters sampled to produce a training LUT #" @param maxval list. maximum value for input parameters sampled to produce a training LUT #" @param typedistrib list. Type of distribution. Either "Uniform" or "Gaussian" #" @param gaussiandistrib list. Mean value and STD corresponding to the parameters sampled with gaussian distribution #" @param parmset list. list of input parameters set to a specific value #" @param nbsamples numeric. number of samples in training LUT #" @param nbsamplesperrun numeric. number of training sample per individual regression model #" @param nbmodels numeric. number of individual models to be run for ensemble #" @param replacement bolean. is there replacement in subsampling? #" @param sailversion character. Either 4SAIL or 4SAIL2 #" @param parms2estimate list. list of input parameters to be estimated #" @param bands2select list. list of bands used for regression for each input parameter #" @param noiselevel list. list of noise value added to reflectance (defined per input parm) #" @param specprospect list. Includes optical constants required for PROSPECT #" @param specsoil list. Includes either dry soil and wet soil, or a unique soil sample if the psoil parameter is not inverted #" @param specatm list. Includes direct and diffuse radiation for clear conditions #" @param path_results character. path for results #" @param figplot boolean. Set TRUE to get scatterplot of estimated biophysical variable during training step #" @param force4lowlai boolean. Set TRUE to artificially reduce leaf chemical constituent content for low LAI #" #" #" @return modelssvr list. regression models trained for the retrieval of inputvar based on brf_lut #" @export train_prosail_inversion <- function(minval = NULL, maxval = NULL, typedistrib = NULL, gaussiandistrib = NULL, parmset = NULL, nbsamples = 2000, nbsamplesperrun = 100, nbmodels = 20, replacement = TRUE, sailversion = "4SAIL", parms2estimate = "lai", bands2select = NULL, noiselevel = NULL, specprospect = NULL, specsoil = NULL, specatm = NULL, path_results = "./", figplot = FALSE, force4lowlai = TRUE) { ###===================================================================### ### 1- PRODUCE A LUT TO TRAIN THE HYBRID INVERSION ### ###===================================================================### # Define sensor characteristics if (is.null(specprospect)) { specprospect <- prosail::specprospect } if (is.null(specsoil)) { specsoil <- prosail::specsoil } if (is.null(specprospect)) { specatm <- prosail::specatm } # define distribution for parameters to be sampled if (is.null(typedistrib)) { typedistrib <- data.frame("CHL" = "Uniform", "CAR" = "Uniform", "EWT" = "Uniform", "ANT" = "Uniform", "LMA" = "Uniform", "N" = "Uniform", "BROWN" = "Uniform", "psoil" = "Uniform", "LIDFa" = "Uniform", "lai" = "Uniform", "q" = "Uniform", "tto" = "Uniform", "tts" = "Uniform", "psi" = "Uniform") } if (is.null(gaussiandistrib)) { gaussiandistrib <- list("Mean" = NULL, "Std" = NULL) } if (is.null(minval)) { minval <- data.frame("CHL" = 10, "CAR" = 0, "EWT" = 0.01, "ANT" = 0, "LMA" = 0.005, "N" = 1.0, "psoil" = 0.0, "BROWN" = 0.0, "LIDFa" = 20, "lai" = 0.5, "q" = 0.1, "tto" = 0, "tts" = 20, "psi" = 80) } if (is.null(maxval)) { maxval <- data.frame("CHL" = 75, "CAR" = 15, "EWT" = 0.03, "ANT" = 2, "LMA" = 0.03, "N" = 2.0, "psoil" = 1.0, "BROWN" = 0.5, "LIDFa" = 70, "lai" = 7, "q" = 0.2, "tto" = 5, "tts" = 30, "psi" = 110) } # define min and max values # fixed parameters if (is.null(parmset)) { parmset <- data.frame("TypeLidf" = 2, "alpha" = 40) } # produce input parameters distribution if (sailversion == "4SAIL") { inputprosail <- get_distribution_input_prosail(minval, maxval, parmset, nbsamples, typedistrib = typedistrib, Mean = gaussiandistrib$Mean, Std = gaussiandistrib$Std, force4lowlai = force4lowlai) } else if (sailversion == "4SAIL2") { inputprosail <- get_distribution_input_prosail2(minval, maxval, parmset, nbsamples, typedistrib = typedistrib, Mean = gaussiandistrib$Mean, Std = gaussiandistrib$Std, force4lowlai = force4lowlai) } if (sailversion == "4SAIL2") { # Definition of Cv && update LAI maxlai <- min(c(maxval$lai), 4) inputprosail$Cv <- NA * inputprosail$lai inputprosail$Cv[which(inputprosail$lai > maxlai)] <- 1 inputprosail$Cv[which(inputprosail$lai <= maxlai)] <- (1 / maxlai) + inputprosail$lai[which(inputprosail$lai <= maxlai)] / (maxlai + 1) inputprosail$Cv <- inputprosail$Cv * matrix(rnorm(length(inputprosail$Cv), mean = 1, sd = 0.1)) inputprosail$Cv[which(inputprosail$Cv < 0)] <- 0 inputprosail$Cv[which(inputprosail$Cv > 1)] <- 1 inputprosail$Cv[which(inputprosail$lai > maxlai)] <- 1 inputprosail$fraction_brown <- 0 + 0 * inputprosail$lai inputprosail$diss <- 0 + 0 * inputprosail$lai inputprosail$Zeta <- 0.2 + 0 * inputprosail$lai inputprosail$lai <- inputprosail$lai * inputprosail$Cv } # generate LUT of BRF corresponding to inputprosail, for a sensor brf_lut <- Generate_LUT_BRF(sailversion = sailversion, inputprosail = inputprosail, specprospect = specprospect, specsoil = specsoil, specatm = specatm) # write parameters LUT output <- matrix(unlist(inputprosail), ncol = length(inputprosail), byrow = FALSE) filename <- file.path(path_results, "PROSAIL_LUT_InputParms.txt") write.table(x = format(output, digits = 3), file = filename, append = FALSE, quote = FALSE, col.names = names(inputprosail), row.names = FALSE, sep = "\t") # Write BRF LUT corresponding to parameters LUT filename <- file.path(path_results, "PROSAIL_LUT_reflectance.txt") write.table(x = format(t(brf_lut), digits = 5), file = filename, append = FALSE, quote = FALSE, col.names = specprospect$lambda, row.names = FALSE, sep = "\t") # Which bands will be used for inversion? if (is.null(bands2select)) { bands2select <- list() for (parm in parms2estimate) { bands2select[[parm]] <- seq(1, length(specprospect$lambda)) } } # Add gaussian noise to reflectance LUT: one specific LUT per parameter if (is.null(noiselevel)) { noiselevel <- list() for (parm in parms2estimate) { noiselevel[[parm]] <- 0.01 } } # produce LIT with noise brf_lut_noise <- list() for (parm in parms2estimate) { brf_lut_noise[[parm]] <- brf_lut[bands2select[[parm]], ] + brf_lut[bands2select[[parm]], ] * matrix(rnorm(nrow(brf_lut[bands2select[[parm]], ]) * ncol(brf_lut[bands2select[[parm]], ]), 0, noiselevel[[parm]]), nrow = nrow(brf_lut[bands2select[[parm]], ])) } ###===================================================================### ### PERFORM HYBRID INVERSION ### ###===================================================================### # train SVR for each variable and each run modelsvr <- list() for (parm in parms2estimate) { colparm <- which(parm == names(inputprosail)) inputvar <- inputprosail[[colparm]] modelsvr[[parm]] <- prosail_hybrid_train(brf_lut = brf_lut_noise[[parm]], inputvar = inputvar, figplot = figplot, nbensemble = nbmodels, withreplacement = replacement) } return(modelsvr) } #" writes ENVI hdr file #" #" @param hdr content to be written #" @param hdrpath Path of the hdr file #" #" @return None #" @importFrom stringr str_count #" @export write_envi_header <- function(hdr, hdrpath) { h <- lapply(hdr, function(x) { if (length(x) > 1 || (is.character(x) && stringr::str_count(x, "\\w + ") > 1)) { x <- paste0("{", paste(x, collapse = ","), "}") } # convert last numerics x <- as.character(x) }) writeLines(c("ENVI", paste(names(hdr), h, sep = "=")), con = hdrpath) return(invisible()) }