Added support for the per-ROI cell_count data produced by the diatom chain counter (Groves et al. 2026, doi:10.1093/plankt/fbaf064) through the ifcb-pytorch-classify pipeline and stored in .mat, .h5 and .csv classification files. Abundance can now be reported in cells rather than images, so a chain of eight cells counts as eight.
New ifcb_summarize_cell_counts() reports cell abundance and chain-length statistics (n_counted, mean, median, max, sd) per sample and class, and abundance per liter when given an hdr_folder. n_counted is the number of ROIs the chain counter measured, including those it found to hold a single cell.
single_cell_values (default c(-1, 0)) sets which cell_count values are read as one cell: by default the ROIs the counter skipped (-1) and those where it found no cells (0). Any other value is used as it stands.
cell_counts is NA rather than 0 for a sample whose classification file carries no cell_count data, since a zero would look the same as a taxon that was genuinely absent. counts still reports the images, and a warning says how many files were affected. A missing value inside a file that does carry cell_count data, such as a blank CSV cell, also makes the affected sample’s cell_countsNA, and is likewise reported with a warning naming the samples and how many ROIs were missing a value.
The SHARK column template (ifcb_get_shark_colnames() and ifcb_get_shark_example()) gains an IMAGE_COUNT column after COUNT. Dividing COUNT by IMAGE_COUNT gives the average cells per image. That is not the same as mean_chain_length, which averages only over the images the counter actually measured, whereas the ratio includes the images it never looked at.
ifcb_extract_biovolumes() and ifcb_summarize_biovolumes() gain two arguments controlling how the Menden-Deuer and Lessard (2000) carbon equations are applied. Both keep the previous behavior by default, so carbon values do not change unless you set one.
diatom_equation chooses the diatom equation. "large" (the default) uses the large-diatom equation, as ifcb-analysis does; "all" uses the all-sizes equation, which gives small cells more carbon; "auto" chooses per ROI by volume, keeping each equation inside the size range it was fitted for. The two disagree at the 3000 micron^3 boundary, predicting about 190 against 135 pgC, so "auto" makes carbon fall as a cell grows past it. That is why it is not the default. New helpers vol2C_diatom() and vol2C_diatom_auto() are exported.
carbon_conversion chooses the volume the equation is applied to. The equations are fitted per cell, but an IFCB biovolume covers a whole image, which for a chain-forming diatom is the whole chain, so converting the chain volume in one go under-reports carbon. "cell" converts per cell and sums over the chain, and requires use_cell_counts = TRUE. See ?ifcb_extract_biovolumes for how large the difference gets.
Biovolume itself is unchanged by either argument. It is still measured per image, so a chain of small cells still gives one large biovolume.
Added ifcb_qc_sample(), which checks raw IFCB samples (the .hdr, .adc and .roi triplet) and returns one row of QC metrics and flags per sample. It looks for missing files, ROI counts that disagree with the header, truncated .roi files, run times that contradict the ADC, implausible analyzed volumes, and ROI dimensions that failed to parse. Bead runs, empty samples, oversized .roi files and high recorded humidity or temperature are reported as advisory flags and do not fail qc_pass. A check that cannot be run on a sample is reported as NA and does not fail it either, so legacy headers without a roiCount field are not penalized for a check that could never apply. The function accepts a directory, sample names with a data_folder, or explicit file paths, and needs no Python.
ifcb_extract_features() gains a multiblob argument. With multiblob = TRUE it additionally writes multiblob/<bin>_multiblob_v4.csv files inside the features folder, holding the per-blob features of every blob in a region of interest with more than one blob (the slim table describes only the largest, plus summed* totals), one row per blob with roi_number, blob_number and 18 morphological columns. This is the sidecar output ifcb-features introduced in v1.2.0, which is also the minimum version required; the function stops with a clear error on older releases, since they never compute per-blob features. As upstream, a bin in which no ROI has more than one blob gets no sidecar file, so the file’s presence means the bin genuinely holds multi-blob ROIs; the skip logic reads the numBlobs column of a bin’s existing feature CSV to tell whether a sidecar is expected, so a re-run with multiblob = TRUE over a directory extracted without it re-extracts only the bins that actually need one, without overwrite = TRUE. ifcb_read_features() reads these files as it stands, with multiblob = TRUE; its multiblob filter now matches on the file name rather than the whole path, so a feature folder located somewhere under a directory with multiblob in its name no longer reads as all-multiblob (or, with the default multiblob = FALSE, as empty).
Minor improvements and fixes
ifcb_save_classification() now reports a missing roi_file before requiring hdf5r. Its tests no longer depend on the suggested hdf5r package, fixing the CRAN check failures on platforms without it.
Tests using use_python = TRUE now skip when scipy is unavailable, instead of letting reticulate download an ephemeral Python environment mid-check.
The native MAT reader no longer runs out of memory on a truncated compressed section when R is built against zlib-ng (the system zlib on recent Fedora), where memDecompress() grows its buffer indefinitely instead of erroring. Such sections now fall through to the incremental recovery path as intended.
ifcb_classify_images(), ifcb_classify_sample() and ifcb_classify_models() now retry transient network failures (dropped connections and HTTP 429/5xx responses) up to four times with exponential backoff. A brief server outage, such as the hosted SciLifeLab Serve instance restarting, used to fail every image it touched with Couldn't connect to server, leaving a run of NA rows mid-sample. Set options(iRfcb.gradio_max_tries = ) to change the number of attempts; a malformed value for either retry option is reported by name rather than surfacing as a cryptic error elsewhere.
ifcb_extract_biovolumes(), ifcb_summarize_biovolumes() and ifcb_summarize_cell_counts() now stop with an error when one sample resolves to more than one classification file, for example a folder holding both a .mat and an .h5 for the same sample, and name the samples involved. Previously both files were read and joined, which silently doubled that sample’s counts, biovolume and carbon.
The classification-file readers no longer trip over non-class .csv files in a class directory. An IFCB Dashboard class_scores export ({sample}_class.csv) could be picked up and then fail with a confusing Unknown or uninitialised column: 'class' error followed by a WoRMS 400 response. Such files are now skipped with a warning naming the missing columns when a folder is supplied, and raise a clear error when passed explicitly. A valid label file named {sample}_class.csv now resolves to the same sample as {sample}.csv, so it joins under the right sample instead of appearing as a phantom {sample}_class sample no HDR file could match. Relatedly, an hdr_folder whose files match none of the classified samples now warns and reports NA per-liter values instead of aborting with an unexplained join error, and the classifier column read from .mat files is a plain character vector rather than a one-column matrix.
ifcb_is_diatom() gains a details argument. When TRUE it returns a data frame with the WoRMS class resolved for each taxon instead of a logical vector. Use it to find genus homonyms, that is diatom genera such as Navicula or Actinocyclus that share a name with an animal and so resolve to a non-diatom class, then add those taxa to diatom_include. A class list in which nothing resolves in WoRMS at all, such as one holding only labels like unclassified or detritus, returns NA classes instead of failing.
ifcb_extract_biovolumes() and ifcb_summarize_biovolumes() report the diatom classification more clearly when verbose = TRUE. The classes treated as diatoms are listed in full, classes missing from WoRMS are listed separately, and the long list of non-diatoms is reduced to a count. Previously that list was printed and truncated, making it hard to see whether a class you expected to be a diatom had been recognized as one.
Removed R.matlab as a dependency. Reading .mat files without Python now uses the native R reader as well, which decodes MATLAB UTF-16 text correctly, so accented class or path names survive where R.matlab::readMat() could mangle them. It reads every numeric storage type MATLAB uses to hold an array compactly (int8 through uint32, single and double), preserving each across a read-write round-trip, including values at the signed and unsigned 32-bit limits that R cannot hold in an integer. Multi-row character matrices, such as the filelistTB sample list that ifcb-analysis summary files hold as one fixed-width row per sample, are read as one string per row, so ifcb_read_summary() handles multi-sample summaries the same way with either reader. The reader also refuses input it cannot represent faithfully, naming the variable and the reason, reports a truncated file instead of reading it back quietly padded with zeros, and rejects declared dimensions that disagree with the data carried rather than recycling values to fill them out. The writer applies the same standard: a value that does not fit the storage type a variable was read with is refused rather than silently wrapped, and ifcb_correct_annotation() and ifcb_replace_mat_values() widen a classlist to double when a new class id outgrows the compact integer type MATLAB stored it in. All of this matters because ifcb_adjust_classes() and ifcb_correct_annotation() write what they read back over the same file. One thing it tolerates rather than refuses: a compressed section that ends without its stream terminator, which some classification files written by MATLAB contain. The data in those is intact, so it is decoded incrementally and read, with a warning naming the file. R.matlab has moved to Suggests.
Note that this makes reading stricter than in 0.9.0, so a file that used to open may now stop with an error. R.matlab would read a .mat containing a struct, an object, a sparse or complex array, a logical array, more than two dimensions, or a MATLAB string array (as saved by class2use = ["a" "b"] in a recent MATLAB release), and iRfcb would then quietly write back something that was not what it read. Those files are now named and refused instead. The formats iRfcb itself deals with are unaffected: manual files, class2use files and classifier output written by MATLAB ifcb-analysis or the Python pipelines all read as before. If you do have a third-party file that no longer opens, ifcb_get_mat_names() and ifcb_get_mat_variable() read it with use_python = TRUE, which goes through SciPy instead, once a Python environment has been set up with ifcb_py_install().
ifcb_extract_features() now works with both raw-data readers used by WHOI ifcb-features. Release v1.1.0 swapped pyifcb for the lighter ifcbkit, which broke iRfcb against v1.1.0 and later, and since ifcb_py_install(features = TRUE) installs the newest release by default this affected new installations. Either reader now works, and a new backend argument (or the IRFCB_IFCB_BACKEND environment variable) pins one. Measurements are unaffected for the D-style bins current instruments produce. The readers differ only on older I-style bins and on ROIs with zero height, so pin a reader if you need results comparable to an earlier I-style run. Installing v1.1.0 or later also drops the h5py dependency, which had restricted which Python versions could be used.
ifcb_extract_features() now works with ifcb-features v1.2.0, which changed what compute_features returns to carry the new multiblob output. Without this change every ROI failed to unpack, and because a failed ROI is skipped rather than fatal, the extraction would report success while writing feature files holding only roi_number and no blob images. All earlier ifcb-features releases still work, and the extracted measurements are unchanged. The per-blob multiblob rows themselves are not used, since they are not part of the slim feature set.
Fixed ifcb_extract_features() writing text instead of numbers into the Eccentricity, MajorAxisLength and MinorAxisLength columns of <bin>_features_v4.csv. On numpy 2.3 and later these values arrived as complex numbers and were written as strings such as (0.797+0j), quietly turning three numeric columns into text. They are numeric again and the values are unchanged, with one exception worth knowing about for size spectra and biovolume sums: a degenerate, one-pixel-wide blob is now measured as 0 rather than NaN, so it contributes a zero instead of a missing value you could filter out. That matches upstream ifcb-features and the summed* columns. To reproduce an earlier run, pin ifcb-features v1.0.0 or numpy < 2.3; see ?ifcb_py_install.
ifcb_extract_features() no longer prints a FutureWarning for every region of interest from recent scikit-image releases, which had been breaking up the progress bar. ifcb_py_install(features = TRUE) also holds scikit-image below 0.28, the release that removes the deprecated functions ifcb_features calls.
ifcb_extract_features() no longer discards a whole sequential run when one bin cannot be read. A corrupt or truncated .roi used to escape as a Python traceback, taking every bin already processed with it. Such a bin is now reported as a per-bin error like any other.
ifcb_extract_features() gains a feature_tag argument controlling the feature file name. The default ("features") writes <bin>_features_v4.csv as before; "fea" writes <bin>_fea_v4.csv, the name the IFCB Dashboard serves.
Fixed ifcb_volume_analyzed_from_adc() failing when given more than one ADC file, or a URL. The existence check was not vectorized, so a vector of files stopped with the condition has length > 1 even though the rest of the function already looped over them, and it rejected URLs outright even though the rest of the function handles remote files. Local paths are now checked one by one, with every missing path reported at once, and URLs pass straight through.
Realigned ifcb_get_runtime() and ifcb_volume_analyzed_from_adc() with their MATLAB ifcb-analysis counterparts (IFCBxxx_readhdr.m, IFCB_volume_analyzed_fromADC.m), fixing five places where the R port had drifted from the reference.
Legacy headers, which carry run time and inhibit time on a single line, were read with the inhibit time set to a copy of the run time, so every legacy sample reported a look time of zero and an analyzed volume of zero. The two values are now read from their own positions on the line, as the reference does.
Header keys are matched at the start of the line, as the reference’s strmatch does. A header holding an extra key that merely ends in runtime: used to make ifcb_get_runtime() return two values, which ifcb_qc_sample() then recycled into a duplicated, failing row for that sample.
The startup-offset correction for instruments whose run/inhibit clocks lead the ADC timestamp called mode(), which in R is base::mode() and not the statistical mode the MATLAB reference computes, so the correction always failed with non-numeric argument to binary operator on exactly the instruments it was written for.
Corrupted inhibit-time rows are repaired the way the reference repairs them, from the last well-behaved row plus one typical increment per bad row, instead of using the corrupted final value as it stands. The timestamp-derived run-time fallback is now applied when the run-time clock disagrees with it, where previously it was computed and then overwritten.
A volume that cannot be derived, for example from an ADC without run/inhibit time columns, is now NA with a warning instead of a silent 0, matching the NaN the reference returns. A zero looked exactly like an instrument that analyzed no water, and made ifcb_qc_sample() fail such samples for an implausible volume.
ifcb_correct_annotation() and ifcb_replace_mat_values() now check their inputs and fail with a message you can act on. An out-of-range ROI number or column_index reports the value and the valid range instead of subscript out of bounds, a classlist with fewer than two columns is reported as such, and a missing input file or a .mat file without a classlist variable is named explicitly. ifcb_adjust_classes() likewise passes the reader’s own refusal message through when it skips a file, instead of relabeling every problem “empty or corrupted”.
Fixed use_python = TRUE being ignored when Python and SciPy were in fact available. ifcb_get_mat_names(), ifcb_get_mat_variable(), ifcb_read_summary(), ifcb_count_mat_annotations(), ifcb_extract_annotated_images() and ifcb_adjust_classes() decided whether they could use Python by looking for scipy in reticulate::py_list_packages(), and fell back to the R reader without saying so when it was not listed. Two things went wrong with that: on a conda environment the listing reports conda’s own base packages, so an installed and perfectly importable scipy was absent from it; and the check did not initialize Python, so in a fresh session it failed whatever the environment held. Availability is now settled by asking Python to import the module, and when use_python = TRUE is requested but Python with scipy is genuinely unavailable, the fallback to the R reader is announced with a warning (once per session) instead of happening silently. This matters most when a .mat file cannot be read by the R reader, since use_python = TRUE is the documented way to read it.
Clarified the deprecation of ifcb_read_hdr_data(hdr_folder = ) and ifcb_annotate_batch(adc_folder = ). Both read as though the package had renamed hdr_folder and adc_folder everywhere, which it has not: those two functions were changed to accept a vector of file paths, so their argument was renamed to match. Functions that genuinely take a single directory, such as ifcb_psd(), ifcb_summarize_biovolumes(), ifcb_summarize_cell_counts() and ifcb_annotate_samples(), keep hdr_folder and adc_folder and are not deprecated.
Corrected the vol2C_lgdiatom() documentation, which said the relationship applied to diatoms above 2000 micron^3. The Menden-Deuer and Lessard (2000) large-diatom equation is for cells above 3000 micron^3.
