[general]
name=PARACUDA-NG
qgisMinimumVersion=3.30
qgisMaximumVersion=4.99
description=Spectral machine learning: build, validate and apply regression models for spectroscopy and multispectral / hyperspectral imagery, integrated with QGIS raster layers.
about=PARACUDA-NG (PARAmetric CUbe Data Analysis, Next Generation) is a spectral machine-learning toolkit for spectroscopy and multispectral / hyperspectral remote sensing, embedded in QGIS as a native, dockable panel. A 7-step wizard (Data, Configuration, Preprocess, Model, Validate, Execution, Apply) takes you from loading spectral data through spectral resampling to a selectable sensor grid (Sentinel-2, Landsat 5/7/8/9, EnMAP, EMIT, PRISMA, DESIS, PlanetScope, VENuS and more), preprocessing (smoothing, continuum removal, derivatives, absorbance, baseline correction), and training or comparing regression models (PLS-R, SVM, Ridge, Lasso, MLR, Elastic Net, Huber, Gradient Boosting, Gaussian Process, Random Forest, XGBoost) with optional Optuna hyperparameter tuning and cross-validation. A trained model can be applied directly to a loaded QGIS raster layer, writing the prediction back as a new layer with a matching colour map, or to an unknown tabular dataset. Additional tools include compositional (CLR/ALR/ILR) modelling for parts that sum to 100 percent, a label-permutation and spectral-mixing integrity suite, cross-sensor spectral harmonisation (transfer functions), a data converter for arbitrary instrument layouts, uncertainty analysis (leverage, Faber-Kowalski, bootstrap, quantile regression forest and conformal prediction intervals, validated with PICP, PIT and CRPS) that is saved with the model and carried into tabular predictions and per-pixel uncertainty rasters, and hover tooltips on every control. PARACUDA-NG needs a few Python packages QGIS does not ship (scikit-learn, joblib, optuna, xgboost, rasterio, spectral); it installs them into QGIS's Python on first run, or via Tools then Install Python dependencies.
version=2.0.0
author=Sharad Kumar Gupta, Jonti Evan Shepherd
email=sharadgupta27@gmail.com
tracker=https://github.com/sharadgupta27/paracuda-qgis/issues
repository=https://github.com/sharadgupta27/paracuda-qgis
homepage=https://paracuda-ng.github.io/
license=MIT
category=Raster
icon=icon.png
supportsQt6=True
hasProcessingProvider=False
experimental=False
deprecated=False
tags=spectroscopy,hyperspectral,multispectral,machine learning,regression,raster,remote sensing,soil,chemometrics,PLS,random forest,XGBoost,resampling,Sentinel-2,EnMAP,EMIT,PRISMA
changelog=2.0.0
    - Uncertainty analysis (step 5 Validate): leverage-based (ASTM E1655), Faber-Kowalski error propagation, bootstrap / bagging, quantile regression forest and conformal prediction intervals at 90, 95 or 99 percent confidence, each explained in one line, with an optional reference-lab error and instrument noise.
    - Training, cross-validation and test sets report bias, RMSEP, SEP, RPIQ, Lin's concordance and the RMSEP2 = SEP2 + bias2 split, plus interval checks (PICP, width, PIT, quantile coverage, CRPS, interval score) in the workbook, the PDF report and the Visualization tab.
    - The uncertainty model is saved with the model: step 7 Apply adds prediction intervals to tabular predictions and writes a per-pixel uncertainty raster (added to QGIS as a layer) with leverage and an outside-calibration-domain flag; files with reference values are validated automatically.
    - Optional SOC stock conversion with GUM + Monte Carlo uncertainty propagation and an uncertainty budget.
    - Hover tooltips on every button, checkbox, drop-down and field, sharing one text table with the desktop tool and the Python package.
    - The Mask background checkbox shows its ampersand again.
    - Corrected methods (results change; 1.x models keep their transform when loaded): Continuum Removal divides by the upper convex hull; linear Baseline Correction subtracts the lower convex hull; MSC reference from training rows only; derivatives use real band spacing; Optimize Components chooses by training-set CV; Leave-One-Out reports pooled RMSE/R2; failed CV shows no result instead of 0.
    - PLS-R never fits more components than the numerical rank of the spectra (extra components fitted rounding noise); default component count is 32 capped at the band count, as in the desktop tool.
    - Integrity tests use shuffled folds (a sorted file could look like "no signal"); permutation p-value is (b+1)/(m+1).
    - Uncertainty: bootstrap intervals no longer too narrow; finite-sample conformal quantile; exact PIT KS distance; tie-aware Spearman; out-of-fold fold count now matches the desktop tool.
    - Warning when an image or table is on a different scale from the training spectra (x100, x10000); micrometre image headers converted to nm and the ENVI reflectance scale factor applied; empty cells in unknown data filled along the spectrum.
    - Prediction maps keep zero and negative values; image prediction uses far less memory.
    - Results workbook now has Train RPD, Train nRMSEP and the component-optimisation sheet, matching the desktop tool.
    - Fixed: PCA crashes, K-Fold with Shuffle off, Spectral Outlier Removal models on images, image without header wavelengths, Data Converter dropping sub-micrometre bands, Excel export with long or '/' names, Dark-theme table striping.
    - Models saved by the desktop tool or the Python package open in the plugin, and the reverse (the uncertainty model inside a 2.0 bundle used to block this).
    - Desktop tool, Python package and plugin verified to produce bit-identical results on the sample data.
    1.0.2
    - Raster-layer filter resolves the QGIS 3.34+ scoped enum first and only falls back to the pre-3.34 unscoped member, so the Qt6 compatibility check is clean while QGIS 3.30-3.33 keeps working.
    - Renamed an ambiguous single-letter local (E741) and replaced a bare except (E722) in the models backend.
    1.0.1
    - Workbook XML in the Data Converter's fast .xlsx reader is now parsed with defusedxml, which refuses the entity declarations behind "billion laughs" expansion; without defusedxml installed the reader stands down and pandas reads the file as before.
    - Replaced every silent try/except/pass with contextlib.suppress so the intent is explicit, and documented the two subprocess call sites that launch pip.
    1.0.0
    - First QGIS plugin release: the full 7-step spectral-ML wizard as a native, dockable Qt6 panel (QGIS 3.30+ and QGIS 4).
    - Apply a trained model to a loaded QGIS raster layer and add the prediction back as a new layer with a matching colour map.
    - Data / Configuration / Preprocess / Model / Validate / Execution / Apply steps with a live Model Development Flow diagram (exportable at 300 DPI).
    - Sensor resampling for Sentinel-2, Landsat 5/7/8/9, EnMAP, EMIT, PRISMA, DESIS, PlanetScope and VENuS, plus custom FWHM/SRF grids and band binning.
    - Single and batch model runs, Find Best Preprocessing search, Optuna tuning, K-Fold / Leave-One-Out / Leave-P-Out cross-validation, and compositional (CLR/ALR/ILR) modelling.
    - Check Spectral Integrity (label-permutation and spectral-mixing tests) and Spectral Harmonization tools, plus a Data Converter.
    - Data loading runs off the UI thread behind a progress dialog and picks the fastest available reader, roughly halving the time to open a large spreadsheet.
    - Data Distribution inspects one property at a time from a drop-down, with a histogram, boxplot, normal Q-Q plot, cumulative distribution and findings.
    - Noisy-edge and water-band exclusion presets follow the selected spectral domain (VSWIR, LWIR or both) and merge into the Omit field instead of stacking duplicates.
    - Automatic first-run installation of the required Python packages; five selectable themes (Ocean, Slate, Forest, Light-Contrast, Dark).
