No packages match

soilKey - Automated Soil Profile Classification per WRB 2022, 'SiBCS' 5 and USDA Soil Taxonomy 13

Implements deterministic classification keys for the World Reference Base for Soil Resources 2022 (4th edition) and the Brazilian System of Soil Classification ('SiBCS', 5th edition). Provides a unified profile representation with explicit per-attribute provenance, multimodal extraction from field reports and photos via vision-language models, spatial priors from 'SoilGrids' and national soil maps, and gap-filling of soil attributes from Vis-NIR or MIR spectra via the Open Soil Spectral Library ('OSSL'). The taxonomic key itself is never delegated to a language model; LLMs are restricted to schema-validated extraction. Each classification result reports a key trace, a provenance-aware evidence grade, and ambiguities that further measurement would resolve.

Last updated

7.28 score 7 stars 42 scripts 404 downloads

soilFlux - Physics-Informed Neural Networks for Soil Water Retention Curves

Implements a physics-informed one-dimensional convolutional neural network (CNN1D-PINN) for estimating the complete soil water retention curve (SWRC) as a continuous function of matric potential, from soil texture, organic carbon, bulk density, and depth. The network architecture ensures strict monotonic decrease of volumetric water content with increasing suction by construction, through cumulative integration of non-negative slope outputs (monotone integral architecture). Four physics-based residual constraints adapted from Norouzi et al. (2025) <doi:10.1029/2024WR038149> are embedded in the loss function: (S1) linearity at the dry end (pF in [5, 7.6]); (S2) non-negativity at pF = 6.2; (S3) non-positivity at pF = 7.6; and (S4) a near-zero derivative in the saturated plateau region (pF in [-2, -0.3]). Includes tools for data preparation, model training, dense prediction, performance metrics, texture classification, and publication-quality visualisation.

Last updated

4.78 score 2 stars 7 scripts 504 downloads

soilVAE - Supervised Variational Autoencoder Regression via 'reticulate'

Supervised latent-variable regression for high-dimensional predictors such as soil reflectance spectra. The model uses an encoder-decoder neural network with a stochastic Gaussian latent representation regularized by a Kullback-Leibler term, and a supervised prediction head trained jointly with the reconstruction objective. The implementation interfaces R with a 'Python' deep-learning backend and provides utilities for training, tuning, and prediction.

Last updated

deep-learningsoil-propertiessoil-sciencesoil-spectroscopytensorflowvariational-autoencoder

4.60 score 8 scripts 360 downloads

GeoVersa - Design-Based Residual-Correction Forests for Digital Soil Mapping

Implements DB-TARF (Design-Based Targeted Adaptive Residual Forest) for large-scale digital soil and ecological mapping evaluated under the design-based paradigm of Wadoux et al. (2021) <doi:10.1016/j.ecolmodel.2021.109692>. A random forest is augmented by a cross-fitted, out-of-fold-selected residual correction (residual forests, ordinary kriging, recalibration), together with design-based conformal prediction intervals.

Last updated

convolutional-neural-networkdeep-learningdigital-soil-mappinggeostatisticskrigingpedometricspytorchremote-sensingspatial-statistics

3.48 score 1 stars 1 scripts