antsaro/taneti-sfi-mapping

Nested-CV Random Forest and regression kriging pipeline for mapping soil fertility across three bioclimatic regions of Madagascar. # Taneti-SFI: Mapping Soil Fertility Across Madagascar's Open Ecosystems Digital soil mapping of a five-depth Soil Fertility Index (SFI) across three bioclimatic regions of Madagascar (Central Highlands, Northwest, Southeast), built on Random Forest tuned through nested cross-validation, cross-checked against a regression-kriging correction on the residuals (optional), and predicted wall-to-wall on GPU. This repository accompanies the manuscript *"Soil fertility variability under tropical open grassland ecosystems of Madagascar"* (Rafidimanantsoa, Ramifehiarivo et al., submitted to Geoderma Regional). See Citation below. ## What is being predicted The response variable is a Soil Fertility Index (SFI), modeled separately at five depth layers: 0-10, 10-20, 20-30, 30-60 and 60-90 cm. Modeling each depth as its own response, rather than fitting one model across all depths, lets the covariate relationships and hyperparameters adapt to how each layer behaves: surface fertility is shaped more strongly by recent land use and organic matter turnover, while deeper layers reflect longer-term parent material and drainage effects. Consistent with that, model accuracy in the underlying study declines steadily with depth (R² = 0.64 at 0-10 cm down to R² = 0.12 at 60-90 cm). ### How the SFI itself was built The SFI is not something this repository computes; it is the target variable, derived beforehand from lab-measured soil properties following the method of Dewi et al. (2024), adapted from Mukashema (2007). Seven physicochemical properties (pH, texture, cation exchange capacity, K, P, C, N) were reduced to six minimum soil fertility indicators (MSFI) after removing one variable from each strongly correlated pair ($|r| > 0.5$), then weighted through a Principal Component Analysis retaining components with eigenvalue $> 1$ (Kaiser criterion). For each observation $i$, the index is $$ \text{SFI}_i = \sum_{j=1}^{N} W_j \cdot S_{ij} \cdot p, \qquad p = \frac{1}{n} $$ where $N …