Dataset: Remote sensing-based prediction of permanganate oxidisable carbon (POXC) in organic cassava plots, Northeast Thailand

Supporting data, code, and figures for: Can Multi-source Remote Sensing Detect Labile Soil Organic Carbon? A Spatially Explicit Evaluation in Organic Cassava Systems, Northeast Thailand (Aumtong, S., Pliumchareorn, C., Damri, N., Kantamang, K.; submitted to Remote Sensing Applications: Society and Environment). This repository contains the complete plot-level dataset, analysis code, and figures needed to reproduce all results in the manuscript. We evaluated whether multi-source remote-sensing (RS) features can predict labile and bulk soil organic carbon across 248 organic cassava plots in Yasothon, Amnat Charoen, and Ubon Ratchathani provinces (dry season 2023–2024; Acrisols/Ultisols, sandy-loam). RS predictors combined Sentinel-2 optical, Sentinel-1 SAR, MODIS land-surface temperature, and SRTM terrain (24 features retained from 39 extracted). Permanganate-oxidisable carbon (POXC; KMnO₄ method) and total organic carbon (Walkley–Black) were the response variables, with four exploratory hypotheses (H1: LDA of management-age groups; H2: bootstrapped mediation of age → NDVI → POXC; H3: Soil Order moderation of age → TOC; H4: Random Forest prediction under both random and spatially explicit cross-validation). Key finding (methodological). Under random 5-fold CV, POXC appeared modestly predictable (R² = 0.245, RMSE = 0.311 g kg⁻¹), but this skill did not survive spatially explicit (leave-block-out) validation: R² = −0.13 for Random Forest and negative across all algorithms (XGBoost −0.20, SVR −0.21, Ridge −0.43), consistent with strong positive spatial autocorrelation (Moran's I = 0.48). Bulk TOC was not predictable (R² = −0.057). LDA discrimination was marginal (CV accuracy 46.3 ± 4.9%, permutation p = 0.035) and the NDVI mediation was borderline (95% CI [−0.015, +0.001], touching zero). Random CV substantially overstates the RS-detectability of labile carbon in clustered tropical sampling designs; spatially explicit validation is essential. Contents. Plot-level dataset (soil properties + RS features + data dictionary); Google Earth Engine extraction script; a single unified Python pipeline reproducing all H1–H4 results plus a figure-regeneration script; canonical results table; spatial-validation table; QC log; an anonymised fertiliser-input summary for the 15-farm purposive subsample; and all six manuscript figures. A full file list and methods are in README_RSASE_Dataset.txt. Privacy. All files are anonymised. The master soil file containing farmer names is retained privately by the research team and is not deposited. Funding: Agricultural Research Development Agency (ARDA), Thailand — Project PRP6707031410.License: Creative Commons Attribution 4.0 International (CC BY 4.0).