Biochar application age in tropical longan orchards: soil physical, chemical and carbon data (version 3)

# Biochar application age in tropical longan orchards: soil physical, chemical and carbon data (version 3) Soil physical, chemical and carbon data from a three-year chronosequence following a singleapplication of EM-inoculated longan-wood biochar in twelve smallholder longan orchards at Mae Pang,Phrao District, Chiang Mai, northern Thailand. All orchards lie on one soil series, the Mae Taengseries (Typic Hapludults, Ultisols). The design is space-for-time. Twelve orchards fall into four classes by time since biochar wasapplied — 0, 1, 2 and 3 years, with three orchards per class; the Year-0 orchards never receivedbiochar and provide the reference. Within each orchard three trees were sampled at four depths(0–15, 15–30, 30–45, 45–60 cm), giving 144 composite samples. Biochar was applied once at 15 kg pertree, placed in a trench cut around each tree at roughly 1 m from the trunk. Because each orchard belongs to exactly one application-age class, orchard identity is confoundedwith the treatment. Any model trained on these data should be validated with the orchard as thegrouping unit, and the identifiers needed to do that are included. ## What is new in version 3 **The carbon basis is corrected, and the correction published in version 2 was wrong.** Version 2 stated that soil organic carbon had been recomputed "using the corrected factor", givingvalues 9.1 % lower than version 1. That change replaced a constant of 1.43 in the Walkley–Blackformula with 1.30, on the understanding that 1.43 was an oxidation-recovery factor. It is not. In thelaboratory worksheet 1.43 sits in the position of the **normality of the ferrous ammonium sulfate**,which is a different quantity entirely, and substituting a recovery factor for it was an error. The ferrous ammonium sulfate actually used was a **0.5 M** solution — the same solution standardisedper batch to 0.5051–0.5128 for the particle-size fraction analyses deposited alongside this dataset(10.5281/zenodo.21423736). Version 3 uses 0.5. | Quantity | Multiplier from version 1 | Multiplier from version 2 | Net effect ||---|---|---|---|| `SOC_g_kg` and the two carbon stocks | 0.5 / 1.43 = 0.34965 | 0.5 / 1.30 = 0.38462 | **65.0 % below version 1** || `SOM_pct` | 0.5 / 0.6993 = 0.71500 | unchanged in v2 | **28.5 % below versions 1 and 2** | Soil organic matter carried the same class of error, with 0.6993 in the normality position on its ownworksheet, and was not corrected in version 2 at all. **No statistical result changes.** Each correction is a single constant multiplier. Random Forest isinvariant to rescaling one predictor, and the four cross-validation outcomes reproduce from this fileexactly as published: four-class stratified 0.798, four-class leave-one-orchard-out 0.350, binarystratified 0.952, binary leave-one-orchard-out 0.855. **The file now carries only what the analysis uses.** Sixteen columns were removed, taking it from 39to 23: - Four labile-carbon pools and their percentages of SOC (`VLB`, `LB`, `LLB`, `NLB` in both forms) were internally impossible. As percentages of SOC they summed to 265 %, with one pool alone exceeding 100 %; in g kg⁻¹ they summed to 39.2 against a total SOC of 19.1.- `SOC_over_clay_silt` was not a carbon quantity despite its name: across all rows it correlated with `clay_g_kg` at r = +1.000 and with SOC at r = +0.003.- `NH4_mg_kg` and `NH4_over_P` — ammonium is not reported anywhere in the associated paper, and these values carried an unexplained 10.2 mg kg⁻¹ offset against the laboratory worksheet together with 36 blank rows.- `C_storage_Mg_ha`, `WEOC_over_SOC`, `WEOC_pct_of_SOC` and `SOC_over_clay` do not appear in the paper.- `SOC_gkg_previous_1p43`, retained in version 2 for transparency, is superseded: the basis it preserved was itself wrong.- `sample_code` is removed because codes run sequentially through the four application-age blocks, so including it as a feature leaks the class label and inflates every accuracy. Orchard, tree and depth together identify each row uniquely. What remains: four identifiers, the classification target, the fifteen model predictors, and the threequantities the paper reports without modelling (`SOM_pct` and the two carbon stocks). **Documentation corrected in version 2 and retained here.** Water-extractable organic carbon is anambient-temperature extraction shaken 16 h, not a hot-water extraction at 80 °C. Soil organic carbonand soil organic matter are independent Walkley–Black determinations on differently sieved soil andare not interconvertible by the Van Bemmelen factor. Sulfate values are verified against the originalabsorbance records for all samples. ## Notes for reuse Sample 224 lacks a saturated hydraulic conductivity measurement and was excluded from the analyses,leaving n = 143; it is not present in this file. Soil organic carbon and water-extractable organic carbon both come from titrations and fall ondiscrete steps — approximately **0.30 g kg⁻¹** and 97.5 mg kg⁻¹ respectively on the corrected basis, acoarser resolution than the reported decimal places imply. Three samples have soil organic carbon below 1 g kg⁻¹ and one water-extractable organic carbon valueis recorded as zero, which should be read as missing rather than as a true absence. Both carbon-stockcolumns are blank on 107 rows, where a profile lacks one of the increments the cumulative figurerequires. Silt and clay differ significantly between application-age classes. Texture cannot respond to anamendment within three years, so this is direct evidence that the orchards were unequal before anybiochar was applied, and it is why the leave-one-orchard-out result is so much weaker than thestratified one. Anyone reusing these data for modelling should treat the four-class age signal aspartly inherited site variation. ## Files | File | Contents ||---|---|| `biochar_longan_chronosequence_v3.csv` | 143 rows, 23 columns || `biochar_codebook_v3.csv` | one row per column, each labelled predictor, target, identifier or reported-only, with method and correction noted || `reproduce_analysis.py` | reproduces all four cross-validation results over seeds 0–9 | Licensed CC BY 4.0.