Modelling Soil Ammonium Nitrogen, Nitrate Nitrogen and Available Phosphorus Using Normalized Difference Vegetation Index and Climate Data in Xizang’s Grasslands

There is still a lack of high-precision and large-scale soil ammonium nitrogen (NH 4 + -N), nitrate nitrogen (NO 3 − -N) and available phosphorus (AP) in alpine grasslands at least on the Qinghai–Xizang Plateau, which may limit our understanding of the sustainability of alpine grassland ecosystems (e.g., changes in soil NH 4 + -N, NO 3 − -N and AP can affect the sustainability of grassland productivity, which in turn may alter the sustainability of livestock development), given that nitrogen and phosphorus are important limiting factors in alpine regions. The construction of big data mining models is the key to solving the problem mentioned above. Therefore, observed soil NH 4 + -N, NO 3 − -N and AP at 0–10 cm and 10–20 cm, climate data (air temperature, precipitation and radiation) and/or normalized vegetation index (NDVI) data were used to model NH 4 + -N, NO 3 − -N and AP in alpine grasslands of Xizang under fencing and grazing conditions. Nine algorithms, including random forest algorithm (RFA), generalized boosted regression algorithm (GBRA), multiple linear regression algorithm (MLRA), support vector machine algorithm (SVMA), recursive regression tree algorithm (RRTA), artificial neural network algorithm (ANNA), generalized linear regression algorithm (GLMA), conditional inference tree algorithm (CITA), and eXtreme gradient boosting algorithm (eXGBA), were used. The RFA had the best performance among the nine algorithms. Climate data based on the RFA can explain 78–92% variation of NH 4 + -N, NO 3 − -N and AP under fencing conditions. Climate data and NDVI together can explain 83–93% variation of NH 4 + -N, NO 3 − -N and AP under grazing conditions based on the RFA. The absolute values of relative bias, linear slopes, R 2 and RMSE values between simulated soil NH 4 + -N, NO 3 − -N and AP based on RFA were ≤8.65%, ≥0.90, ≥0.91 and ≤3.37 mg kg −1 , respectively. Therefore, random forest algorithm can be used to model soil available nitrogen and phosphorus based on observed climate data and/or normalized difference vegetation index in Xizang’s grasslands. The random forest models constructed in this study can be used to obtain a long-term (e.g., 2000–2020) raster dataset of soil available nitrogen and phosphorus in alpine grasslands on the whole Qinghai–Tibet Plateau. The raster dataset can explain changes in grassland productivity from the perspective of nitrogen and phosphorus constraints across the Tibetan grasslands, which can provide an important basis for the sustainable development of grassland ecosystem itself and animal husbandry on the Tibetan Plateau.