Identifying the Underlying Factors and Variables Governing Macronutrients in Cultivated Tropical Peatland Using Regression Tree Approach

Abstract

The capability of machine learning/ML algorithms to analyze the effect of human and environmental factors and variables in controlling soil nutrients has been profoundly studied over the last decades. Unfortunately, ML utilization to estimate macronutrients and their governing factors in cultivated tropical peat soil are extremely scarce. In this study, we trained regression tree/RT, ML-based pedotransfer models to predict total N, P, and K in peat soils based on oil palm/OP and OP+bush datasets. Our results indicated that the dataset might contain outliers, non-linear relationships, and heteroscedasticity, allowing RT-based models to perform better compared to multiple linear regression/MLR models (as a benchmark) in estimating total N and P in both datasets, contrastingly, not in K. The difference of important variables in each RT-based model partially showed the vital role of land use in nutrient modeling in peat. The depth of sample collection, organic C, and ash content became the prominent factor and variables in regulating the entire predicted nutrients. Meanwhile, the distance from the oil palm tree and pH were the salient features of total P prediction models in OP and OP+bush sites, respectively. This study proposed employing ML-based pedotransfer models in analyzing and interpreting complex tropical peat data as an alternative to linear-based regression. Our initial study also shed more light on the development possibility of the pedotransfer models that agricultural practician, researchers, companies, and farmers can use to predict macronutrients, both in tabular and spatial terms, in cultivated tropical peatlands

Prediction and Interpretation of Total N and Its Key Drivers in Cultivated Tropical Peat using Machine Learning and Game Theory

“…& Chambers (2021), the RT algorithm is easy to implement and captures non-linear relationships among covariates. Unfortunately, the RT-based PfF model's low predictive performance may arise from smooth, gradual changes in the continuous covariates, as can be observed in Pulunggono et al (2022a) results. Furthermore, this present study hypothesized that using more advanced ML algorithms such as ensemble trees (e.g., random forest/RF, gradient boosting machine/GBM, and extreme gradient boosting/XGB), which are considered more complex than RT algorithm, thereby can handle previous constraints and may improve N predictive performance.…”

“…A single block (around 30 hectares) is usually established as the smallest sampling unit for OPP management. Nevertheless, previous research highlighted the spatiotemporal variability of peat nutrients at sub-block scales, along with the gradients of the distance from the oil palm tree, canal and mineral soil border, peat thickness, sampling depth, oil palm age, season, and land use (Pulunggono et al 2016;Pulunggono 2019;Pulunggono 2020;Pulunggono 2021;Pulunggono 2022a;Pulunggono 2022b). Soil samples must be collected enormously to satisfy the sampling design for this detailed approach, especially for large-scale OPP.…”

“…Currently, ML-based PtFs are widely used in predicting soil properties, such as soil texture , soil temperature (Abimbola et al 2021), aggregate stability (Boushilim et al 2021, soil organic C/SOC (Sakhaee et al 2022), cation exchange capacity (Ma et al 2021), soil reaction (Baltensweiler et al 2021), available P (Kaya et al 2022) and other soil macro-and micronutrients (Shi et al 2022;Sihag et al 2019). Eventhough the research for cultivated tropical peatlands was scarce (e.g., Pulunggono et al 2022a), PtFs employed with ML algorithms have been successfully implemented in estimating N in a maize field (Mashaba-Munghemezulu et al 2021), rangelands (Parsaie et al 2021), paddy soils (Xu et al 2021) and other various land uses (Forkuor et al 2017;Hounkpatin et al 2022;Liu et al 2022) which is entirely located in mineral soils.…”

“…In a similar location to this study, Pulunggono et al (2022a) initially tried regression tree/RTbased PtF models for OP-cultivated tropical peat total N, P, and K contents. However, their report only described model performances and interpretation of a single type of tree-based ML algorithm and its comparison to multiple linear regression/MLR.…”

“…Selain itu, mengekstraksi informasi yang berarti dari model 'kotak hitam' ini sangat penting, terutama terkait kompleksitas algoritmik dan sifat hubungan kovariat tanah yang tidak linier. Penelitian ini menggunakan dataset Pulunggono (2022a) dan metode bootstrapping, untuk (1) mengembangkan dan mengevaluasi tujuh model PtF, termasuk model linear umum (GLM) dan regresi machine learning (ML) untuk mengestimasi total nitrogen (N) pada gambut tropis yang telah didrainase dan dibudidayakan untuk kelapa sawit (OP) di Riau, Indonesia, serta (2) menjelaskan fungsi model dengan menggabungkan Shapley Additive Explanation (SHAP), sebuah perangkat yang berasal dari teori permainan koalisi. Studi ini menunjukkan kinerja prediksi yang unggul dari PtF berbasis ML dalam mengestimasi total N dibandingkan dengan algoritma GLM.…”