Estimating soil erodibility and soil aggregate stability from basic soil properties using statistical and machine  learning techniques for kandi region of Punjab

Manpreet Singh

Estimating soil erodibility and soil aggregate stability from basic soil properties using statistical and machine learning techniques for kandi region of Punjab

Date

2022

Authors

Manpreet Singh

Publisher

Punjab Agricultural University

Abstract

Accelerated soil erosion in the lower Shiwalik kandi region of Punjab results in deterioration of soil physical quality. Soil aggregation and soil erodibility are important parameters indicating soil physical quality but the quantification of these parameters is complex so efforts have been done to estimate these properties from easily measurable soil characteristics by using pedo-transfer functions (PTFs). Statistical PTFs are available for estimating of these properties but they don't explainthe sufficient variability in data. Machine learning techniques may play an important role in this context. Therefore, the present study was planned to compare PTFs developed using statistical and machine learning techniques for kandi region of Punjab. The basic soil physico-chemical properties were measured across four locations with five land uses at each location at three depths and with three replications. Three data sets were prepared for these soil properties. When dataset 1, having six basic soil properties, was used for estimation of mean weight diameter (MWD), water stable aggregates (wSA) and erodibility (K), the prediction using artificial neural network (ANN) was slightly better than generalized linear model (GLM). When dataset 2, having those six basic soil properties which were having high correlation with soil structural parameters, was used for estimation of MWD, WSA and K, the prediction using GLM was slightly better than ANN. When dataset 3, having all 11l basic soil properties, was used for estimation of MWD, WSA and K, the prediction using ANN was significantly better than GLM. When using data available from literature, ANN performed better for prediction of MWD and WSA whereas GLM performed better for prediction of K. So, it may be concluded that ANN performs better for a large set of data and for a complex system having a greater number of variables whereas for small set of data and for simple system having less variables, the statistical methods perform better.

Citation

Manpreet Singh (2022). Estimating soil erodibility and soil aggregate stability from basic soil properties using statistical and machine learning techniques for kandi region of Punjab (Unpublished M.Sc. thesis). Punjab Agricultural University, Ludhiana, Punjab, India.