Improving the Efficiency of Soil Geochemical Parameter Approximation by a Multilayer Perceptron Through Radial-Convergent Anisotropy-Based Training Set Formation (Case Study: Available Phosphorus)
This study investigates whether preprocessing training data using radially convergent anisotropy (RCA) improves the ability of a multilayer perceptron (MLP) to approximate the spatial distribution of soil geochemical parameters. The proposed approach accounts for the geometry of spatial variability in the parameter of interest to make the input features more informative. Numerical experiments were conducted for available phosphorus using a synthetic field measuring 5000 × 2500 m and comprising 125,000 grid nodes with known reference values. Training datasets containing 50, 100, 400, and 800 geochemical sampling points were preprocessed using RCA before MLP training. Prediction accuracy was evaluated by comparing predicted and reference values at all grid nodes using RMSE, MAE, and R2. Kriging served as the benchmark for spatial prediction. With the smallest training dataset of 50 points, the differences between the RCA-preprocessed MLP and kriging were no greater than 12% for either RMSE or MAE. With 800 training points, the RCA-preprocessed MLP achieved a 13% lower RMSE and an 18% lower MAE than kriging. These results demonstrate the effectiveness of the proposed approach for the synthetic field studied and suggest that RCA can be used to prepare neural network training datasets for digital agrochemical mapping and precision agriculture.
Authors
- Akhmedov Daulet
- Moldir Tanashova (ORCID: https://orcid.org/0009-0007-9036-8923)
- Baurzhan Bekmukhamedov (ORCID: https://orcid.org/0009-0008-3893-047X)
- Madina Batyrbayeva
Institutions
- Al-Farabi Kazakh National University (KZ)
- Almaty Technological University (KZ)
- National Center of Space Research and Technology (KZ)
Publication Details
- Journal
- Computation
- Published
- 2026-11-20
- DOI
- https://doi.org/10.3390/computation14110302
- Primary Topic
- Soil Geostatistics and Mapping
- Type
- article
- Field-Weighted Citation Impact
- 0.00