A novel approach to soil nutrients prediction model for Bezuidenhout Park, Johannesburg, Gauteng Province, South Africa: attention temporal neural networks (ATNN)

Description

Dada and colleagues predict soil nutrient levels for a Johannesburg park in 2026 with a hybrid model that pairs an attention-based temporal neural network with gradient boosting. Farmers need field-level nutrient maps to dose fertilizer well, and the model digests elevation data, satellite imagery from two missions, and weather records together. The attention mechanism lets the network weigh which time points and signals matter for each nutrient.

For aluminium the hybrid posted a prediction error near 1.98 parts per million, an average percentage error of 2.81 percent, and a fit score of 0.69 against laboratory measurements. Agreement with lab values reached 0.76 on the concordance scale, and the authors report clear gains over random forest, plain gradient boosting, and AdaBoost baselines across the nutrients tested. They pitch the pairing as a template for digital soil mapping that can guide fertilizer decisions across South African farms facing thin sampling.

Catalog record

Source: https://europepmc.org/articles/PMC13482613