Soil fertility prediction is crucial for sustainable agriculture, yet data scarcity in climate-vulnerable regions hampers machine learning (ML) model development. We introduce a conditional diffusion model tailored for generating synthetic soil fertilitydatasets under extreme climate scenarios (e.g., drought, heatwaves). By conditioning on scenario embeddings, our model produces high-fidelity synthetic samples that capture nutrient correlations and temporal dynamics. We demonstrate thataugmenting real datasets with our synthetics improves downstream crop recommendation classifiers by 15–20% in accuracy and F1-score, particularly in low-sample regimes. Evaluations on public Kaggle datasets show superior performance overbaselines like SMOTE and CTGAN, with Fréchet Inception Distance (FID) scores below 10 and Kolmogorov-Smirnov (KS) p-values indicating distributional fidelity. This approach offers a zero-cost, scalable solution for equitable ag-tech, aligningwith UN SDG 2 goals.
S M Shahriar Hossain (Sat,) studied this question.