dez-data/sub-saharan-soil-analysis

Soil chemistry analysis and crop recommendation logic across sub-Saharan Africa using the iSDA dataset # Sub-Saharan Soil Analysis: Agricultural Recommendation Feature Engineering Analysis of the iSDA Soil Dataset, covering 42 sub-Saharan African countries at 30m resolution, built to evaluate how soil chemistry data can drive crop recommendation logic for agricultural advisory systems. ## Why This Dataset Soil nutrient profiles determine which crops can realistically succeed at a given site, yet most recommendation systems in the region work off country-level averages rather than measured site data. The iSDA dataset closes that gap: topsoil and subsoil readings for nitrogen, organic carbon, pH, calcium, magnesium, potassium, aluminium, iron, and manganese across the continent. That granularity is what makes site-specific, rather than regional, recommendation logic possible. Sourced from Kaggle, originally derived from iSDA Africa's machine-learning soil property predictions. ## What's in This Repo - **`notebooks/soil_analysis.ipynb`** — full analytical pipeline: cleaning, EDA, feature engineering, and a rule-based crop recommendation function, with code and inline insights. - **`reports/Soil_Analysis_Report.docx`** — formal write-up covering all five analysis stages plus an LLM-assisted insight generation exercise. ## Data Cleaning Six columns (boron, electrical conductivity, phosphorus, sodium, sulphur, zinc extractable) were fully null across the dataset and were dropped rather than imputed; imputing a column with zero observed values fabricates data rather than recovering it. `carbon_total` nulls were handled with median imputation where the gap was partial. Outliers in key nutrient columns were capped using the IQR method rather than removed, since soil data legitimately contains extremes at sites with unusual parent material, and removing those records would shrink geographic coverage exactly where more data is needed, not less. ## Key Findings Nutrient distributions are right-skewed across the board except pH, meaning most sites carry low-to-moderate …