Operational Value of Earth Observation and Reanalysis Predictors for Subnational Maize Yield Forecasting in Nigeria

Accurate, spatially explicit crop yield forecasts are critical for food security monitoring in climate-vulnerable regions, yet the operational value of different Earth observation (EO) data streams remains poorly quantified. Using a Random Forest model with a predictor-block ablation design evaluated under strict Leave-One-Year-Out cross-validation, this study isolates the marginal contribution of four EO inputs: MODIS NDVI phenology, CHIRPS rainfall, ERA5-Land air temperature, and MODIS land surface temperature (LST), for subnational maize yield forecasting across Nigeria. Two findings are especially salient for operational forecasting. First, MODIS LST added negligible forecast skill beyond ERA5-Land air temperature (ΔRMSE = 0.001 t ha⁻¹), an outcome explained by strong collinearity between the two variables at this scale (Pearson r = 0.89). Second, a simple persistence baseline based on prior-year state yield outperformed the full EO-only model, establishing a stringent benchmark for seasonal satellite and reanalysis predictors. The best EO-only configuration achieved RMSE = 0.475 t ha⁻¹ and MAPE = 29.4%, while adding soil properties and lagged yield reduced these to 0.407 t ha⁻¹ and 22.6%, respectively. Detrended analysis confirmed that EO variables capture genuine interannual anomaly signal beyond the secular trend. Forecast skill varied sharply across agroecological zones, with zone-level RMSE nearly perfectly predicted by yield variability. The EO-only model failed to beat a constant-mean forecast in three of nineteen evaluation years, each associated with distinct non-climatic disruptions. These findings clarify which EO data streams justify inclusion in operational forecasting systems and show that effective crop monitoring depends on selective predictor design rather than data-maximal accumulation.