ClimaCropNet: A Two-Stage Hybrid Deep Learning and Stacked-Ensemble Framework for Crop Yield and Market-Price Forecasting under Climate Variability in a Data-Scarce Tropical District

Climate variability threatens agricultural productivity and food security across Sub-Saharan Africa, where rain-fed smallholder systems dominate. Early warning of crop conditions, yields, and prices could support planning, yet it is unclear whether deep-learning models provide this information better than simple methods in data-scarce districts. This study developed and rigorously evaluated ClimaCropNet, a two-stage hybrid framework that integrates multi-source satellite remote sensing and climate reanalysis data to forecast crop yields and agricultural market prices in the Edo-North Senatorial District, Nigeria. Twenty-five years (2001 to 2025) of Moderate Resolution Imaging Spectroradiometer vegetation and land-surface-temperature products and ERA5-Land climate variables were acquired through Google Earth Engine and combined with official annual yields of four staple crops for Nigeria (cassava, yam, rice and maize; 21 years each) and the inflation-adjusted wholesale maize price at Dawanau market, Kano, the national reference price that southern markets closely follow (19 years). Stage 1 applies a Convolutional Neural Network, Bidirectional Long Short-Term Memory, and attention network, stacked with tree ensembles and simple reference forecasts, to forecast the normalized difference vegetation index one to three months ahead from monthly climate and vegetation sequences; Stage 2 supplies these forecasts, with early-season climate, to a stacked ensemble of reference forecasts, ridge regression, Random Forest, Extreme Gradient Boosting, and Gradient Boosting regressors for annual prediction. ClimaCropNet was benchmarked against naive, statistical, machine-learning and deep-learning models using expanding-window validation, Diebold-Mariano tests and Wilcoxon signed-rank tests. Over 210 test months per lead time, ClimaCropNet was the most accurate learned model at every lead time, significantly more accurate than nine of fifteen benchmarks at one month, and reduced the mean squared error of the monthly climatology by 11 percent at one month and 2 percent at two months, with 90 percent prediction intervals covering 86 to 89 percent of observations. Its forecast-level stacking reduced the error of a feature-level stacking design by 15 to 20 per cent. For annual national yields, ClimaCropNet was significantly more accurate than eight of ten benchmarks and statistically indistinguishable from persistence. It also forecasts the real maize price with 20 per cent lower means squared error than persistence, although nine test years are too few for this difference to be significant. Leave-one-out cross-validation with a year trend raised the apparent coefficient of determination for the nominal maize price from 0.30 to 0.81, showing how non-temporal validation inflates skill. ClimaCropNet thus provides a transparent, leakage-free and computationally light forecasting framework built entirely on public data for data-scarce agricultural districts.