Enhancing Cropland Mapping in Nigeria via Ensemble Deep Learning and Uncertainty Quantification to Support SDG 2 in Data-Limited Regions

Accurate cropland mapping is essential for food security monitoring and Sustainable Development Goal 2 tracking, yet data scarcity in developing regions limits machine learning applications. This study compares three deep learning architectures for cropland mapping in Nigeria using 1,822 georeferenced samples with 12-month Sentinel-1, Sentinel-2, climate, and topographic features: Long Short-Term Memory networks, Temporal Convolutional Networks, and Vision Transformers. We developed an ensemble framework integrating uncertainty quantification through inter-model disagreement and Monte Carlo Dropout. TCN achieved the highest individual F1-score of 0.808, followed by LSTM at 0.799 and Vision Transformer at 0.779, revealing limitations of attention-based architectures on small datasets despite having 402,050 parameters versus 216,258 and 198,210 for LSTM and TCN. The LSTM-TCN ensemble achieved F1-score of 0.819, improving 1.30% over the best individual model. The ensemble detected 13.6% more cropland than ESA WorldCover 2020 (recall 0.863 vs 0.760), critical for food security monitoring. Uncertainty quantification identified 3.74% of predictions requiring verification, with accuracy improving to 0.854 on certain predictions. Monte Carlo analysis revealed aleatoric uncertainty of 0.378 for LSTM and 0.260 for TCN, with modest inter-intra model correlation (ρ=0.445). Performance varied across agro-ecological zones, with Southern Guinea Savanna achieving F1-score of 0.903 versus 0.667 in Rainforest zones. This work demonstrates that ensemble methods combining architecturally diverse models with uncertainty quantification provide transparent, operationally deployable solutions for agricultural monitoring in data-limited regions, supporting SDG 2 achievement. • Systematic comparison of LSTM, TCN, and Vision Transformer for cropland mapping on small Nigerian dataset (1,822 samples). • LSTM-TCN ensemble achieves highest F1-score of 0.819 using probability averaging and uncertainty quantification. • Dual uncertainty framework (inter-model disagreement + Monte Carlo Dropout) flags 3.74% of predictions as high-uncertainty (U > 0.5), with accuracy of 0.647 vs 0.854 for certain predictions. • Ensemble improves cropland recall by 13.6% over ESA WorldCover 2020, enhancing food security monitoring. • Demonstrates efficient, transparent deep learning solutions for agricultural mapping in data-limited regions supporting SDG 2.