Stage-wise ConvLSTM Sequential Transfer Learning for Hierarchical Crop Type Mapping in Senegal’s Smallholder Systems

Earth Observation data (EO) can support food-security decision making in Sub-Saharan Africa, yet operational crop-type mapping in dryland smallholder systems remains challenging under rainfed-season cloud cover, heterogeneous cropping calendars, and small, irregular fields. Model performance is further degraded by scarce and noisy labels, mixed-cropping and intercropping practices, and strong domain shift across agro-ecologies. A particularly consequential failure mode is confusion between cropped and fallow parcels, where vegetated fallows can mimic crop spectral–temporal signatures and bias cropland statistics and downstream indicators used for early warning, input targeting, and program planning. A hierarchical, stage-wise sequential transfer-learning framework built on Convolutional Recurrent Neural Networks (ConvRNNs/ConvLSTMs) to improve robustness in data-scarce smallholder landscapes is proposed. The approach learns reusable spatiotemporal representations in a coarse-to-fine curriculum and transfers them across tasks of increasing label granularity. Stage 1 produces a cropland mask by classifying cropland versus other land uses (explicitly including fallow), targeting the crop–fallow confusion that dominates errors in dryland settings. Stage 2 refines cropland into agronomic family groups (e.g., cereals, legumes, vegetables), preserving interpretable subclass structure that is often sufficient for operational monitoring when fine labels are sparse. Stage 3 resolves fine-grained crop types and mixed-dominant intercropping states. The ConvLSTM backbone is trained stage-wise: parameters learned at a coarser stage initialize the next stage, while stage-specific classification heads are optimized for the current hierarchy level. The framework is demonstrated in Senegal using Planet NICFI monthly composites (~5 m; RGB+NIR) and in situ polygon labels collected during the 2020 and 2023 rainfed seasons. Training samples are built as ~0.5 ha image patches (14×14 pixels) extracted from interior points within polygons, with sampling density scaled by polygon area to better represent large fields while maintaining coverage of small parcels. The dataset, 6,978 labeled polygons in 2020 and 5,827 in 2023 generate 18,380 and 18,378 patches for September and October 2020 (no August imagery), and 13,733/13,623/13,524 patches for August/September/October 2023. To address severe long-tail imbalance typical of regional crop inventories, offline quota-based corpus curation, online weighted sampling, and consolidate ultra-rare fine-grained labels into an “OTHER” class at Stage 3 to stabilize training, are combined. The staged framework is benchmarked against machine and deep learning baselines (Random Forest, XGBoost, CNN, and single-stage recurrent models) using macro-averaged metrics and precision–recall behavior, selecting operating points that favor higher precision for operational mapping. Results show robust cropland maps with stable accuracy under limited labels and small, irregular fields, while preserving subclass structure; cereals and legumes remain identifiable at Stage 2 (validation accuracy ≈ 0.59). At Stage 3, precision is highest for major crops, Groundnut 0.83, Millet 0.72, and Maize 0.69, and moderate for Cowpea 0.51 and Rice 0.42. Remaining errors are primarily driven by data imbalance, mixed-cropping systems, and spectral confusion, highlighting priority areas for improving long-tail supervision and intercropping representation.