Adaptive Local Aggregation for Personalized Federated Learning in Agricultural Terrain Mapping

Statistical heterogeneity and non-independent and identically distributed (non-IID) data remain major challenges in federated learning, particularly for agricultural terrain mapping where geographically distributed sensing devices collect highly diverse remote sensing data. To address these challenges, this paper proposes an Adaptive Local Aggregation (ALA)-based personalized federated learning framework that improves client-specific model initialization while preserving the standard federated learning communication and global aggregation protocol. Rather than directly replacing local models with the received global model, the proposed framework adaptively initializes selected higher-layer parameters using learnable client-specific aggregation weights, thereby balancing globally shared knowledge with local personalization prior to client-side training. The proposed framework is evaluated on four agricultural remote sensing datasets, namely AgriFieldNet, BigEarthNet v1, BigEarthNet v2, and CropHarvest, together with three benchmark image classification datasets, MNIST, CIFAR-10, and CIFAR-100, under heterogeneous federated learning settings. Experimental results demonstrate consistent improvements in global classification accuracy and convergence behaviour compared with state-of-the-art federated learning methods, highlighting the effectiveness of the proposed framework for privacy-preserving agricultural terrain mapping under heterogeneous client data distributions.