Ensembled Deep Learning-Based Optimisation of Multi-Robot Communication Networks in Agricultural Environments

In this paper, a novel ensemble deep learning framework is presented for communication management in agricultural multi-robot systems. The framework uses Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), and Graph Neural Networks (GNN) to classify link quality, predict the future evolution of link quality, and inform routing decisions in wireless networks used in agricultural scenarios. The models were trained on the Vineyard-2021 LoRa RSSI dataset, comprising real wireless measurements gathered during deployment in vineyards in Northern Italy, preprocessed and annotated across seven sensor nodes and 1,877 hourly timesteps. To prevent target leakage, the prediction task was framed as a future-prediction task: models predict the link-quality class at time t+1 using only measurements available up to time t. The CNN reached a test accuracy of 83.69%, the LSTM 81.14%, and the GNN 85.82%. An unweighted ensemble of the three models reached a test accuracy of 89.32%. All classification errors occurred between adjacent classes, and no Weak-to-Strong or Strong-to-Weak errors were observed. This ensemble was then used to drive routing decisions in a multi-robot simulation built on the NetworkX platform, with fleet sizes ranging from 5 to 30 robots. Adaptive routing driven by the ensemble reached a Packet Delivery Ratio of 99.68%, compared with 95.33% for a distance-only baseline, while slightly increasing latency (18.03 ms vs. 12.80 ms) and energy consumption (1.81 vs. 1.36 units). The paper presents a documented baseline for deep-learning-based link-quality classification on agricultural LoRa data, a novel ensemble design combining spatial, temporal, and topological features, and an evaluation methodology for communication management in simulated multi-robot networks. Limitations include the static deployment of the sensors used for training, the geographic and seasonal specificity of the dataset, and the use of literature-derived packet-success probabilities in the simulation. Future work will gather data from moving robots in Nigerian agricultural environments and validate the framework using ROS/Gazebo.