A convolutional neural network for acoustic-based intensity rainfall detection in the Amazon rainforest

Accurate rainfall monitoring is vital for agricultural planning and natural disaster prevention. This study introduces a novel convolutional neural network (CNN) architecture for passive acoustic-based rainfall detection in the infrastructure- scarce Amazon rain forest. For the binary wet/dry classification task, both the proposed and baseline architectures achieved metric values of approximately 99.00%, with the proposed model reducing showing 36.14% fewer false alarms. Furthermore, on the multi-class rainfall-intensity task, the proposed CNN outperformed the baseline within the ambiguous moderate precipitation class, raising its recall rate from 48.90% to 60.70%.