Exploration of a Neural Network Transfer Learning Approach to Estimate Aboveground Biomass with Insufficient Temporal Data

Forests provide an effective natural climate solution by absorbing and storing atmospheric C02 and thus combat climate warming. Tropical forests, i.e., the Amazon rainforest, host the majority of forest biomass. However, the estimate of aboveground biomass stored in the Amazon rainforest is highly uncertain. This paper employs a transfer learning approach, initially pretraining the neural network on Congo rainforest data and then applying it to estimate Amazon rainforest biomass. By using advanced Artificial Neural Networks (ANN s) and transfer learning to extrapolate aboveground biomass estimates in the Amazon rainforest, this work mitigates the temporal data insufficiency experienced by remote sensing techniques, particularly NASA's Global Ecosystem Dynamics Investigation (GEDI). The model achieves an acceptable R2 score of 0.36 and estimates that the total biomass within the Amazon rainforest ranges between 118.8 to 157.0 Pg. This paper addresses the critical role of rainforests as global carbon sinks and the necessity of accurately monitoring their biomass for effective climate management. By effectively extrapolating in under-sampled regions, the proposed model demonstrates the potential to enhance accuracy in biomass estimations, thus informing better rainforest management strategies. These findings underscore the urgency of addressing the climate crisis and emphasize the pivotal role of accurate biomass estimation.