Part of Advances in Neural Information Processing Systems 36 (NeurIPS 2023) Datasets and Benchmarks Track

Part of Advances in Neural Information Processing Systems 36 (NeurIPS 2023) Datasets and Benchmarks Track

Authors: Andrea Nascetti, Ritu Yadav, Kirill Brodt, Qixun Qu, Hongwei Fan, Yuri Shendryk, Isha Shah, Christine Chung

Abstract:
Above-ground biomass plays a critical role in mitigating climate change, as forests serve as efficient, natural carbon sinks. Traditional field-based and airborne LiDAR measurements have proven reliable for estimating forest biomass, but scaling these techniques across large areas is often both costly and challenging. Satellite data, especially from sources like Sentinel-1 SAR and Sentinel-2 MSI, have emerged as valuable tools for global biomass estimation, yet the full potential of dense, multi-modal satellite time-series data combined with modern deep learning approaches remains largely untapped.

The "BioMassters" data challenge and benchmark dataset aim to explore the use of these multi-modal satellite datasets to estimate forest biomass at a large scale, with reference data from the Finnish Forest Centre’s open forest and nature airborne LiDAR data. Our results show that deep learning models can produce accurate, high-resolution biomass maps, as demonstrated by the performance of the top three baseline models.