KombinatorNet: A Multisensor Deep Learning Model

KombinatorNet, created by the Earth Data Hub team at Sylvera, is a semantic segmentation Deep Learning model characterised by a multi-encoder, single-decoder architecture designed to support multi-sensor inference. Its multisensor design is specifically crafted to overcome the constraints associated with relying on a single input satellite source. This approach allows for temporally consistent and reliable time-series mapping regardless of sensor availability and data artefacts. Preliminary assessments of forest and non-forest pixel classification, including analysis over time series, demonstrate notable advancements in results when compared to outcomes obtained using a single sensor modality. The primary advantage of KombinatorNet lies in its ability to generate harmonious predictions even without consistent observations.