Improving Rare Tree Species Classification Using Domain Knowledge

Forest inventory forms the foundation of forest management. Remote sensing (RS) is an efficient means measuring parameters at scale. Remotely sensed species classification can be used to estimate abundances, distributions, and better approximate metrics such as above ground biomass. State art methods RS rely on deep learning models convolutional neural networks (CNN). These have 2 major drawbacks: they require large samples each classify well lack explainablity. Therefore, rare are poorly classified causing poor approximations their associated parameters. We show that improved by much 8 F1-points using a neuro-symbolic (NS) approach combines CNNs with NS framework. The framework allows for incorporation domain knowledge into model through use mathematically represented rules, improving explainability.