Real-Time Feature Extraction and Topological Change Detection of Multi-Biome Forests on Resource-Constrained Systems

Feature extraction and detection of changed topological characteristics of forest structures are essential in combating illegal logging and providing ecosystem parameters for environmental scientists. However, to this date, most models based on UAV and satellite imagery, which have been developed in this regard, are not oriented towards real-time processing and inference. A key application of such lightweight models would be deploying them on a UAV to predict changes and extract features in real-time during flight. Therefore, this research paper aims to discover, develop and deploy semantic segmentation machine learning models which are lightweight enough to run on resource-constrained systems, which could be mounted on a real UAV to perform these tasks. Towards this end, the following paper proposes 3 lightweight CNN models trained on multi-biome forest datasets, capable of being deployed on two popular target microcontrollers. Moreover, a deployment pipeline for a drone capable of feature extraction and topological change detection is introduced as well.