Fuel detection in forest environments training deep learners with smartphone imagery

Abstract. Unmixing mixtures in images is one of the challenges for extracting information from data. Forest environments are particularly complex due to the relatively irregular structure of trees, shrubs and low vegetation. The amount and condition of vegetation, i.e. thin vs thick branches, trunk vs leaves, understorey and litter provide information to infer the amount of burnable fuel and consequently a key factor for predict fire behaviour. In this work we test a deep learning framework for training and testing the performance of detecting logs and litter of broadleaves and conifers in imagery of forest environments recorded through smartphones. Roboflow and YOLOv8 were employed, using a dataset of forest images manually segmented in four classes: “broadleaf-litter”, “broadleaf-logs”, “conifer-litter” and “conifer-logs”. The results indicate that the "Extra-large Instance Segmentation" model achieved the best performance with F1-score value of 0.79 at a confidence of 0.763 on familiar images in the validation phase with 214 epochs, whereas the "Large Instance Segmentation" model was more effective on new images in the test phase, as expected with a lower F1-score of 0.24 and a confidence value of 0.492. It was observed that this was due mostly to omission errors due to low light conditions in the forestry environment. We conclude that segmenting key elements and including varied images in terms of seasonality and lighting conditions could potentially improve performance. This work lays a useful foundation for refining the use of AI in forest fuel monitoring.