Low Latency Forest Logging Monitoring Using Differ-Modality Learning Approach

As monitoring of logging detection and tree retention through manual means is often resource-expensive, satellite-borne synthetic aperture radar (SAR) and optical sensor modalities have been utilized extensively. Furthermore for industrial forestry management and operations, low latency logging monitoring commensurate with logging schedules is required. However, satellite remote sensing data maybe unavailable at the desired frequency, e.g., optical modality maybe hampered by clouds, or stronger post-flood ground reflections in SAR modality may hamper forest area segmentation. Hence, to perform forest segmentation and logging detection, we utilize "differ-modality learning" (DML) to overcome limitations associated with either modality and their availability at a given time-epoch. DML allows training of the model on a particular modality (e.g., optical) and still make predictions using the other modality (e.g., SAR). Consequently, the model allows us to predict the unlabeled SAR/optical pairs to generate the pseudo labels by a majority vote. We evaluated the model on the SpaceNet-6 challenge dataset and compared its performance to the recently released Meta’s Segment Anything Model (SAM). IoU higher than 0.7 was achieved by training on SAR modality and testing on optical modality as well as training on optical modality and testing on SAR modality. The originality of this work is to facilitate utilization of DML approach for operational forestry monitoring.