Abstract: Understanding the spatial distribution of canopy gaps holds significant importance for the conservation and sustenance of forest ecosystems.In the task of extracting canopy gap information based on GF-2 remote sensing imagery,given the extensive and complex distribution of cannopy gaps within forest systems,traditional remote sensing interpretation methods are inefficient and prone to misclassification and omission.Therefore,an improved canopy gap information extraction model based on PSPNet is proposed.This model replaces the backbone network to lighten its load,incorporates the CBAM attention mechanism,and refines the loss function,thereby enhancing the model's ability to learn forest gap information and addressing the issue of inaccurate identification of canopy gap edge details caused by an imbalance between positive and negative samples.Compared to the original PSPNet model,the improved PSPNet model exhibits an increase in the average intersection over union (IoU) by 3.12 percentage point,an improvement in average pixel accuracy by 3.6 percentage point,and a 65.43% increase in detection speed.This demonstrates the effectiveness of the proposed method for canopy gap information recognition.

Key words: object-oriented classification, membership function, deep learning, semantic segmentation, canopy gap