MCACD: A Multi-Scale Convolutional Attention Network for Forest and Grassland Change Detection

Forests and grasslands are crucial components of terrestrial ecosystems. Their stability and health play an essential role in maintaining ecological balance. However, with the intensification of human activities, these ecosystems are facing increasing risks of degradation. Forest and grassland change detection is vital for monitoring and management by governmental agencies, requiring reliable and automated detection models. Currently, most deep learning-based change detection models face significant limitations in achieving precise detection of forest and grassland changes. These challenges are particularly evident in addressing incomplete segmentation objects and accurately discriminating hard/challenging samples, highlighting substantial room for further research and improvement. To address these challenges, we have developed MCACD. This model effectively preserves detailed information by combining high-resolution feature maps with dynamic upsampling segmentation heads. It also integrates an efficient feature fusion module, designed by us, to enable the integration of dual-temporal features. Additionally, we introduce a highly efficient multi-scale cascade fully convolutional attention decoder, which progressively enhances spatial features at multiple scales and resolutions. A deep supervision strategy within the decoder further optimizes the feature learning process. Experimental results show that, compared to other advanced change detection methods, MCACD performs exceptionally well on our custom-built forest and grassland change detection dataset. It achieves the best F1 score and IoU values, demonstrating significant potential for practical applications in forest and grassland change detection.