Optimisation of landscape feature intelligent recognition algorithm based on deep learning

This paper presents a lightweight dynamic feature pyramid network (DFPN) for intelligent landscape feature recognition, tackling computational redundancy and boundary ambiguity in complex scenes.DFPN employs an adaptive scale-aware module to dynamically weight and fuse ResNet-18's multilevel features (C1-C4), while a squeeze-and-excitation (SE) channel attention mechanism amplifies critical feature responses.Deformable convolutions enable adaptive receptive fields, enhancing multi-scale landscape perception and computational efficiency.For sharper boundaries, DFPN integrates atrous spatial pyramid pooling (ASPP) with edge priors and adopts a boundary-aware dice loss to mitigate edge ambiguity.On urban, complex, and rural landscape datasets, DFPN attains 93.82% pixel accuracy and 79.1% IoU at 42.3 FPS.In cross-dataset evaluation, its mIoU drops only 12.3% on the City dataset -substantially better than U-Net's 18.7% -confirming superior generalisation.Through co-designed architecture and loss optimisation, DFPN strikes an effective balance between accuracy and efficiency, proving robust in diverse, complex landscapes.