Coniferous tree species classification based on DMCA-Unet network model with UAV multispectral imagery

• Replacing ordinary convolution with the combination of depthwise convolution and dilated convolution with consecutive dilation rates for convolution operations, so as to improve feature extraction capability, expand the receptive field, enhance model efficiency, and avoid the grid effect caused by dilated convolution. • Adopting a multi-head cross-attention mechanism to handle information interaction between different inputs in the encoder-decoder architecture. • Substituting the standard ReLU activation function with the LeakyReLU function, which effectively overcomes the common neuron inactivation problem of ReLU during training and further improves the model's fitting performance. • Employing the Adam optimizer to adaptively adjust the learning rate for each parameter. • Using a hybrid loss function that fuses the cross-entropy loss function and the Dice loss function. This hybrid loss function can focus on the classification accuracy of each pixel while emphasizing the accuracy of region overlap and shape, thereby further enhancing the model's segmentation accuracy As a critical component of global forest ecosystems, the classification of coniferous tree species facilitates precise assessment of forest resources and enhances forest As critical components of global forest ecosystems, accurate classification of coniferous tree species facilitates precise assessment of forest resources and enhances forest management efficiency, with substantial implications for ecological conservation and carbon sequestration evaluation. However, existing studies predominantly rely on high-cost, processing-intensive hyperspectral data, restricting large-scale practical applications. To address conventional methods' limitations in characterizing complex coniferous species and overcome hyperspectral dependency, this study proposes DMCA-Unet—a semantic segmentation model utilizing cost-effective unmanned aerial vehicle (UAV) Multispectral imagery. Our approach exhaustively exploits spatial-semantic- spectral features of UAV images for tree species classification in heterogeneous forests. Employing multispectral data from the Millennium Xiulin Experimental Zone (Xiongan New Area), the model adopts a VGG-16-based encoder backbone. The framework integrates deep dilated convolution modules that combine depth-wise convolutions with progressive dilation rates and LeakyReLU activations, substituting standard convolution operations to elevate training efficiency, accuracy, and convergence. Furthermore, a multi-head cross-attention mechanism captures interdependencies among heterogeneous features, strengthening classification capability in complex backgrounds. For optimized parameter learning, a combined loss function integrating cross-entropy and Dice loss resolves class imbalance issues during training. Experiments demonstrate: (1) DMCA-Unet achieves 94.45% overall accuracy, surpassing DeepLabV3+,ViT(Vision Transformer),HRNet, PSPNet, Unet-VGG16, Unet-ResNet50, and U-Net by +4.6%, +27.78%, +2.7%, +11.85%, +2.31%, +2.42%, and +3.54% respectively at lower computational cost; (2) The hybrid loss markedly enhances accuracy under sample imbalance while stabilizing training. This methodology provides technical support for forest resource monitoring and sustainable management.