CasFormer: A Cascaded Transformer Network with Self-Cross Attention for Forest Multi-View Reconstruction and Tree Parameter Extraction

Image-based forest 3D reconstruction from multi-view imagery can generate stand-level point clouds and provide data support for extracting individual-tree structural parameters, such as diameter at breast height (DBH) and tree height. Three-dimensional reconstruction of forest ecosystems is fundamental to forest resource monitoring and biomass assessment. However, forest scenes are characterized by frequent branch and leaf occlusions, repetitive bark and foliage textures, and complex illumination conditions. These factors often introduce ambiguities into feature matching and cost aggregation in convolutional neural network (CNN)-based multi-view stereo (MVS) methods, resulting in incomplete point clouds and depth estimation errors. To address these challenges, this study proposes CasFormer, a cascaded multi-view 3D reconstruction model designed for near-ground handheld forest imagery. Built upon a cascaded cost-volume framework, CasFormer constructs a Self-and-Cross Attention Feature Pyramid Network (SC-FPN) and incorporates a Positional Encoding Generator (PEG) to inject camera geometric priors into feature representations, thereby enhancing feature discriminability in repetitive weak-texture regions. In addition, a pixel-wise visibility-weighted cost volume is used to adaptively adjust source-view aggregation weights according to matching confidence, reducing the influence of occluded or low-quality source views on the cost volume and improving the stability of depth estimation in forest scenes. Experiments across three forest plots show that CasFormer achieves the highest overall single-tree matching rate (99.14%) and outperforms the competing methods in tree-height accuracy and point-cloud density at all three plots. Its mean DAE is 92.87%, exceeding FS-MVSNet and MVSFormer by 0.97 and 2.78 percentage points, respectively. CasFormer attains the lowest DBH RMSE at Plots 1 and 2; at Plot 3, its RMSE is comparable to those of MVSFormer and FS-MVSNet. Its point-cloud densities are 3.87, 2.17, and 4.87 × 106 points/m2 at Plots 1, 2, and 3, respectively, the highest among the methods evaluated at each plot. These results demonstrate that CasFormer improves plot-scale forest 3D reconstruction using handheld RGB imagery that can be acquired with minimal effort. It provides an effective image-based reconstruction method for extracting individual-tree parameters and supports sustainable smart forestry monitoring.