Abstract: Forest tree breeding is characterized by long cycles, high heterozygosity, and significant genotype-by-environment interactions. Traditional genomic selection (GS) models face limitations when dealing with high-dimensional small-sample data and complex non-additive effects. Deep learning, with its nonlinear modeling capability based on multi-layer neural networks, offers new approaches for dissecting the complex relationships between genotype and phenotype. This technology has been widely applied in crop GS, and various models incorporating attention mechanisms and lightweight designs have been developed. Meanwhile, joint modeling of multi-omics and environmental data, as well as the construction of selection indices for multi-trait synergistic optimization, are emerging as research hotspots. However, for forest tree breeding—characterized by long-term environmental interactions, highly heterozygous genomic backgrounds, and multi-objective selection demands—the adaptability of deep learning techniques and their integrative frameworks still lack a systematic review. This paper reviews the applications of deep learning in forest tree GS. The main contents include: (1) the evolution of genomic input features from single nucleotide polymorphisms to k-mers and graphical pan-genome node types, along with associated computational challenges; (2) modeling methods for multi-omics and environmental data; (3) structural characteristics, applicable scenarios, and hyperparameter tuning strategies of mainstream deep learning models; (4) the theoretical progression from single-trait genomic estimated breeding value prediction to multi-trait selection indices. This review also analyzes current major challenges, including data sparsity, model overfitting, and insufficient interpretability, and identifies transfer learning, semi-supervised learning, and mechanism-guided modeling incorporating biological priors as potential solutions. Finally, we envision the further integration of high-throughput phenotyping with deep learning and propose the construction of a smart breeding platform that integrates multi-omics data management, automated analysis pipelines, and breeding decision support, thereby facilitating the transformation of forest tree breeding toward genome-wide intelligent design.