The present application relates to the technical field of
vegetation aboveground biomass prediction, in particular to a
vegetation aboveground biomass prediction method based on multi-scale adaptive topology and cross-
modal iterative interaction, aiming at solving the problem that
remote sensing images and structured
sample plot data are only simply spliced and insufficiently fused, resulting in insufficient prediction accuracy and generalization ability, the present application first collects multispectral
remote sensing images and spatially registered
sample plot attribute data of a target area, completes
radiation correction, data enhancement and other preprocessing; adopts visual
Transformer to extract image global semantic features, constructs an adaptive topology
adjacency matrix for table data, obtains fine-grained features through a graph neural network, generates multi-scale graph representations through grouping and global aggregation, realizes double-
modal multi-round iterative deep interaction through cross-attention, fuses and extracts global features for splicing, inputs a multi-regression head for parallel prediction, and outputs the final result through
adaptive integration and prior knowledge calibration, thereby improving prediction accuracy and generalization.