Landscape Visual Perception Prediction and Spatial Optimization Using XGBoost Ensemble Learning

Abstract

To overcome the difficulty of accurately analyzing complex nonlinear coupling mechanisms in Landscape Visual Perception (LVP) assessment and the lack of scientific decision-making support at the spatial optimization level, this research constructs a set of technical methods that integrate high-precision prediction and intelligent diagnosis to optimize space. This research proposes an integrated learning model based on eXtreme gradient boosting that deeply integrates three-dimensional spatial morphology and visual field perception indicators, and introduces 2nd-order Taylor expansion and regularization techniques to optimize the solution of the loss function. Furthermore, it uses the Shapley additive exPlanations algorithm to build a complete logical chain from feature attribution to spatial intervention. The results revealed that this method achieved a test-set coefficient of determination of 0.812 and a median absolute error of 0.25 in a heterogeneous multi-source data environment, significantly outperforming deep neural networks and traditional regression methods. In the spatial optimization simulation experiment, differentiated and precise intervention implemented based on the system attribution results increased the perceptual scores of typical low-quality block samples by an average of 37.2%, with a maximum improvement of 124.2%, verifying its decision-making efficiency under limited resources. This research forms a closed-loop path of "problem diagnosis-attribution interpretation-precise intervention", providing a scientific basis for urban stock renewal and refined management of human settlements.