Reliable wireless connectivity in precision agriculture requires accurate propagation models that remain robust across varying orchard corridor geometries within a given vegetation canopy type. This study characterizes radio propagation at 18 GHz (FR3 band) through an extensive measurement campaign (N = 17,269 samples) in a custard apple orchard, systematically varying corridor width \(W \in [1.3,\,3.8]\) m and transmitter height \(h_{tx} \in [1.0,\,2.03]\) m across nine geometric configurations. The standard Close-In (CI) model captures the distance-dependent path loss (global fit \(n = 2.51\) ) but leaves substantial geometry-dependent variability unexplained (RMSE = 3.91 dB). When the CI model is fitted per configuration, the parameters vary markedly ( \(n \in [1.93,\,2.68]\) and \(\sigma \in [2.25,\,5.88]\) dB), indicating that a single averaged exponent should not be interpreted as geometry-invariant. To address this limitation, we propose a Hybrid Linear+XGBoost framework that combines physically interpretable geometric corrections with nonlinear residual learning. Under leave-one-scenario-out cross-validation, the hybrid model achieves RMSE = 2.97 dB, a 24.0% improvement over the CI baseline. Crucially, while pure ensemble methods (Random Forest, Gradient Boosting, XGBoost) exhibit performance degradation up to 1.05 dB when extrapolating to unseen geometric configurations, the hybrid architecture demonstrates a 0.35 dB improvement, achieving better accuracy on novel corridor configurations than on interpolated data. Corridor width emerged as the dominant factor influencing shadow fading (SF) within the studied custard apple canopy. These findings establish that geometric parameters must be explicitly incorporated into channel models and that hybrid physics-ML architectures offer robust methodological generalization across corridor geometries, whereas the fitted numerical coefficients are expected to depend on vegetation descriptors such as leaf area index, leaf orientation, and moisture content.
