Fuzzy Cognitive Map Optimisation via Genetic Algorithms for Non-Destructive Maturity Prediction of Hass Avocados Using Bioimpedance Spectroscopy

• FCM-GA achieves MAE = 0.053, outperforming five baselines (p < 0.01) • Leakage-free leave-one-fruit-out protocol ensures genuine forecasting • Soil N and humidity are dominant impedance drivers (S > 0.28) • Impedance plateau (|ΔZ100| < 0.02) identifies harvest in 92% of fruits • Framework costs ∼ USD 30 per BIS unit; no laboratory analysis required Accurately predicting Hass avocado maturity before harvest is essential for supply-chain efficiency and post-harvest quality, yet standard indices (dry-matter content, oil content, firmness) require destructive sampling. We propose a non-destructive, interpretable framework combining low-cost bioimpedance spectroscopy (BIS) with a Genetic Algorithm-optimised Fuzzy Cognitive Map (FCM–GA) operating in a true predictive mode. Impedance is used as a non-destructive proxy for maturity; validation against destructive reference indices (dry-matter content, oil content) remains essential future work. Longitudinal BIS measurements at 50 kHz ( Z 50 ) and 100 kHz ( Z 100 ) were collected from 150 Hass avocado fruits on 10 trees at Finca Nueva Esperanza, Facatativá, Colombia (September–December 2023) over four 15-day visits. Soil nitrogen (N), phosphorus (P), potassium (K), and ambient humidity (H) are treated as exogenous inputs (measured and clamped at each visit), while the FCM predicts the two impedance concepts. The weight matrix W is re-estimated within each leave-one-fruit-out fold from training fruits only (no leakage), and the GA uses only the first measurement visit to optimise A (0); prediction is evaluated on visits 2–4. In this rigorous forecasting setup, the FCM–GA achieves MAE = 0.053 ± 0.015 on held-out impedance visits, outperforming a classic expert-initialised FCM (0.114 ± 0.028; − 53.5 % ; p < 0.001), pseudoinverse FCM learning (0.071 ± 0.019; − 25.4 % ), and tabular baselines (Random Forest: 0.078; linear regression: 0.089; persistence: 0.102). Sensitivity analysis identifies N and H as dominant impedance drivers ( S ¯ > 0.28 ), K as the principal inhibitor ( S ¯ = − 0.24 ), and P as agronomically negligible ( S ¯ < 0.05 ). Convergence diagnostics confirm that the GA reaches fitness ϕ ≥ 0.90 within a mean of 2, 140 ± 680 generations at ∼ 90 s per fruit on commodity hardware. The framework costs ≈ USD 30 per BIS unit and requires no laboratory analysis of the sampled fruit. The harvest-timing indicator is presented as a preliminary observation; validation against independent destructive maturity measurements is required to confirm the impedance plateau as a reliable harvest signal.