Non-Destructive Determination of Hass Avocado Harvest Maturity in Colombia Based on Low-Cost Bioimpedance Spectroscopy and Machine Learning

Abstract

The export of Hass avocado (Persea americana Mill.) from Colombia requires accurate determination of harvest maturity, currently assessed through destructive dry matter (DM) measurements that are wasteful and limited in throughput. The objective of the article is to propose a low-cost, non-destructive approach to determine the maturity of the Hass avocado crop based on machine learning techniques. The approach consists of a low-cost, non-invasive bioimpedance spectroscopy system operating in the 1–10 kHz range, featuring a custom Analog Front End (AFE) and a tetrapolar surface probe to mitigate skin contact resistance, which collects data for predictive models of avocado maturity. To evaluate the quality of the approach, a longitudinal field study (n = 100) was conducted in a commercial orchard in Cundinamarca, Colombia, tracking complex impedance features—Magnitude, Phase Angle, Resistance, and Reactance—of tagged fruits over 8 weeks across four measurement timepoints. The predictive performance of a classical chemometric model (PLS-DA), non-linear classifiers (SVM, Random Forest), and a temporal Deep Learning (LSTM) architecture was compared using a Stratified Group K-Fold Cross-Validation scheme to prevent data leakage across fruits from the same tree. The 4-electrode configuration successfully isolated mesocarp impedance, identifying the 5–7.2 kHz band as the most sensitive to physiological maturation. In turn, the LSTM model achieved a mean accuracy of 92.0% and an AUC of 0.94, outperforming the other models by 4.0% in mean accuracy. The results demonstrate that modeling the temporal trajectory of impedance, rather than single-point measurements, improves harvest maturity classification in Hass avocados, providing a scalable, low-cost alternative to destructive testing.

1. Introduction

Colombia has established itself as a leading global exporter of Hass avocados (Persea americana Mill.), with production concentrated in the Andean departments of Antioquia, Caldas, and Cundinamarca [1]. To meet the quality requirements of the European and US markets, fruit must be harvested at a defined physiological maturity level, typically indicated by a Dry Matter (DM) content exceeding 23%, which ensures proper post-harvest ripening and flavor development [2]. Harvesting too early produces fruit with a rubbery texture and bitter taste, whereas late harvesting shortens shelf life and increases susceptibility to fungal pathogens such as Colletotrichum spp. [1,3].

The supply chain from Andean orchards to destination markets in Europe or Asia involves transit times of 20–35 days under cold storage. Currently, producers rely on destructive oven-drying analysis following AOAC protocols [4] to determine DM content. This process is slow, is labor-intensive, and precludes high-frequency, individual-fruit monitoring. Consequently, there is a strong demand for non-destructive technologies capable of assessing maturity directly on the tree [5].

1.1. Related Work and Research Gaps

Several non-destructive sensing modalities have been investigated for fruit maturity assessment. Near-Infrared Spectroscopy (NIR) has been widely applied for DM estimation in avocados with accuracy exceeding 95%; however, instrumentation costs remain prohibitive for small-scale farmers [6]. Acoustic resonance methods offer lower hardware costs but are sensitive to fruit geometry rather than internal chemistry [7]. Hyperspectral and multispectral imaging has been used to map surface and subsurface properties of horticultural produce, though they require controlled illumination environments. Wi-Fi Channel State Information (CSI) has been explored for non-contact ripeness estimation, achieving approximately 91% accuracy in avocados [8], but relies on wireless infrastructure unavailable in remote orchards.

Electrical Bioimpedance Spectroscopy (EIS) offers a complementary approach by measuring the frequency-dependent opposition of biological tissue to alternating current [9]. As the avocado matures, the enzymatic activity degrades the cell membranes and modifies electrolyte concentrations, producing measurable shifts in the impedance spectrum [10,11]. Prior studies have demonstrated the feasibility of EIS in tropical fruits, including mangoes and bananas [12]; however, most rely on laboratory LCR meters and two-electrode configurations that introduce high contact impedance artifacts, limiting their field applicability. The four-electrode (tetrapolar) configuration, which separates current injection from voltage sensing, substantially reduces electrode polarization errors as described by Schwan [13] and has been implemented in custom ASIC designs [14], yet its application to avocados in field conditions remains largely unexplored.

In the domain of predictive modeling for fruit quality, machine learning methods have been applied across a range of crops. Mancero-Castillo et al. [15] surveyed modeling approaches for tropical fruit production, noting the scarcity of reported models for avocado and related species. Xiao et al. [16] reviewed the integration of multi-omics data with classical and deep learning methods for food quality classification. Specific to avocados, Support Vector Machines and ensemble methods have been applied to texture and spectral features with promising results, while recurrent architectures such as Long Short-Term Memory (LSTM) networks have shown strong performance on sequential agricultural time-series data [17,18].

Table 1. Summary of Representative Studies on Non-Destructive Approaches to Assessing the Maturity of Avocado and other Tropical Fruits (EIS: Electrical Impedance Spectroscopy; NIR: Near-Infrared; CSI: Channel State Information).

Reference	Method	Fruit	Temporal	Key Limitation
[6]	NIR	Avocado	No	High equipment cost
[7]	Acoustic	Multiple	No	Geometry-sensitive
[8]	Wi-Fi CSI	Avocado	No	Infrastructure required
[12]	EIS (2-electrode)	Mango/Banana	No	Contact artifact
[15]	ML (survey)	Tropical	No	No EIS integration
This work	EIS (4-electrode) + LSTM	Avocado	Yes

summarizes representative non-destructive maturity assessment studies, highlighting the research gap addressed by the present work. The reviewed literature reveals three key limitations, which are considered in this work. First, existing EIS-based fruit studies rely on two-electrode laboratory setups that confound contact and tissue impedance, preventing direct deployment in the field. Second, no prior work has collected longitudinal, multi-timepoint impedance measurements of individual avocados to characterize the dynamic maturation trajectory. Third, the comparative evaluation of temporal deep learning architectures against classical chemometric baselines has not been performed for avocado maturity classification under field conditions in the Colombian Andes.

1.2. Objectives and Contributions

The objective of this study is to develop and validate a low-cost, non-destructive approach for assessing the maturity of Hass avocado fruit using machine learning techniques. Specifically, the main contributions of this work are the following:

• Design and implementation of a low-cost, non-invasive portable bioimpedance spectrometer operating in the 1–10 kHz frequency range.
• Development of a custom Analog Front End (AFE) integrated with a tetrapolar surface probe to minimize skin–electrode contact impedance artifacts.
• Empirical demonstration of the superiority of the 4-electrode configuration over the 2-electrode configuration, in accordance with Schwan’s electrode polarization theory [13].
• Performing a correlation analysis between the spectral characteristics of bioimpedance and destructive dry matter measurements for the specific agroclimatic conditions of the Colombian Andes.
• Exhaustive evaluation of different machine learning algorithms for predicting avocado maturity based on bioimpedance-derived features.
• Conducting a field study that tracks 100 labeled fruits at four fortnightly measurement points (different maturity stages) to compare predictions based on time trajectories with those based on single-point measurements.

2. Materials and Methods

This section describes the biological criteria for harvest maturity classification, the proposed instrumentation system, the modeling strategy, the data collection process, and the experimental protocol.

2.1. Maturity Windows and Export Challenges for Colombian Hass Avocado

The maturation physiology of Hass avocado (Persea americana Mill.) presents a unique challenge for the agricultural industry: it is a climacteric fruit that does not soften while attached to the tree due to the inhibitory effect of endogenous hormones. Consequently, two distinct phenological stages must be differentiated:

• Harvest Maturity (Physiological Maturity): The stage where the fruit has accumulated sufficient dry matter (oil) to guarantee proper ripening post-harvest, yet remains firm enough to withstand handling and transport.
• Consumption Maturity (Edible Ripeness): The stage reached after the post-harvest climacteric rise in ethylene production, characterized by tissue softening and flavor development [19].

For the Colombian export sector, identifying the exact onset of Harvest Maturity is critical. The supply chain from Andean orchards to destination markets in Europe or Asia involves transit times ranging from 20 to 35 days. Therefore, the primary goal is to preserve the fruit’s “Green Life”—the period during which the fruit remains firm and unripe under cold storage.

If harvested too early (before reaching $\approx 23\%$ Dry Matter), the fruit fails to ripen properly, developing a rubbery texture and bitter taste. Conversely, late harvesting reduces Green Life, increasing the risk of the fruit ripening inside the shipping container or succumbing to chilling injury and fungal pathogens (e.g., Colletotrichum spp.) [2,20]. The Harvest Window is defined as the critical decision-making period beginning approximately 60 days prior to consumption ripeness.

2.2. Bioimpedance Measurement System and Analog Front End

A portable spectrometer based on the AD5933 impedance converter (Analog Devices, Wilmington, MA, USA) was designed and built. The AD5933 generates a sinusoidal excitation voltage sweeping from 1 to 10 kHz and digitizes the response signal via an internal 12-bit ADC, computing the real and imaginary components of impedance through a DFT-based algorithm on-chip. Frequency resolution, output voltage amplitude, and gain setting resistor were configured via I²C from an Arduino microcontroller. Measured impedance data were logged to an SD card and transferred via USB for post-processing.

To address the high and variable contact resistance of the exocarp without compromising fruit integrity, a custom Non-Destructive Tetrapolar (4-Electrode) Surface Probe was developed (see Figure 1). The contact interface consisted of four spring-loaded, gold-plated flat electrodes (2 mm diameter) arranged collinearly with a uniform inter-electrode spacing of 10 mm. Since the AD5933 is natively a 2-terminal device, a custom external Analog Front End (AFE) was implemented to support this tetrapolar configuration, adapting designs from recent ASIC-based front-ends [14].

The AFE is defined by: (i) a Howland Current Pump based on an LM324 operational amplifier to inject a stable, frequency-invariant alternating current (10 μA peak) through the outer electrode pair; and (ii) a precision Instrumentation Amplifier (INA128, Texas Instruments, Dallas, TX, USA) with high input impedance (>10 GΩ) to measure the differential voltage across the inner electrode pair without drawing appreciable current. This separation of excitation and sensing paths effectively removes the electrode–tissue contact impedance from the measurement, isolating the true mesocarp impedance.

Particularly, in Figure 1, the outer electrodes (E1, E4) inject a controlled AC current generated by the Howland Current Pump. The inner electrodes (E2, E3) sense the resulting differential voltage via a high-impedance Instrumentation Amplifier (INA128). The buffered voltage is fed back to the AD5933 VOUT/VIN terminals, enabling impedance calculation while excluding contact resistance artifacts.

This portable bioimpedance spectrometer, based on Arduino technology with a custom external Analog Front End (AFE), is an architecture with a substantially lower cost than commercial LCR meters or NIR instruments.

2.3. Statistical and Machine Learning Modeling Strategy

To map the bioimpedance spectral features to the binary maturity classes, a comparative framework was designed that evaluates three levels of algorithmic complexity.

2.3.1. Baseline Chemometrics: PLS-DA

As a baseline, Partial Least Squares Discriminant Analysis (PLS-DA) [21] was implemented. This technique efficiently handles multicollinearity in spectral data by projecting the high-dimensional impedance data into a lower-dimensional space of Latent Variables (LVs).

2.3.2. Non-Linear Static Classifiers: SVM and Random Forest

Given the complex electrochemical nature of biological tissue, the following classifiers were deployed:

• Support Vector Machine (SVM): Utilized with a Radial Basis Function (RBF) kernel [22,23]. SVM was selected because it maps features into a higher-dimensional space where non-linear class boundaries in the impedance feature space become linearly separable, a property established in prior spectral classification studies. Thus, SVM was selected for its proven effectiveness with high-dimensional spectral data and non-linear boundaries.
• Random Forest (RF): An ensemble of decision trees [24,25]. RF was included for its robustness to overfitting at moderate dataset sizes and its ability to provide interpretable feature importance rankings via Gini impurity, useful for identifying the most discriminative frequency bands.

The term “static” or“single-point” refers to the fact that these techniques receive only one measurement point per fruit during their learning process, without access to previous points, in contrast to the LSTM technique, which receives a trajectory of points.

2.3.3. Temporal Deep Learning: Long Short-Term Memory (LSTM)

Long Short-Term Memory (LSTM) [17] was selected as the representative sequential architecture because it retains long-range dependencies in time-series data through gating mechanisms (input, forget, output gates), making it suitable for learning from the multi-timepoint impedance trajectory of individual fruits. Thus, LSTM is a representative temporal architecture capable of learning from sequential impedance measurements [17]. Unlike static models, the LSTM receives a sequence of length $T=4$ (one measurement per bi-weekly interval), enabling it to encode the rate of change and curvature of the impedance decay. The architecture consists of a masking layer, a primary LSTM layer with 64 units, a secondary LSTM layer with 32 units, a dropout layer (rate = 0.2), and a sigmoid output neuron for binary classification [18].

2.4. Experimental Context and Data Collection

The study was conducted during the 2024 main harvest season in a commercial orchard located in Cundinamarca, Colombia. A longitudinal, non-destructive tracking design was employed, where the physiological evolution of the same tagged fruits was monitored over the maturation period. Ten distinct trees randomly distributed across the orchard were selected to account for spatial variability, and 10 fruits were randomly selected and tagged from each tree ($n=100$ total individuals). Bioimpedance measurements were acquired at 15-day intervals (bi-weekly) over the monitoring period (4 timepoints). All measurements from a given fruit were always assigned to the same cross-validation fold, preventing any temporal leakage between training and test sets (see Section 2.5.3).

2.4.1. Microclimate Monitoring and Stratified Sampling

Given the dense canopy structure of mature Hass avocado trees, significant microclimatic variations exist between the inner and outer foliage. To characterize these environmental gradients, portable data loggers (DHT22 sensors interfaced with Arduino) were installed at two distinct canopy depths. The air temperature ($T_{air}$) and relative humidity ($RH$) were recorded to calculate the Vapor Pressure Deficit (VPD) [26]. To ensure that the bioimpedance dataset captured this intra-tree variability, a stratified random sampling method was applied (Figure 2).

2.4.2. Dataset Structure and Preprocessing

The final dataset comprised 400 total bioimpedance measurements, derived from the longitudinal tracking of 100 fruits across four 15-day intervals (4 timepoints × 100 fruits), each linked to its final Dry Matter label. The final destructive analysis yielded a class distribution of 58 fruits labeled Harvest-Ready and 42 fruits labeled Immature. Class 1 fruits were more prevalent in the sun-exposed upper canopy (35 vs. 23), confirming non-uniform ripening rates.

Table 2. Representative Sample of the Bioimpedance Dataset. (Values at 7.2 kHz. $\left|Z\right|$: impedance magnitude ($\mathsf{\Omega }$); $\theta$: phase angle (°); R: resistance ($\mathsf{\Omega }$); X: reactance ($\mathsf{\Omega }$).)

Fruit ID	Tree	Week	$\left|\mathit{Z}\right|$	$\mathit{\theta }$	R	X	Class
F-001	T-01	1	4820	−8.3	4768	−694	–
F-001	T-01	2	4510	−9.1	4450	−714	–
F-001	T-01	3	4190	−9.8	4128	−711	–
F-001	T-01	4	3870	−10.4	3803	−698	1 (Ready)
F-012	T-02	1	6230	−5.2	6207	−564	–
F-012	T-02	2	6080	−5.6	6051	−593	–
F-012	T-02	3	5910	−5.9	5877	−607	–
F-012	T-02	4	5780	−6.1	5748	−614	0 (Immature)

presents a representative sample of the dataset illustrating the feature structure and class labels.

To ensure reproducibility, the raw data from AD5933 were converted into the full complex impedance spectrum. For each frequency point, the input feature vector was constructed using four components: Magnitude ($\left|Z\right|$), Phase Angle ($\theta$), Resistance (R), and Reactance (X). Outliers were screened using a Z-score threshold of $\pm 3\sigma$. Subsequently, Min-Max normalization was applied to scale all features to the $[0,1]$ range. For the LSTM architecture, the dataset was reshaped into a 3D tensor of shape $(Samples\times TimeSteps\times Features)$, where $TimeSteps=4$.

2.5. Experimental Protocol

2.5.1. Calibration Protocol and Validation

To rigorously correct for the parasitic impedance of the cabling and the AFE phase shift, a classic Open–Short–Load (OSL) compensation strategy was applied.

• Open/Short Correction: Measurements were taken with probes separated (Open) and shorted (Short) to characterize leakage admittance and residual series impedance, respectively.
• Load Correction (Gain Factor): A bank of precision metal-film resistors ($0.1\%$ tolerance) covering the physiological range (2–$10\mathrm{k}\mathsf{\Omega }$) was used to compute the System Gain Factor.
• Phase Validation (RC Dummy Cell): A dummy cell consisting of a $5.6\mathrm{k}\mathsf{\Omega }$ resistor in parallel with a $10\mathrm{nF}$ capacitor was tested. The system achieved a phase accuracy of $\pm 0.5^{\circ }$ compared to the theoretical Bode plot of the RC network.

2.5.2. Reference Method: Dry Matter Analysis

To establish the ground truth for the predictive models, destructive analysis was performed only at the end of the longitudinal tracking period (Week 4) for all 100 tagged fruits, following the AOAC standard method [4]. A 10 g sample of the mesocarp was weighed ($W_{wet}$), dried in an oven at 105 °C for 24 h, and weighed again ($W_{dry}$). The Dry Matter percentage was calculated as $\%DM=(W_{dry}/W_{wet})\times 100$. Fruits with $\%DM\ge 23\%$ were labeled Harvest-Ready (Class 1); others were labeled Immature (Class 0) [2].

2.5.3. Validation Strategy: Stratified Group K-Fold Cross-Validation

A Stratified Group K-Fold Cross-Validation scheme ($k=5$) was implemented [27]. Unlike standard random splitting, this approach grouped samples by Tree ID. In each iteration, data from 8 trees (80%) were used for training, while data from the remaining 2 trees (20%) were held strictly for validation. Importantly, all four timepoint measurements from any individual fruit always remained within the same fold—either entirely in training or entirely in validation. This grouping explicitly prevents temporal leakage: no information from a fruit’s early timepoints can appear in training while its later timepoints appear in testing.

2.5.4. Hyperparameter Tuning

Hyperparameter optimization was performed using an exhaustive Grid Search strategy within the inner loops of cross-validation [25]. For the LSTM network, the Adam optimizer ($\eta =0.001$) and Binary Cross-Entropy loss were employed, with Early Stopping (patience = 10 epochs) to prevent overfitting.

Table 3. Hyperparameter Search Space and Optimal Configuration (The optimal values represent the mode selected across the 5 cross-validation folds).

Model	Hyperparameter	Grid Search Space	Optimal Value
PLS-DA	Number of Components	$[2,3,\dots ,10]$	4
SVM	Kernel	{‘linear’, ‘rbf’, ‘poly’}	‘rbf’
	C (Regularization)	{0.1, 1, 10, 100, 1000}	10
	$\gamma$ (Gamma)	{‘scale’, ‘auto’, 0.01, 0.1}	‘scale’
Random Forest	N Estimators	$\{50,100,200,500\}$	200
	Max Depth	$\{\mathrm{None},10,20,30\}$	20
	Min Samples Split	$\{2,5,10\}$	5
LSTM	Hidden Units (L1)	$\{32,64,128\}$	64
	Hidden Units (L2)	$\{16,32,64\}$	32
	Dropout Rate	$\{0.1,0.2,0.3,0.5\}$	$0.2$
	Batch Size	$\{8,16,32\}$	16

details the search spaces and optimal configurations.

Table 4. LSTM Hyperparameter Sensitivity (Mean accuracy (±SD) across a 5-fold CV at each tested value, holding other parameters at their optimal values. The values highlighted in bold are the best values for each parameter).

Hyperparameter	Values Tested	Optimal	Best Acc.	Sensitivity
Hidden Units (L1)	32, 64 , 128	64	$0.92\pm 0.01$	Medium
Hidden Units (L2)	16, 32 , 64	32	$0.92\pm 0.01$	Low
Dropout Rate	0.1, 0.2 , 0.3, 0.5	0.2	$0.92\pm 0.01$	Medium
Batch Size	8, 16 , 32	16	$0.92\pm 0.01$	Low
Learning Rate	0.0001, 0.001 , 0.01	0.001	$0.92\pm 0.01$	High

presents the sensitivity analysis for the LSTM model, quantifying the sensitivity of mean cross-validation accuracy to each major hyperparameter.

3. Results and Discussion

3.1. Preliminary Analysis Based on Feature Engineering

This section presents the analysis of variables prior to evaluating avocado maturity prediction performance. Several feature engineering processes [23] are applied to determine the most relevant variables.

3.1.1. Analysis of Agronomic Covariates and Feature Importance

To determine the relative contribution of environmental vs. physiological variables, the study incorporated auxiliary datasets including cumulative precipitation and soil nutrient concentrations (N, K, P) [28]. Feature importance analysis using Random Forest (Figure 3) revealed that bioimpedance features (specifically at 7.2 kHz and 5 kHz) accounted for over 85% of the model’s decision-making weight. Soil nutrients showed a negligible contribution to the specific task of identifying the immediate harvest maturity window.

3.1.2. Impact of Electrode Configuration

The reliability of the data acquisition system was heavily dependent on the probe design. Consistent with Schwan’s theory on electrode polarization [13], the 4-electrode configuration proved to be superior. As shown in Figure 4, the 2-electrode setup exhibited a high variance ($\sigma \approx \pm 1250\mathsf{\Omega }$) due to skin resistance. The 4-electrode setup effectively isolated the mesocarp impedance, significantly reducing noise ($\sigma \approx \pm 120\mathsf{\Omega }$). Figure 5 shows the bioimpedance trajectories of maturation at different stages of maturation, comparing ready and immature harvests.

3.1.3. Spectral and Temporal Analysis

Feature engineering revealed specific frequency bands sensitive to maturity. Figure 6 illustrates the correlation ($R^2$) between impedance and dry matter. Two distinct peaks were observed at 5000 Hz and 7200 Hz, aligning with Equivalent Circuit Modeling (ECM) approaches [11]. Figure 7 highlights the temporal evolution of the signal, showing a consistent decay in impedance magnitude at 7.2 kHz.

3.2. General Analysis and Model Benchmarking

3.2.1. Robustness and Stability Analysis

The application of Group K-Fold Cross-Validation revealed the stability of the evaluated architectures. Figure 8 illustrates the distribution of accuracy scores. Boxplots represent the spread of classification accuracy across 5 folds. The LSTM model (green) achieves the highest median accuracy and lowest variance. While PLS-DA showed significant variance between folds, the LSTM model demonstrated superior stability with a compact distribution ($92.0\%\pm 1.0\%$), which confirms effective generalization.

Table 5. LSTM Per-Fold Cross-Validation Metrics (MAE: Mean Absolute Error; RMSE: Root Mean Square Error; $R^2$: coefficient of determination between predicted probability and binary label).

Fold	Accuracy	AUC	MAE	RMSE	${\mathit{R}}^2$
1	0.93	0.95	0.09	0.18	0.72
2	0.91	0.93	0.11	0.21	0.68
3	0.92	0.94	0.10	0.19	0.70
4	0.93	0.95	0.09	0.18	0.73
5	0.91	0.93	0.11	0.20	0.69
Mean	0.92	0.94	0.10	0.19	0.70
SD	0.01	0.01	0.01	0.01	0.02

provides per-fold regression metrics for the LSTM model.

In the case of the best model (LSTM), overfitting risk was managed through three mechanisms: (1) Group K-Fold ensures the model is evaluated on entirely unseen data; (2) dropout regularization (rate 0.2) reduces co-adaptation of LSTM units; and (3) Early Stopping prevents training beyond the point of validation loss improvement. Despite these strategies, the dataset is limited to a single orchard ($n=100$), and generalizability to other agro-climatic zones—where soil salinity, rootstock, and altitude differ—cannot be assumed without additional validation studies.

3.2.2. Model Performance Benchmarking

Table 6. Comprehensive Model Benchmarking. (Accuracy reported as Mean ± SD with 95% CI. MAE, RMSE, and $R^2$ computed on predicted class probabilities against binary labels. MedAE: Median Absolute Error.)

Model	Accuracy [95% CI]	Prec.	Rec.	F1	AUC	MAE	RMSE	${\mathit{R}}^2$	MedAE
PLS-DA	0.78 ± 0.04 [0.73, 0.82]	0.75	0.72	0.73	0.79	0.21	0.35	0.38	0.19
SVM	0.85 ± 0.03 [0.81, 0.88]	0.83	0.81	0.82	0.86	0.15	0.28	0.55	0.13
Random Forest	0.88 ± 0.02 [0.85, 0.90]	0.86	0.85	0.85	0.90	0.12	0.24	0.63	0.10
LSTM	0.92 ± 0.01 [0.90, 0.94]	0.91	0.91	0.91	0.94	0.10	0.19	0.70	0.08

details the metrics for all the evaluated models. The LSTM architecture, which exploits the temporal sequence of impedance measurements, achieved the highest performance across all metrics. The superior accuracy of machine learning approaches over PLS-DA confirms the non-linear nature of the impedance–maturity relationship.

It is important to note that the $R^2$ values reported in Table 6 (ranging from 0.38 for PLS-DA to 0.70 for LSTM) should not be interpreted under the conventional regression criterion of $R^2\ge 0.95$, since the models addressed in this work are binary classifiers, not continuous regressors. In this context, $R^2$ was computed between the predicted class probabilities (continuous output in [0, 1]) and the binary ground-truth labels (0 or 1), which inherently limits the achievable coefficient of determination regardless of model quality: a perfect classifier that outputs exactly 0 or 1 still does not yield $R^2=1$ when the label distribution is unbalanced (58% vs. 42% in this study). The appropriate performance indicators for evaluating binary classification are Accuracy, AUC, F1-Score, Precision, and Recall, all of which are reported in Table 6 and in Figure 9. By these primary metrics, the LSTM achieves strong performance (Accuracy = 92%, AUC = 0.94, F1 = 0.91), and the moderate $R^2$ values are fully consistent with these results. The $R^2$ column is included as a supplementary descriptor of the probability-output calibration, and its relatively low values reflect the structural mismatch between a regression-oriented metric and a classification task, not a deficiency in model performance.

3.2.3. Signal Tracking and Model Fidelity

Figure 10 compares raw bioimpedance measurements and LSTM predictions. Environmental measurement noise is attenuated in the LSTM output, with the physiological impedance decay trajectory reconstructed more smoothly. In this way, the LSTM successfully tracks the general downward trend of maturation while ignoring transient disturbances.

3.2.4. Impact of Longitudinal Data Sequence Length

A critical hypothesis of this study is that the history of the fruit’s development contains predictive information unavailable in a single measurement. To quantify the value of the longitudinal dataset size (temporal depth), the performance of the LSTM model (trained on the full sequence of 4 bi-weekly measurements) was compared against single-point (“static”) baseline models trained exclusively on the final snapshot data (Week 4).

Table 7. Impact of Temporal Depth on Model Accuracy (Comparison between single-point (static) and multi-point (longitudinal) inputs).

Dataset Scope	Model	Mean Accuracy	AUC	Gain
Single-Point (Week 4 only)	Random Forest	$0.88\pm 0.03$	0.90	–
Longitudinal (Weeks 1–4)	LSTM	$0.92\pm 0.01$	0.94	$+4.0\%$ *

demonstrates that restricting the dataset to a single time-point ($T=1$) results in a performance drop. The inclusion of the full temporal sequence ($T=4$) improves Mean Accuracy by +4.0% classifier. To assess the statistical significance of this improvement, a paired t-test [29] was conducted on the fold-level accuracy scores of the LSTM ($T=4$) versus Random Forest ($T=1$): $t\left(4\right)=4.21$, $p=0.014$ (<0.05), confirming that the improvement is statistically significant.

3.3. Comparative Analysis with Other Avocado Maturity Assessment Technologies

This section presents a qualitative comparison of our approach with similar studies. The criteria considered include the cost of implementing the technology, whether it involves destroying or touching the fruit, the accuracy of the maturity measurements obtained, whether it takes into account the temporal aspect of avocado maturity, and whether special field preparation is required for its use.

Table 8. Comparison of Avocado Maturity Assessment Technologies.

Method	Type	Accuracy	Cost	Temporal	Field Ready
Dry Matter (Oven)	Destructive	100% (Ref)	Low	No	No
NIR Spectroscopy	Non-Destructive	>95%	High	No	Yes
FruitSense (Wi-Fi) [8]	Non-Contact	≈$91\%$	Medium	No	Yes
Proposed (EIS+LSTM)	Non-Destructive	92%	Low	Yes	Yes

shows the results of this comparison. While Wi-Fi CSI and NIR eliminate contact, they require expensive infrastructure and do not consider the temporal aspect in avocado maturity assessment. Also, Wi-Fi CSI, NIR, and EIS+LSTM require field preparation for data collection. Dry Matter is destructive and does not consider the temporal aspect. The proposed EIS+LSTM approach offers a cost-effective solution for field measurements where infrastructure is limited.

4. Conclusions

This study developed a low-cost, field-deployable bioimpedance spectroscopy system, and implemented four predictive models on the longitudinal impedance data provided by the system for non-destructive classification of Hass avocado harvest maturity in Colombia. The system has a custom tetrapolar AFE, designed around the AD5933 converter, which was validated through Open–Short–Load calibration, successfully isolating mesocarp impedance from surface contact artifacts. The identified 5.0–7.2 kHz frequency band exhibited the strongest correlation with destructive DM measurements, consistent with known electrochemical changes associated with cell membrane degradation during physiological maturation.

From a predictive perspective, among the evaluated classifiers, the LSTM architecture achieved the highest mean accuracy (92.0%, AUC = 0.94), outperforming the linear PLS-DA baseline by 14 percentage points and the best single-point classifier by 4.0%. These results indicate that harvest maturity is more effectively characterized as a dynamic trajectory than as a scalar threshold at a single observation.

The preliminary findings are subject to important limitations. The dataset comprises 100 fruits from a single high-altitude Andean orchard, and reflects the specific agro-ecological conditions of a region The bioimpedance baselines are sensitive to agro-ecological factors, including soil salinity, seasonal rainfall, rootstock variety, and altitude. Generalization to other production regions requires additional calibration studies to adjust the predictive models to the environmental values in each region and avoid degrading their performance. Finally, the AD5933 hardware limits exploration to frequencies below 100 kHz; the exploration of high-frequency phenomena (>100 kHz) could offer additional tissue insight.

Future work will focus on extending the dataset to multiple orchards and implementing Transfer Learning protocols to reduce the cost of deployment in new sites. For example, the definition of a pre-trained LSTM core that is fine-tuned with small calibration datasets from new orchards minimizes the need for extensive destructive testing or learning processes. Additionally, the integration of this sensor technology into autonomous IoT nodes or drone-mounted manipulators is planned. This would enable the creation of real-time “Maturity Maps” for large-scale plantations, allowing producers to execute selective harvesting strategies based on precise physiological data rather than calendar estimation.