📄 agriculture

Integrating postharvest quality assessment and deep learning to identify the optimal harvest maturity stage of apricot fruit

This study identifies Stage 3 as the optimal harvest maturity for apricots based on superior postharvest quality retention and develops a lightweight deep learning model using RGB images to rapidly and non-destructively classify maturity stages for improved harvest decision-making.

Original authors: Shima Naeimifar, Hojatollah Bodaghi, Seyed Iman Saedi

Published 2026-09-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shima Naeimifar, Hojatollah Bodaghi, Seyed Iman Saedi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every year, orchards around the world produce millions of apricots, a fruit celebrated for its sweet flavor and rich nutrition. Yet, these fruits face a relentless race against time. Unlike many crops that can be picked and stored for months, apricots are soft, perishable, and ripen quickly once they leave the tree. If picked too early, they remain hard and tasteless; if picked too late, they become mushy and spoil before they reach the market. The key to solving this problem lies in finding the exact moment of harvest, a stage of development where the fruit is ripe enough to taste good but still firm enough to survive the journey to a grocery store. For decades, farmers have relied on their eyes and experience to guess this moment, a method that is often inconsistent and prone to error. Now, a new approach combines the careful study of fruit chemistry with advanced computer vision to find a more reliable answer.

Researchers at Shahrood University of Technology in Iran set out to solve this puzzle by first understanding the biology of the apricot and then teaching a computer to recognize the right moment to pick it. They focused on a specific variety known as Zafarani, growing in the Shahrood region. The team began by harvesting fruit at five distinct stages of growth, ranging from very immature to fully ripe. They placed these fruits in a cold storage environment, mimicking the conditions used by commercial distributors, and watched how they changed over thirty-five days. Every week, the scientists measured a wide range of qualities: how hard the fruit was, how much weight it lost as water evaporated, the level of sweetness, the amount of acid, and the precise color of the skin. They were looking for the single stage where the fruit held its shape and flavor the best while resisting the natural decay that comes with time.

The results of this long observation revealed a clear winner. The fruits harvested at the third stage of maturity, a point that is neither too green nor fully soft, proved to be the most resilient. These apricots kept their firmness better than those picked earlier or later. They lost less weight, meaning they stayed plump and juicy longer. Their chemical balance, including their sweetness and acidity, remained more stable, and their color changed less dramatically than the others. Fruits picked too early struggled to hold their quality, changing rapidly and losing firmness quickly. Those picked too late were already too soft to withstand the storage period, deteriorating almost immediately. The study confirmed that this specific window of time offers the best compromise, allowing the fruit to be stored for weeks without losing the qualities that make it desirable to eat.

Having identified this ideal harvest stage, the researchers turned to the second part of their work: teaching a machine to find it. They realized that if a computer could look at an apricot on the tree and instantly know if it was at that perfect third stage, farmers could harvest with much greater precision. To do this, they took over a thousand photographs of the apricots in the orchard under natural sunlight, capturing the fruit exactly as it appeared in the real world, with all the shadows and varying angles that a farmer would encounter. They did not use special lights or controlled setups; they wanted the system to work in the messy reality of an orchard.

The team then built a custom computer program, a type of artificial intelligence known as a deep learning model, specifically designed to sort images. This program was trained to look at the photos and categorize the fruit into one of the five maturity stages. The model was designed to be lightweight, meaning it could run quickly on standard devices without needing massive computing power. After training on thousands of images, the computer was tested on a new set of photos it had never seen before. It successfully identified the correct maturity stage in the vast majority of cases. When it made mistakes, they were usually small errors, such as confusing a fruit that was just about to reach the perfect stage with one that had just passed it. This showed that the computer understood the gradual changes in the fruit's appearance rather than just memorizing random patterns.

The significance of this work lies in its ability to bridge the gap between biological science and practical farming. By first proving which stage of growth leads to the longest shelf life, and then creating a tool to identify that stage instantly, the researchers have offered a path toward reducing waste and improving quality. The computer vision system does not replace the farmer's knowledge but augments it, providing an objective, rapid way to make harvest decisions. While the study was conducted on a single variety of apricot in one location, the method suggests a powerful future for agriculture. If this technology can be adapted to other fruits and different growing conditions, it could transform how produce is harvested, ensuring that the fruit reaching consumers is as fresh and flavorful as the day it was picked. The study concludes that while the tool is promising, further testing across different orchards and seasons is needed to confirm its reliability on a commercial scale.

Technical Summary: Integrating Postharvest Quality Assessment and Deep Learning to Identify the Optimal Harvest Maturity Stage of Apricot Fruit

Problem Statement
Apricot fruit (Prunus armeniaca cv. Zafarani) is highly perishable, characterized by rapid softening and quality deterioration at ambient temperatures. The harvest maturity stage is a critical determinant of postharvest performance, storage life, and consumer acceptance. However, determining the optimal harvest time currently relies heavily on human visual estimation, which is subjective, uncertain, and inefficient. Traditional image processing methods for fruit classification often lack robustness under natural orchard conditions due to their dependence on controlled lighting and fixed camera geometries. Furthermore, while deep learning has advanced agricultural monitoring, its specific application to non-destructive, real-time apricot maturity classification under unstructured field conditions remains relatively unexplored. There is a need to integrate physiological quality data with advanced computer vision to establish a reliable, non-destructive method for identifying the specific maturity stage that maximizes storage potential.

Methodology
The study was conducted in two integrated phases:

Physiological Quality Assessment:
- Experimental Design: Apricot fruits were harvested at five distinct physiological maturity stages (Stages 1–5) based on Edwards' maturity scale.
- Storage Protocol: Fruits were stored at 6°C for 35 days. Quality attributes were evaluated at weekly intervals (0, 7, 14, 21, 28, and 35 days).
- Metrics: The study measured firmness, weight loss, total soluble solids (TSS), titratable acidity (TA), pH, and color attributes ( $L^*$ , $C^*$ , $h^\circ$ , and $\Delta E$ ).
- Statistical Analysis: Data were analyzed using a factorial design in a completely randomized layout (CRD) with three replications, utilizing SAS software for mean comparisons (Duncan's test, $p < 0.05$ ).
Deep Learning Model Development:
- Data Acquisition: 1,032 RGB images of apricots were captured under natural orchard conditions (June–July 2024) using a Samsung Galaxy A54 smartphone. No control was exerted over lighting, distance, or angle.
- Dataset Partitioning: The dataset was split into training (660 images), validation (165 images), and independent test (207 images) sets.
- Architecture: A custom lightweight dual-path Convolutional Neural Network (CNN) was developed. The architecture features two parallel branches:
  - Path 1: Utilizes regular convolutional layers.
  - Path 2: Utilizes dilated convolutional layers (dilation factor = 2) to expand the receptive field without reducing resolution or increasing computational load, thereby recovering spatial information lost during pooling.
  - Integration: Feature maps from both paths are concatenated and passed to dense layers with a Softmax activation for classification.
- Training: The model was trained using the Nadam optimizer, categorical cross-entropy loss, a batch size of 8, and 80 epochs within the Google Colaboratory environment (Python, Keras/TensorFlow).
- Evaluation: Performance was assessed using precision, recall, F1-score, and confusion matrices.

Key Results

Physiological Findings:
- Optimal Stage: Fruits harvested at Stage 3 demonstrated the most favorable overall postharvest performance.
- Firmness: Stage 3 fruits exhibited superior firmness retention compared to Stages 4 and 5, which softened rapidly, and Stages 1 and 2, which showed significant changes.
- Weight Loss: Stage 3 showed the lowest cumulative weight loss over the 35-day period.
- Chemical Stability: Stage 3 fruits maintained the smallest relative changes in TSS (4.16%), pH (6.54%), and titratable acidity, indicating greater compositional stability.
- Color: While all stages underwent color changes, Stage 3 displayed a balanced progression in hue angle ( $h^\circ$ ) and total color difference ( $\Delta E$ ), avoiding the instability seen in earlier stages and the advanced senescence of later stages.
Deep Learning Performance:
- The dual-path CNN successfully classified the five maturity stages under natural field conditions.
- Metrics: On the independent test set, the model achieved a weighted precision of 0.81, recall of 0.78, and F1-score of 0.79.
- Error Analysis: Misclassifications predominantly occurred between adjacent maturity stages (e.g., Stage 2 vs. Stage 3), reflecting the gradual biological transition between stages rather than distinct visual discontinuities. The model demonstrated good generalization with no significant overfitting.

Key Contributions

Identification of Optimal Harvest Window: The study empirically identifies Stage 3 as the specific physiological maturity stage that offers the best compromise between ripeness and storage stability for 'Zafarani' apricots, based on a multi-parameter quality assessment.
Novel Architecture: The development of a lightweight, dual-path CNN incorporating both regular and dilated convolutions specifically tailored for fruit maturity classification in unstructured environments.
Integrated Approach: The research successfully bridges the gap between traditional postharvest physiological assessment and modern deep learning, using the former to define the target classes for the latter.
Non-Destructive Field Application: The model was trained and validated on images captured under natural, uncontrolled orchard conditions, demonstrating feasibility for real-world deployment.

Significance and Claims
The paper claims that integrating physiological quality assessment with deep learning provides a robust framework for rapid, non-destructive maturity-based harvest decision-making. The primary significance lies in the ability to objectively identify Stage 3 maturity in the field, enabling growers to initiate harvest when fruits possess the highest potential for extended storage life.

The authors emphasize that the deep learning model was designed to recognize the maturity stage identified as optimal by the physiological data, rather than independently determining optimality. They conclude that while the approach shows promise, further validation across different cultivars, orchards, seasons, and variable imaging conditions is necessary to establish the robustness and practical applicability of the system for commercial deployment. The study does not claim immediate commercial readiness but rather establishes a proof-of-concept for a scalable, AI-driven solution to apricot postharvest management.

More like this