Flexible Visible Spectral Sensing for Chilling Injuries in Mango Storage

Abstract

Mango, as an important economic crop in tropical and subtropical regions, suffers from chilling injuries caused by postharvest low-temperature storage, which seriously affect its quality and economic benefits. Traditional detection methods have limitations such as low efficiency and strong destructiveness. This study designs and implements a flexible visible light spectral sensing system based on visible light spectral sensing technology and low-cost environmentally friendly flexible circuit technology. The system is structured based on a perception-analysis-warning-processing framework, utilizing laser-induced graphene electroplated copper integrated with laser etching technology for hardware fabrication, and developing corresponding data acquisition and processing functionalities. Taking Yunnan Yumang as the research object, a three-level chilling injury label dataset was established. After Z-Score standardization processing, the prediction accuracy of the SVM (Support Vector Machine) model reached 95.5%. The system has a power consumption of 230 mW at 4.5 V power supply, a battery life of more than 130 days, stable signal transmission, and a monitoring interface integrating multiple functions, which can provide real-time warning and intervention, thus offering an efficient and intelligent solution for chilling injury monitoring in mango cold chain storage.

1. Introduction

Mangoes are known as the ‘king of tropical fruits’ and are loved by people for their excellent taste and aroma. They are also one of the best-selling fruits in the world [1]. As a typical respiratory burst tropical fruit, chilling injuries during post-harvest storage are a key bottleneck limiting the sustainable development of the industry. Although low-temperature storage can inhibit respiratory metabolism and extend shelf life, it is difficult to precisely control the temperature threshold. When the temperature falls below the physiological critical value, it can easily cause an imbalance in cellular reactive oxygen species metabolism and membrane lipid peroxidation, leading to irreversible damage such as water-soaked spots on the fruit skin and fibrosis of the fruit flesh [2]. According to FAO statistics, economic losses caused by chilling injuries to mangoes after harvest account for 30% to 50% of total production worldwide [3]. Traditional testing methods are destructive and inefficient, with a single sample testing cycle lasting up to 2–4 h, a sample loss rate exceeding 80%, and a labor cost per sample of $15–20. Non-destructive technologies such as near-infrared spectroscopy and magnetic resonance imaging avoid sample damage, but their equipment costs exceed 100,000 USD and 1 million USD, respectively, and the instruments typically weigh over 50 kg. Data collection rates are only 10–15 samples per hour, significantly below the real-time online monitoring technical requirements of 20–30 samples per minute for smart agriculture production lines [4,5].

In the interdisciplinary field of agricultural engineering and intelligent detection technology, visible light spectroscopy sensing has emerged as an ideal technical approach for non-destructive monitoring of chilling injuries in mangoes due to its natural compatibility with the absorption characteristics of plant pigments. Abnormal pigment metabolism in mango fruit peel during the early stages of chilling injury (such as chlorophyll degradation and carotenoid structural changes) can be accurately characterized by differences in the reflectance spectrum in the 400–760 nm wavelength range. This technology, with its advantages of low equipment cost and fast detection speed, has shown potential in fruit and vegetable quality monitoring [6]. However, traditional rigid spectral systems are constrained by geometric structural limitations, resulting in 15–20% path length errors during detection due to insufficient contact between the probe and the mango’s ellipsoidal surface, leading to significant spatial deviations in reflectance spectral data. In contrast, flexible visible light spectroscopy technology achieves curved surface adaptation through a deformable substrate, controlling path length fluctuations to below 5%, effectively reducing error interference in traditional spectroscopy detection and providing technical support for the precise acquisition of multi-wavelength spectral parameters [7]. At the same time, traditional single-waveband spectral mode analysis lacks coupled modeling of multi-band parameter changes, resulting in insufficient generalization capabilities of chilling injury prediction models [8].

Innovations in flexible electronics technology provide innovative solutions to the above problems. For example, Wang et al. used a flexible organic thin-film sensor array to successfully monitor the hardness of different areas on the surface of apples in real time, verifying the feasibility of flexible sensor technology in fruit and vegetable quality inspection [9]. Wang et al. developed a gas sensing system sensitive to banana ripeness based on flexible graphene/PDMS composite membranes, which accurately determines banana ripeness by detecting changes in ethylene concentration [10]. In terms of chilling injury monitoring, Xia et al. used a flexible pressure sensor array to monitor mechanical damage during fruit storage. Its high sensitivity enabled the capture of minute deformation signals, providing a new approach to detecting tissue softening caused by chilling injuries [11]. The above research indicates that flexible electronics technology has promising applications in agricultural product quality monitoring. Its non-invasive and high-adhesion characteristics provide technical reference for monitoring chilling injuries in mangoes.

In addition, machine learning algorithms have become the core technology for non-destructive testing of fruit and vegetable quality in the intelligent transformation of agricultural engineering. Song et al. achieved an accuracy rate of 97.7% in apple brown spot disease identification by automatically extracting deep features from spectral data using CNN [12]. Relevant studies have demonstrated that by leveraging the kernel function mapping of Support Vector Machine (SVM), a determination coefficient (R2) of 0.8 was achieved in the prediction of soluble solid content in kiwifruits [13]. Additionally, transfer learning has been applied to egg freshness classification, where fine-tuning of pre-trained models reduced the mean absolute error (MAE) of predictions from 3.16 to 1.39 [14]. These research findings provide critical technical paradigms and methodological references for the monitoring of mango chilling injuries.

This study integrates visible light spectral sensing with flexible electronic technology to develop a monitoring system for mango chilling injury. The flexible sensing system based on laser-induced graphene technology enables in-situ dynamic acquisition of spectral data at 8 characteristic wavebands, while the prediction model constructed by support vector machine (SVM) algorithm achieves an accuracy of 95.5% for Yunnan Yumang samples and a robustness of 88.89% against ambient light interference. The system conducts automated “perception-analysis-warning” processes, addressing the issues of destructive measurement, high cost, and poor real-time performance in traditional detection methods. It provides an accurate solution for mango chilling injury monitoring and offers references for the development of intelligent preservation technologies for tropical fruits and the digital transformation of the fruit and vegetable industry.

2. Materials and Methods

2.1. System Architecture

This study proposes a flexible visible light spectral sensing system combined with machine learning to achieve real-time monitoring of chilling injuries in mangoes during post-harvest storage. The system aims to monitor whether mangoes have suffered chilling injuries, predict future chilling injuries, and provide specific information on their extent in real time. As shown in Figure 1a, chilling injury is a critical factor affecting mango quality during the post-harvest period from transportation to sales. However, since chilling injuries result from the cumulative physiological changes at the cellular level of mangoes, they are difficult to detect early. To accurately capture this information, as shown in Figure 1b, we selected visible light spectral reflectance data for experimental purposes. Traditional rigid circuit boards can cause damage to mangoes; as shown in Figure 1c, we selected flexible circuits to build the monitoring system. As shown in Figure 1d, the flexible spectral data is transmitted wirelessly to the upper-level computer platform for processing. As shown in Figure 1e, the obtained data is subjected to feature extraction and a chilling injury prediction model is established. As shown in Figure 1f, all relevant data is visualized and displayed in real-time on a cloud platform, enabling precise prediction of mango chilling injuries.

2.2. FVSS (Flexible Visible Light Spectroscopy System) Implementation

Flexible circuits have obvious advantages when applied to irregular surfaces such as fruits, but their preparation is a key technical challenge. Wan et al. have proposed a low-cost, small-scale flexible circuit manufacturing process that is particularly suitable for the production of flexible visible light spectroscopy system circuits on the surface of mangoes [15]. This paper prepared FVSS according to the above method. As shown in Figure 2, LIG was prepared by irradiating a PI film with a CO₂ laser, then conductive silver paste was applied to both sides of the LIG and fixed with PTFE. It was then placed in an electroplating solution for copper plating. After copper plating, the circuit was engraved with a UV nanosecond laser to remove excess copper and LIG. Finally, the electronic components were soldered, and the outer layer was encapsulated with transparent silicone.

As shown in Figure 3a, both the microcontroller module and the sensor module are composed of a PI substrate, LIG, copper, electronic components, and an encapsulation layer. Their 3D simulation diagrams, physical diagrams, and flexible display and testing are shown in Figure 3b–d, respectively, verifying the excellent flexibility, stability, and ease of use of the FVSS, which can closely adhere to the surface of a mango to collect spectral data. The core component of the system is the AS7341 multispectral sensor (Austrian manufacturer ams OSRAM) from AMS. It features 11 optical channels, 9 filters, and a detection wavelength range of 350 nm to 1000 nm. The 8 visible light channels (Channels F1–F8 correspond to 405 nm to 425 nm, 435 nm to 455 nm, 470 nm to 490 nm, 505 nm to 525 nm, 545 nm to 565 nm, 580 nm to 600 nm, 620 nm to 640 nm, and 670 nm to 690 nm, respectively) are critical for analyzing the physical and chemical composition of mangoes. The sensor processes optical channel information through 6 ADCs. Data from the 11 channels is input into the ADCs via a multiplexer, and the spectral digital signals are transmitted to the microcontroller via an I2C interface composed of SCL and SDA. The system schematic diagram, where the microcontroller module and sensor module form the system, enables real-time collection of spectral reflectance data from the mango surface, and uploads it to the host computer for processing.

Figure 1. Overall system architecture. (a) The process from mango harvesting to consumption; (b) and (c) spectral sensor detection principle; (d) sensing system logic diagram; (e) prediction model confusion matrix; (f) cloud platform monitoring interface.

Figure 1. Overall system architecture. (a) The process from mango harvesting to consumption; (b) and (c) spectral sensor detection principle; (d) sensing system logic diagram; (e) prediction model confusion matrix; (f) cloud platform monitoring interface.

Figure 2. Flexible circuit production flowchart.

Figure 2. Flexible circuit production flowchart.

Figure 3. Hardware implementation of flexible sensing system. (a) Exploded view of MCU and sensor module; (b) circuit 3D diagram; (c) actual photo; (d) system flexibility demonstration and actual measurement.

Figure 3. Hardware implementation of flexible sensing system. (a) Exploded view of MCU and sensor module; (b) circuit 3D diagram; (c) actual photo; (d) system flexibility demonstration and actual measurement.

2.3. Chilling Injury Model

This study employs a supervised learning approach to establish a chilling injury monitoring model. As shown in Figure 4, the overall acquisition scheme of mango chilling injury data is presented. The mangoes used in the chilling injury data acquisition experiment were divided into four parts: mangoes at 13 °C for feature acquisition, mangoes at 5 °C for feature acquisition, mangoes for 13 °C for label establishment, and mangoes at 5 °C for label establishment. On the first day of the acquisition experiment, all mangoes were placed in the corresponding constant temperature chambers. From day 2 to day 15, spectral acquisition was performed on all feature acquisition mangoes. The spectral data of mangoes stored at the same temperature collected on the same day were put into the same group and given numbers such as An or Bn, where A represents the storage temperature of 13 °C, B represents the storage temperature of 5 °C, and n is the number of days minus one. After feature acquisition using the feature acquisition mangoes, several label establishment mangoes were taken out and placed in a 25 °C constant temperature chamber, and the corresponding group labels were established after waiting for 2 days.

Before labeling, mangoes must be categorized. The storage process is divided into three categories. The first category consists of mangoes stored at 13 °C under normal conditions. The second category includes mangoes stored at 5 °C in a low-temperature environment but without chilling injury. These mangoes are stored below 13 °C but have not experienced chilling injury, and their flavor and quality have not been significantly impaired. If the low temperature persists, chilling injury will occur. The third category consists of mangoes that have already suffered chilling injury, which occurred due to prolonged storage at low temperatures. These categories correspond to labels 0, 1, and 2, respectively. To avoid subjective bias, the determination of chilling injury is made through a multi-person visual assessment method, with the formula as follows:

$$CI={\displaystyle \frac{\sum Number of levels\times Corresponding number of fruits}{Total number of fruits\times Highest level number}}$$

(1)

$$Chilling injury labels=\left\{\begin{array}{c}If CI<0.5, and k=0, Then 0\\ If CI<0.5, and k=1, Then 1\\ If CI\ge 0.5, Then 2\end{array}\right.$$

(2)

CI stands for chilling injury index. Number of levels represents the number of mango grades with or without chilling injury, with 0 indicating no chilling injury and 1 indicating chilling injury. The corresponding number of fruits represents the number of mangoes in that chilling injury grade, Total number of fruits represents the total number of mangoes in the group used to assess the chilling injury grade, and The highest level number represents the highest grade of chilling injury in the mangoes, with chilling injury grades being 0 or 1. Mangoes without chilling injury are classified as Grade 0 mangoes, while those with chilling injury are Grade 1 mangoes. If CI is greater than or equal to 0.5, the group of mangoes has suffered chilling injury; otherwise, they have not. k represents the storage temperature of the mangoes, where k = 0 indicates a storage temperature of 13 °C, and k = 1 indicates a storage temperature of 5 °C.

Using the above method, the collected mango reflectance spectral data were labeled and used for training the chilling injury monitoring model.

2.4. Experimental Scheme

On 22 January 2025, 70% ripe Yumang (a specific mango cultivar) were harvested from a professional cultivation zone in Yunnan Province. The collection was executed during the winter harvest season (January–February) as the first batch of 2025 production for this cultivar. Samples were airlifted to Beijing for experimental purposes within a 48-h transportation window to maintain physiological integrity. The experiment selected mangoes at 70% maturity, as they had not yet entered the respiratory burst phase, resulting in low ethylene release and high starch content. This helps maintain skin hardness and physiological stability, ensuring the accuracy of spectral measurements. Fruits at this maturity level can be stored at room temperature for 10–15 days and at low temperatures for 21–28 days, thereby reducing transportation losses and extending the experimental period. Their compact cellular structure can withstand mechanical damage, and their slow maturation process and stable metabolic levels ensure the reliability of experimental data.

After selection, they were placed in a thermostatic chamber pending the experiment. The selected mangoes were stored at 13 °C for 12 h and then divided into two groups, A and B. Group A (A1: 34 pieces; A2: 70 pieces) and Group B (B1: 34 pieces; B2: 70 pieces) were stored at 13 °C and 5 °C, respectively. Before the experiment, the mangoes were labeled. Mangoes in groups A1 and B1 were taken out daily and their reflection spectra at 405–690 nm were measured with a flexible visible light spectral sensing system in a dark box, with 3 points for each mango, and then put back to their original environment for 14 consecutive days. All groups of mangoes were stored at 90% relative humidity, and the average light intensity in the laboratory environment was 280 lux.

(1)

Spectral characteristics acquisition

During the 14-day experiment, the reflectance spectra of all mangoes in the A1 and B1 groups were measured daily using a flexible visible light spectroscopy system. After measurement, the mangoes were returned to their storage environment. Meanwhile, five mangoes were taken from the A2 and B2 groups each day and placed in a 25 °C environment. After two days, a visual assessment was conducted. The measurement of reflectance spectrum data was conducted in a dark box environment using flexible sensors attached to the surface of the mangoes. Reflectance spectra were obtained from three points on the mango peel. These three points were selected at approximately 10 cm from the straight line between the top and bottom of the mango, and at approximately 3 cm to the left of the midpoint of this line segment. Each sample was collected from the same part of the mango. The measurement results were the average of three measurements.

(2)

Physical and chemical indices

The physical and chemical indices measured during the experiment were color difference and weight loss rate.

(a)

Color difference

Significant changes in color difference also occur during mango chilling injury and ripening and can be used as training data for judging and predicting chilling injury. While measuring the reflected spectrum data, a CR-400 color difference meter (Konica Minolta in Japan) was used to measure the a*, b*, and L* values at three points on the mango skin. The measurement results were taken as the average of three measurements.

(b)

Weight loss rate

The weight loss rate of mangoes was measured using a JA5003 high-precision electronic scale (Shanghai Jingke, China). The weight of the same mango on the current day and the previous day was recorded. The weight loss rate formula is as follows:

σ = (G₀ − G)/G₀ × 100%

(3)

σ represents the weight loss rate of mangoes, G₀ is the weight of mangoes on the previous day, and G is the weight of mangoes on the current day.

(3)

Model evaluation

In this study, data were collected from mangoes stored at 13 °C, mangoes stored at 5 °C for 1–6 days, and mangoes stored at 5 °C for 7–14 days, with 1428, 612, and 816 data points, respectively. The visible light spectrum wavebands collected included eight wavebands, namely 405–425 nm, 435–455 nm, 470–490 nm, 505 nm, 545–565 nm, 580–600 nm, 620–640 nm, and 670–690 nm. The software used for data processing and analysis was Origin2024. In this study, the spectral data were randomly divided into a training set and a test set, with an 80% to 20% ratio. The performance evaluation metrics selected for the mango prediction model include accuracy, precision, recall, and F1-Score [16]. The formulas are as follows:

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(4)

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(5)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(6)

$$F1 Score=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(7)

Among them, TP refers to true positive cases, TN refers to true negative cases, FP refers to false positive cases, and FN refers to false negative cases.

(4)

Flexible sensing system performance

By using an electrochemical workstation to measure the system operating current at varying voltages, we evaluated the power consumption of FVSS. To characterize the signal transmission performance, we assessed the system PLR (Packet Loss Rate) by conducting measurements at multiple communication distances, while maintaining consistent transmission power and environmental conditions. Ambient light interference evaluation aims to determine the accuracy of the chilling injury prediction model under two conditions: a dark box and a laboratory environment.

3. Results and Discussion

3.1. Spectral Characteristics Analysis

As shown in Figure 5a, the skin color of mangoes stored at 13 °C changed from a darker green to a yellowish-green. Mangoes stored for a shorter period of time remain darker in color after being left at room temperature for two days, but by the sixth day, there are signs of the skin turning yellow. By the 10th day, most mangoes exhibited obvious signs of ripening, characterized by a brighter and yellower skin. The surface condition of mangoes stored at 5 °C and then placed in a room-temperature environment for 2 days is shown in Figure 5b. It can be observed that under low-temperature refrigeration, the surface state of mangoes from days 1 to 5 is not significantly different from those at 13 °C. However, unlike at 13 °C, mangoes stored at 5 °C do not exhibit a noticeable yellowing of the peel until after the sixth day, with clear yellowing beginning only on the 12th day. Upon visual assessment, chilling injury symptoms began to appear on the surface of the mangoes from the 7th day onwards, with the symptoms gradually worsening thereafter.

The changes in reflectance values across eight spectral wavebands over a 14-day period at 13 °C and 5 °C were summarized separately, as shown in Figure 5c,d. It can be observed that the reflectance values of mangoes stored at 5 °C exhibited a consistent downward trend across all eight spectral wavebands over the 14-day period, while those stored at 13 °C showed either a slow upward trend or relatively stable trends. After the first day of storage, the spectral characteristics of the mango surfaces stored at 13 °C and 5 °C have already shown significant differences. The reflectance values of the mango peel stored at 13 °C were generally higher, while those at 5 °C were lower. As storage time increases, the reflectance values of mango surfaces stored at 13 °C generally showed a trend of first increasing and then decreasing, while those stored at 5 °C exhibited an overall decreasing trend. This indicates that mango surfaces stored at 13 °C appear brighter, whereas lower temperatures reduce the glossiness of mango surfaces, resulting in a duller color [17].

As shown in Figure 5e, the trends in surface reflectance values of mangoes across different wavelength wavebands are similar, with no significant differences between wavebands. However, storage temperature had a more pronounced effect on reflectance values: mangoes stored at high temperatures exhibited higher reflectance values than those stored at low temperatures, and the difference became more pronounced as storage time increased. Based on the changes in weight loss rate shown in Figure 6a, this phenomenon may be attributed to changes in moisture content. During storage, water loss in mangoes causes a consistent change in light absorption and scattering across all visible light wavebands—water loss reduces fruit transparency, enhances light scattering, and thereby increases the intensity of reflected light across all wavebands. At normal storage temperatures, water loss in mangoes occurs more rapidly, resulting in higher reflectance values across all wavebands compared to the low-temperature storage group [18]. At the same time, the similarity in the trends of the reflected light spectra at different wavelengths may also be due to changes in the fruit tissue structure during long-term storage, leading to an increase in the degree of light scattering and reflection at all wavelengths [19].

In addition, a phenomenon was observed in the figure: during storage, regardless of whether the temperature was 13 °C or 5 °C, the reflectance values for almost all wavelengths showed wave-like changes rather than a steady increase or decrease. This may be due to dynamic changes in the surface color of mangoes caused by metabolic activities at different stages after harvest [20].

3.2. Physical and Chemical Indices Analysis

As shown in Figure 6a, the weight loss rate of mangoes during a 14-day storage period can be observed. In the initial days, the weight loss rates of mangoes stored at 13 °C and 5 °C were similar, maintaining a relatively stable growth rate. However, after the seventh day, the weight loss rate of mangoes accelerated significantly. This may be due to the weakened inhibitory effect of the temperature at 13 °C on the respiratory activity of the mango fruit, causing the mango to enter the respiratory transition phase and begin the ripening process, resulting in a significant increase in respiratory rate and a corresponding increase in water evaporation rate [21]. When mangoes are stored at 5 °C, the low temperature continues to inhibit their respiration, and the rate of weight loss remains almost stable throughout the 14 days.

L* represents the brightness of mangoes and is used to quantify the lightness or darkness of color. As shown in Figure 6b, the trend of L* changes in mangoes stored at 13 °C and 5 °C is entirely different. Mangoes stored at 13 °C exhibit a consistent upward trend in L*, indicating a continuous increase in skin brightness, which aligns with the brightness changes observed during the mango ripening stage. In contrast, mangoes stored at 5 °C showed a continuous decrease in L* over the 14-day measurement period, indicating that the surface becomes increasingly dull and lackluster, which is consistent with the characteristics of chilling injury in mangoes.

a* represents the red-green ratio of mangoes. As shown in Figure 6c, a* is negative, indicating that the mango peel tends to be green. As time increases, the a* values of mangoes stored at two temperatures show an increasing trend, meaning that the green components of the peel decrease while the red components increase. This suggests that regardless of mango ripening or chilling injury, the content of chlorophyll a and chlorophyll b in the peel decreases with time. The a* value of mangoes stored at 13 °C increases at a faster rate, which may indicate that the maturation process has a greater impact on the degradation of chlorophyll on the mango surface [22].

b* represents the yellow-blue hue of mangoes. As shown in Figure 6d, all b* values are positive, indicating that the skin color of mangoes leans more toward yellow than blue. Over time, the b* value of mangoes stored at 13 °C gradually increases, reflecting the continuous accumulation of β-carotene and lutein in the mango skin. In contrast, the b* value of mangoes stored at 5 °C decreases continuously. This may be due to enzymatic browning occurring in the low-temperature environment, where polyphenol oxidase (PPO) catalyzes the oxidation of polyphenols to form melanin, which absorbs light and inhibits the increase in b* value [23].

3.3. Model Evaluation

After training models using SVM, RF, MLP, and XGBoost algorithms with spectral data as features, the results are compared in

Table 1. Performance parameters of each model.

Algorithm	Accuracy	F1 Score
		Normal	ULT-Damage	Chilling Injury
RF	77.6%	0.85	0.61	0.79
MLP	87.1%	0.9	0.76	0.90
XGBoost	78.7%	0.87	0.64	0.78
SVM	95.5%	0.97	0.93	0.96

, which shows the accuracy and F1-score of the algorithm-trained models.

All algorithms performed significantly worse in the low-temperature non-chilling category compared to the other two categories, as this category overlaps with the feature space of the normal and chilling-damaged categories, with its spectral features lying between those of chilling-damaged mangoes and mangoes stored at 13 °C. The normal category achieved the best classification performance, primarily due to the larger training sample size. The SVM algorithm achieved an accuracy rate of 95.5%, with an F1-Score of 0.93 for the low-temperature non-chilling damage category, far exceeding other algorithms; while the MLP algorithm lagged behind SVM, it outperformed the remaining algorithms. Based on the excellent performance of SVM and MLP in training nonlinear models, it is inferred that spectral chilling injury features exhibit nonlinear separability, requiring complex decision boundaries [24]. Figure 6e,h show the SVM classification test confusion matrix and the importance of visible light spectral wavebands, respectively.

Using color difference (L*, a*, b*) as features, the results of training models using SVM, RF, MLP, and XGBoost are shown in

Table 2. Performance parameters of each color difference model.

Algorithm	Accuracy	F1 Score
		Normal	ULT-Damage	Chilling Injury
RF	64.0%	0.75	0.39	0.60
MLP	60.5%	0.74	0.03	0.51
XGBoost	64.3%	0.74	0.36	0.62
SVM	55.2%	0.70	0.10	0.48

. The accuracy rates did not exceed 70%, with XGBoost performing the best (64.34%) and SVM the worst (55.24%). This may be due to sample imbalance or the difficulty of accurately reflecting mango chilling injury using color difference features. As shown in Figure 6f, the confusion matrix of the XGBoost model with the highest accuracy rate indicates severe misclassification of low-temperature non-chilling injury, with a recall rate of only 27.7%, and the model is almost unable to distinguish between normal and low-temperature non-chilling injury categories.

The model based on color difference achieved a maximum accuracy rate of only 64.34%, making it difficult to effectively distinguish between the three types of mangoes. However, the model trained using SVM with visible light spectra as features demonstrated excellent ability to distinguish between the three types of mangoes, achieving an accuracy rate of 95.5%, which meets production requirements.

3.4. Flexible Sensing System Performance

(1)

Power consumption

The power consumption of a flexible visible light spectrum sensing system is a key parameter affecting its operating time. It must have low power consumption characteristics to meet the requirements of long-term operation under battery power supply [25]. When testing voltages of 4 V, 4.5 V, and 5 V, it was found that while the 4 V power supply provided stable operation, the LED light flickered, making it unsuitable for stable system operation. As shown in Figure 7, the current curves for 4.5 V and 5 V power supplies are similar, with the 5 V power consumption being approximately 255 mW and the 4.5 V power consumption being approximately 230 mW. Based on a 200 mAh 5 V battery, the system can operate continuously for up to 130 days, far exceeding the mango refrigeration cycle. Conservatively estimated, the interval between two charges/battery replacements exceeds 150 days.

The formula for calculating power consumption (P) and battery life (T) is:

$$P=V\times I$$

(8)

$$T={\displaystyle \frac{200 \mathrm{m}\mathrm{A}\mathrm{h}}{I_{avg}}}\times 0.8$$

(9)

where V is the power supply voltage, I is the operating current, I_avg is the average operating current, and 0.8 is a coefficient added to account for battery efficiency.

(2)

Signal transmission

The purpose of this test is to verify the transmission stability and anti-interference capabilities of WiFi signals in simple and complex scenarios. Good FVSS signal transmission should have a long transmission distance and good anti-interference capabilities [26]. The test results are shown in

Table 3. Communication performance parameters.

	Outdoors				Indoors
Communication distance (m)	20	40	60	80	1	5	10	15	20
PLR	0	0	0	0.2%	0	0	0	0	0.3%

. As can be seen, when the WiFi and flexible sensing system are placed in an open outdoor environment, the effective communication distance is within 80 m. When the distance exceeds 80 m, the packet loss rate (PLR) significantly increases, making stable communication impossible. In a laboratory environment with multiple device interferences, packet loss begins to occur when the communication distance reaches 20 m, and communication becomes unstable beyond 20 m. In actual mango storage cold storage facilities, where device density is lower than in the laboratory, communication quality and distance should be better.

(3)

Ambient light interference

Machine learning models have high accuracy in predicting chilling injuries to mangoes in dark rooms. Considering that mangoes may be exposed to light during packing or unpacking, environmental light interference tests were conducted to assess the robustness of the model [27].

Table 4. Average reflectance values in darkroom and laboratory environments.

Waveband (nm)		405–425	435–455	470–490	505–525	545–565	580–600	620–640	670–690
Reflectance Value	Dark box	41.3	49.8	100	193.1	427.3	518.6	588.5	445.2
	Lab	41.8	50.3	101.2	195.5	433.1	523.7	595.6	452.3

shows that the reflectance values for each waveband under laboratory ambient light conditions are higher than those in a darkroom (increase of 0.5–7.1, within 2% of the original wavelength value). This is because the sensor is tightly adhered to the mango, and external light is reflected through the skin into the sensor, significantly reducing its intensity. After inputting the ambient light data into the model,

Table 5. Classification test of mango chilling injuries under ambient light.

Category	Accuracy	Recall	F1 Score
Normal	0.76	1.00	0.87
ULT-damage	0.97	0.69	0.81
Chilling injury	0.99	0.98	0.99

shows an accuracy rate of 88.89% (compared to 95.5% for the dark box data). Although there is a slight decrease, the system still maintains good judgment capabilities, indicating that the flexible visible light spectral sensing system has a certain degree of robustness against ambient light interference.

4. Conclusions

This study developed a flexible visible light spectroscopy sensing system for post-harvest mango chilling injury monitoring. Integrating visible light spectroscopy with low-cost flexible circuit technology, it overcomes traditional detection limitations by enabling non-destructive, real-time monitoring through a sensing-analysis-alert-handling workflow. The system innovative hardware design uses novel flexible circuit manufacturing, reducing costs while fitting fruit surfaces. An SVM-based prediction model achieved 95.5% accuracy using spectral data and maintained 88.89% accuracy under ambient light, showing strong robustness. Comprehensive testing verified stable cloud connectivity, real-time monitoring, and environmental interference resistance. However, there are still many challenges in terms of commercial deployment, such as the frequent maintenance requirements of flexible circuits, the high cost of large-scale system deployment, and the significant impact of complex ambient light on spectral data in real-world environments. Future research will focus on expanding sample diversity, integrating multi-source data, achieving hardware miniaturization, developing calibration algorithms, and enhancing communication capabilities to further improve system performance and practicality.