A Remote Strawberry Health Monitoring System Performed with Multiple Sensors Approach

Abstract

Temperature is a key physiological indicator of plant health, influenced by factors including water status, disease and developmental stage. Monitoring changes in multiple factors is helpful for early diagnosis of plant growth. However, there are a variety of complex light interference phenomena in the greenhouse, so traditional detection methods cannot meet effective online monitoring of strawberry health status without manual intervention. Therefore, this paper proposes a leaf soft-sensing method based on a thermal infrared imaging sensor and adaptive image screening Internet of Things system, with additional sensors to realize indirect and rapid monitoring of the health status of a large range of strawberries. Firstly, a fuzzy comprehensive evaluation model is established by analyzing the environmental interference terms from the other sensors. Secondly, through the relationship between plant physiological metabolism and canopy temperature, a growth model is established to predict the growth period of strawberries based on canopy temperature. Finally, by deploying environmental sensors and solar height sensors, the image acquisition node is activated when the environmental interference is less than the specified value and the acquisition is completed. The results showed that the accuracy of this multiple sensors system was 86.9%, which is 30% higher than the traditional model and 4.28% higher than the latest advanced model. It makes it possible to quickly and accurately assess the health status of plants by a single factor without in-person manual intervention, and provides an important indication of the early, undetectable state of strawberry disease, based on remote operation.

1. Introduction

Strawberries are widely grown worldwide for their considerable gross value of production and nutritional value, which has attracted the attention of most growers [1,2]. However, strawberry growth has high environmental requirements and is greatly affected by external environmental factors such as light and temperature [3,4]. The existing detection and analysis methods need to optimize the detection process and index selection according to different environmental conditions to improve the accuracy and reliability of detection [5,6,7,8]. Traditional artificial health monitoring is labor-intensive and subjective [9]. Although RGB imaging can be automated for visual inspection [10], the problem of difficult phenotypic feature acquisition for plant health monitoring due to leaf occlusion still requires robotic or human intervention [11]. In order to overcome the difficulties, an online plant health monitoring method based on an infrared thermal imaging sensor is applied to strawberry cultivation. Canopy temperature is used as a scalable alternative term to avoid the waste of resources, reduce annotation dependency while maintaining deployability and realize remote monitoring of strawberry health status in a wide range [12].

Early plant health monitoring methods were mostly based on traditional laboratory methods. Mass spectrometry or nuclear magnetic resonance spectroscopy are used to measure isotope distribution in steady-state metabolite repositories [13]. Homeostatic labeling studies have quantitatively assessed the metabolism of developing seeds, and isotope nonstationary metabolic flux analysis (INST-MFA) has recently been applied to plant leaves to evaluate metabolic analysis and functional description of plant cells [14]. At present, the preferred method for detection of plant non-volatile compounds is matrix-assisted laser desorption ionization mass spectrometry (MALDI-MSI) [13]. Dalisay et al. analyzed the health of intact Linum usitatissimum capsules and seed tissue at different developmental stages by using the comprehensive omics method [15]. With the development of technology and the demand for on-site detection, the detection of volatile organic compounds (VOCs) has gradually been used as an important biomarker of plant health and surrounding environmental conditions [16]. Wang H et al. developed a bioelectronic nose based on gas biosensor array and signal processing model for non-invasive diagnosis of strawberry health conditions [17]. However, it is difficult to meet the needs of health assessment of large-scale plant communities due to its high cost and loss in practical application.

Nowadays, the remote monitoring of large-scale plant community growth has attracted extensive attention. Ampatzidis et al. developed a data acquisition and image processing technique using a small unmanned aerial vehicle (UAV) to assess the phenotypic characteristics of citrus crops [18]. However, there are many obstacles in the greenhouse, leading to detection problems and shadows on the target [19,20,21], which makes it difficult to implement the detection method based on RGB. Hyperspectral imaging has advantages such as collecting and processing images over a wider spectral range in the visible and infrared ranges [22,23,24]. Yiting Xie et al. used hyperspectral imaging technology to reveal biological stresses related to photosynthesis and realize early disease detection [25]. Since Yilian Tang et al. discovered the relationship between temperature and plant health in greenhouses, and infrared thermal imaging has great potential in thermal detection [26,27,28,29,30,31,32], Nicholle Hatton et al. obtained a direct correlation of 0.7404 between soybean fungal infection status and canopy temperature, using thermal infrared imaging [33]. However, the acquisition of phenotypic characteristics for plant health monitoring is limited by light interference in the greenhouse and often requires manual assistance. Therefore, solving the problem of light interference caused by light conditions and plant morphology is a problem worthy of more attention in greenhouse strawberry detection.

The combination of hyperspectral imaging with traditional imaging technology and the application of deep learning technology can reduce the influence of light interference caused by light conditions and be used to analyze the health status of the measured object more comprehensively [34,35,36,37,38]. Jensen et al. used the visual and thermal images of barley’s late growing season to draw the two growth stages of barley by discriminant function analysis, and the classification accuracy was 83.5% [39]. Harshana and Ogawa et al. try to use a deep learning algorithm to identify two types of agricultural products (ripe strawberries and immature strawberries, respectively), and the evaluation accuracy rate is 82.62% [40]. However, in the automatic monitoring of plant health in the greenhouse, the light interference caused by the overlapping crown leaves of strawberries is often ignored [41]. To solve the above problems, Grant et al. used leaf temperature to indirectly measure the water use efficiency of strawberries [42,43,44]. While deep learning methods like attention-based transformers [25] or wearable sensors [45] offer high-resolution plant phenotyping, they demand significant computational resources, annotation effort and energy-limiting scalability in resource-constrained greenhouses. Based on this, we propose a soft-sensing method for leaves, which can perform adaptive screening of thermal images and realize remote monitoring of strawberry fruit health in a greenhouse.

In this paper, an adaptive leaf soft-measurement method based on an Internet of Things system and infrared thermal imaging sensor for large-scale strawberry health monitoring is proposed. Firstly, the relationship between strawberry health and canopy temperature was studied to solve the problem of incomplete fruit characteristic information caused by strawberry crown leaf covering, and a soft-sensing method of strawberry leaf based on infrared thermal imaging sensor was proposed. Secondly, the laws of radiation (mainly solar zenith angle) and reflection (mainly flowers and leaves; greenhouse roof also have an effect) are studied, interference items are analyzed, a fuzzy comprehensive evaluation model is established and a temperature correction algorithm is designed to reduce the influence of prediction errors. Finally, a control strategy is proposed to only wake up the image acquisition node when the data collected by the local solar altitude sensor and the environmental information acquisition node meets the requirements, to reduce the interference of the periodic change in solar zenith angle and environmental temperature and humidity on the strawberry health monitoring system.

2. Materials and Methods

2.1. Equipment Conditions and Materials

The test experiment was conducted in parallel with the control group. The control experiments were carried out in an ambient room, where air temperature, humidity and light intensity were precisely controlled. Infrared thermal imager was used to photograph strawberry crown leaves, and specific strategies are shown in Figure 1a. The experimental group test was conducted in the strawberry greenhouse planting area of Jiangsu Agricultural Expo Garden in China, which took half a month. The whole planting area is a 10 m × 10 m greenhouse, in which nine strawberry ridges are planted: each ridge is 0.8 m wide and 8 m long, and the distance between the ridges is 0.2 m. The range in temperature, humidity and light intensity of strawberry greenhouse in one year was 15~30 °C, 70~80% RH and 15,000~45,000 lx, respectively. The maximum temperature during the test was 21~25 °C, and the daytime temperature range was ±2~3 °C. The temperature of different places in the greenhouse at a given time is considered to be the same.

In addition, strawberries are cultivated in a nutrient solution, so they are considered to be free of water stress and fertilizer imbalance. A total of 500 images were taken. Grouped according to health status, 94 canopy images showed diseases of different degrees, and the rest were healthy strawberry canopy images. Disease samples were augmented via ±1.5 °C thermal noise injection to simulate infection gradients, preserving physiological plausibility. Grouped by maturity, 94 crown leaf photos were of almost all young strawberries, 256 crown leaf photos were of high proportion of fruit during development and 150 crown leaf photos were of high proportion of mature fruit.

2.2. Strawberry Health Monitoring IoT System Platform

The hardware part of the Internet of Things monitoring system is composed of environment detection node and image acquisition node. The temperature and humidity sensor, light intensity sensor and sun angle sensor constitute an environment detection node to collect information such as temperature, humidity, sun angle and weather. As shown in Figure 2, DHT11 (Guangzhou Aosong Electronics Co., Ltd., Guangzhou, China) is selected as the temperature and humidity sensor, and its collection range is 0~50 °C and 20~95% RH, with errors of ±2 °C and ±5% RH, respectively. Five photodiodes are placed in a blind cylinder with small holes to form a solar angle sensor. The GY-30 (Shenzhen Zhichuang Micro Intelligent Electronics Co., Ltd., Shenzhen, China) illuminance sensor has an acquisition range of 1~65,535 lx with an error of ±1 lx. Considering that the temperature and humidity in the greenhouse change little in a short time, the allowable error range of our algorithm is ±2.5 °C and ±6% RH, respectively. The measurement range and error of the above sensors prove that these sensors can indeed measure the temperature accurately to meet the constraint judgment. As shown in Figure 2, OV5640 visible light image acquisition module (OmniVision, shanghai, China) and lepton3.5 infrared thermal image acquisition module (FLIR Systems, Inc., La Jolla, CA, USA V1.3.0) constitute image acquisition nodes.

Two systems of environmental parameter and image information processing are constructed on the cloud platform. Each environmental detection node and image acquisition node is equipped with a BC26 module (Shenzhen Boshi Jie Technology Co., Ltd., Shenzhen, China, V1.0.3) for data transmission. The environment acquisition node can send the collected environmental information to the cloud regularly, and the cloud can also send commands to control the image acquisition module for image acquisition. We split the visible and thermal images into multiple packets and sent them to the cloud. We added a CRC check method to the information to ensure that the cloud could accurately receive the image data. The proposed strawberry health monitoring IoT system framework is shown in Figure 1. We adjusted the position of both cameras so that they could capture the same area. The arranged sensor nodes interact twice with the processing system in the cloud. The workflow of the overall strawberry health monitoring IoT system is shown in Figure 1b.

2.3. Data Preparation

The image processing algorithm runs on the server side. All images received should be pre-processed prior to the screening process. First, the input visible image is converted to gray image, and the gray connection is established with the corresponding thermal image. Secondly, median filter and Gaussian filter are used to smooth the noise generated by the imaging sensor. Then, Otsu algorithm is used to separate the canopy from the background soil cover. Finally, the threshold value calculated by Otsu algorithm is used to binarize the image.

The health monitoring system is implemented based on image processing technology, as shown in Figure 1a, including three steps of screening, calculation and diagnosis. Firstly, the information of interference items on the image is extracted, which is substituted into the interference item evaluation model for calculation, and the images with high interference index removed. Then, the temperature correction algorithm of steps ③–⑤ in Figure 1a is used to calculate the average temperature of the canopy leaf in the region of interest. In order to eliminate the measurement error caused by the change in air temperature, normalization is realized. The linear mapping relationship between temperature information and gray level was established using Lepton 3.5’s factory calibration data [46]. To validate thermal calibration accuracy, we placed reference objects (blackbody sources at 20 °C, 25 °C, 30 °C) within the field of view during imaging. The mean absolute error (MAE) between sensor-reported and ground-truth temperatures was 0.48 °C, confirming sufficient precision for agricultural applications. Finally, the average temperature calculated from the uploaded images is used for health analysis. If the monitoring area is determined to be unhealthy, the server, together with other IoT nodes, will send the obtained diagnostic information such as the disease area, status and possible causes to the client.

2.4. Fuzzy Comprehensive Assessment of Environmental Factors

The growth pattern is affected by many factors such as sun altitude angle, temperature, weather conditions and strawberry plant morphology. Within a certain influence range, the blade soft-measurement method is accurate, but beyond this influence range, it is not accurate. In order to evaluate the comprehensive influence of different factors on the blade soft-sensing method, a fuzzy comprehensive evaluation method is proposed to quantify these interference terms. The evaluation indicators are as follows:

$$F=\left\{SZA,AT,WC,IT\right\}$$

(1)

where $SZA$ is an abbreviation of the solar altitude angle, $AT$ is an abbreviation of the air temperature, $WC$ is an abbreviation of the weather condition and $IT$ is an abbreviation of the interference term. The evaluation set is given by the following set:

$$V=\left\{LI,MI,HI\right\}$$

(2)

where LI represents low impact, MI represents moderate impact and HI represents high impact. The manually designed evaluation index weight vector is given by the following equation:

$$A=\left(w_1,w_2,w_3,w_4\right)$$

(3)

where $w_i$ ($i=1,2,3,4$) means the weight of the corresponding indicator in the evaluation indexes set (1). The range of above $w_i$ is from 0 to 1 and the sum of all $w_i$ is 1. The manually designed evaluation matrix is given by the following matrix:

$$R=\left[\begin{array}{c}0.300.300.40\\ 0.250.500.25\\ 0.250.250.50\\ 0.300.300.40\end{array}\right]$$

(4)

where each row of the matrix $R$ represents an evaluation vector of one evaluation index in the evaluation indexes set. The comprehensive evaluation model is given by the following formula:

$$B=Normalize\left(A·R\right)$$

(5)

Finally, by multiplying the evaluation level vector given by the following:

$$V^{’}=\left(\mathrm{25,70,100}\right)$$

(6)

a comprehensive evaluation value is obtained to evaluate the comprehensive influence of environmental factors on the model results.

2.5. Methods of Strawberry Health Status Assessment

The leaf soft-measurement method proposed in this paper is a method to evaluate the health status of strawberry by measuring the average temperature of the canopy and leaves of strawberry plants. Leaf area index (LAI) is an important ecological and biophysical parameter related to leaf growth. Yoshihiro Hirooka et al. obtained the following growth model by studying the relationship between effective cumulative canopy leaf temperature and leaf area index [47]:

$$LAI=\mathrm{\alpha }{\mathrm{e}}^{\beta T}$$

(7)

where $\mathrm{\alpha }$ corresponds the initial $LAI$ value at $T=0°\mathrm{C}$ (${iLAI}_{exp}$); $\beta$ (°C⁻¹) is the parameter referred to as the relative leaf growth rate (RGR) and $T$ (°C) is effective cumulative canopy temperature.

The health evaluation of a single strawberry fruit is based on the appearance characteristics of the fruit in different growth stages, such as color, shape, size, lesion and maturity [48]. For the whole plant, overall health is assessed by the ratio of the number of healthy fruits to the total number of all incubated fruits. After standardized operation, the overall health of the strawberry plant is limited to a range of 0 to 1. The above definition process can be given by the following equation:

$$H_{oa}={\displaystyle \frac{n_{H_{sg}}}{N}}$$

(8)

where $H_{oa}$ represents the overall health of a strawberry plant, $n_{H_{sg}}$ defines the number of healthy single strawberry fruit and $N$ is the total number of all gestated fruits.

For strawberry plants bearing multiple fruits, LAI was positively correlated with fruit load [49]. As the strawberry fruit ripens, the fruit load of the plant increases. The greater the LAI, the greater the degree of leaf cross, the lower the photosynthetic efficiency and the increase in respiration, resulting in higher crown leaf temperature rise. In this article, the definition of strawberry plant health is shown in Equation (8). In conclusion, fruit load, crown leaf respiration and crown leaf temperature increase with the ripening of strawberry fruits. Therefore, there was a positive correlation between strawberry plant health and crown leaf temperature.

2.6. Methods Evaluation

In order to further verify the effectiveness of the method for monitoring strawberry health status through temperature, the strawberry crown leaf temperature before and after correction was substituted into Equation (13), and the calculated strawberry health status was compared with the overall health evaluation value of the comprehensive variable, so as to obtain a positive correlation between the average strawberry crown leaf temperature and strawberry health status. Jun Steed Huang et al. proposed the semi-covariance (nonlinear) coefficient of comparison with Pearson’s (linear) correlation coefficient, as shown in Equations (9) and (10), to study the positive correlation of variables in Equations (13) and (8) [50].

$$S_{pos}\left(X,Y\right)={\displaystyle \frac{E\left[ReLU\right((X-E[X\left]\right)(Y-E[Y\left]\right)\left)\right]}{\sqrt{E\left[X^2\right]-{\left(E\left[X\right]\right)}^2}\sqrt{E\left[Y^2\right]-{\left(E\left[Y\right]\right)}^2}}},$$

(9)

$$S_{neg}\left(X,Y\right)={\displaystyle \frac{E\left[ReLU\left(-\left(X-E\left[X\right]\right)\left(Y-E\left[Y\right]\right)\right)\right]}{\sqrt{E\left[X^2\right]-{\left(E\left[X\right]\right)}^2}\sqrt{E\left[Y^2\right]-{\left(E\left[Y\right]\right)}^2}}},$$

(10)

where $S_{pos}$ is the positively correlated covariance coefficient or convergent part, and $S_{neg}$ is the negatively correlated covariance coefficient or divergent part. All variables were divided into four quadrants based on the mean of every two variables. $S_{pos}$ belongs to the first and third quadrants, and $S_{neg}$ belongs to the second and fourth quadrants.

Finally, model robustness was evaluated via 5-fold cross-validation and bootstrap resampling (1000 iterations). Classification metrics (accuracy, precision, recall, F1) were computed. Environmental variability was controlled through multi-condition thresholds: solar zenith angle (36–70°), illuminance (30,000 ± 500 lx) and tri-daily sampling (10:00, 13:00, 15:00).

3. Results

3.1. Data Preparation and Temperature Correction

Part of the experimental results on sunny and cloudy days are shown in Figure 3a, where maturity is obtained by Equation (8). The temperature data collected in the experiment was normalized to eliminate the overall changes in canopy temperature caused by periodic changes in air temperature. The z-score normalization conversion function is as follows:

$$x^{*}={\displaystyle \frac{\left(x-\overline{x}\right)}{\sigma }}$$

(11)

where $x$ is the original sample data, $\overline{x}$ is the average value of all sample data and $\sigma$ is the standard deviation of all sample data.

Figure 3a,b show the normalization process and results of partial sample data. As shown in Figure 3a, the unprocessed temperature data collected on different dates is grouped. The average temperature of unprocessed temperature data collected in high-temperature weather is higher than that collected in low-temperature weather. The unprocessed temperature data was normalized to eliminate the air temperature dimension, and the temperature distribution after processing was consistent, as shown in Figure 3b.

In the thermal image of grayscale palette, the grayscale distribution of the thermal image is the same as that of the grayscale image. As shown in Figure 3c, the gray value corresponding to the lowest temperature of the thermal image is 0, the gray value corresponding to the highest temperature is 255 and the gray value corresponding to each temperature is between the two.

After temperature normalization and linear mapping between temperature information and gray level were established, a temperature correction algorithm was used to calculate the average temperature of canopy leaves in the region of interest (ROI). Figure 3d–i show the results of some key steps in the temperature correction algorithm. The images captured by the infrared thermal imager are mainly divided into shading conditions (Figure 3d) and sunshine conditions (Figure 3g). As shown in Figure 3e, shadow conditions contribute little to the choice of ROI. The algorithm can accurately identify ROI even if the shading condition includes the illumination condition. It can be seen from Figure 3f that the average temperature of shaded crown leaves is lower than that of sun-exposed crown leaves. As shown in Figure 3h, although the high coronal leaves produce shadows on the lower coronal leaves, the shaded areas on the lower coronal leaves are basically recognized. The test results show that the algorithm is robust to illumination conditions. Sunlight will cause the temperature of the sun-cured crown leaf to increase, which will affect the calculation of the average temperature.

3.2. Environmental Disturbance Item Evaluation Model

In order to eliminate the influence of interference terms on average temperature calculation, the evaluation model of interference terms is established. From the results of Figure 4a, it can be seen that the mean temperature of coronaries containing flowers, fruits and soil cover is 2.6 °C higher than that of coronaries containing only leaves. Figure 4b shows the optimal fitting curve of the blade and other terms. A low R² value indicates that there is no correlation between the two. The closer R² is to 0, the lower the correlation between leaf temperature and other interference terms. Therefore, it is difficult to compensate for the measurement errors caused by interference terms through calculation. To sum up, the temperature of the leaves in the canopy image is lower than other items, and if other interference terms are considered in the detection process, the regression analysis cannot be performed well.

The interference term index is calculated by the ratio of the interference term area to the total pixel area, to assess the effect of the interference term on the calculated canopy temperature. The calculation is given by the following:

$$\mathrm{I}\mathrm{d}\mathrm{x}={\displaystyle \frac{\sum f_{\mathrm{x}}}{\sum f}}$$

(12)

where $\mathrm{I}\mathrm{d}\mathrm{x}$ refers to the interference term index, $f_{\mathrm{x}}$ defines the pixels in interference term area and $f$ represents the total pixel area in a thermal infrared image. Considering that the area of the soil coverings has been segmented by the thresholding step in the temperature correction algorithm procedure, only flowers and fruits need to be detected. Figure 4d–i show the results of detecting the flowers and fruits area, which are used to calculate the interference term index. The detection of strawberry fruit and flowers was based on color, and the three components of H, S and V were threshold-divided in HSV space (illustrated in Figure 4e,h). After grayscale operation and Gaussian blurring, the image was converted to a binary image with a simple grayscale threshold. A dilation was utilized to fill the holes in the strawberry fruit and the flowers. After that, two erosion operations were performed to remove the individual holes that were incorrectly identified and two expansion operations were then used to restore the size of the foreground target (demonstrated in Figure 4f,i). Finally, the interference term index is obtained by calculating the ratio of the interference area to the total ROI. To evaluate HSV-based segmentation robustness, we manually annotated 200 images for flowers/fruits. The Intersection-over-Union (IoU) was 0.82 ± 0.07 under controlled lighting, but decreased to 0.61 ± 0.12 under strong shadows, indicating lighting sensitivity. For high-accuracy screening, we thus restricted image capture to solar zenith angles > 36° (Section 3.5).

The data set is filtered several times with different interference indicators, and the data with no more than the set indicators is used as the training set. The error value obtained by dividing the data set with different interference items is the Root Mean Squared error of the regression accuracy, and the change in the regression accuracy under certain interference items is observed. Figure 4c depicts a graph of this change. As shown in Figure 4c, the change curve first decreases and then increases, showing a saddle shape. The reason for this phenomenon is that the number of samples divided by the index of small interference items in the training set is very small, which leads to the unbalanced distribution of samples. With the increase in the number of samples, the error value gradually increases and tends to be stable in the range of 0.02–0.03. When the interference term index exceeds 0.03, the error value begins to increase significantly, which indicates that the interference term has a great influence on the calculation of canopy temperature. In summary, the thermal image screening conditions should make the interference index not exceed 0.03. The change curve in Figure 4c is used as the evaluation model of interference terms to eliminate the unpredictable interference caused by the complexity of images and guide the adaptive screening of thermal images. The IoU reduction from 0.82 to 0.61 under shadowed conditions justifies our solar angle constraint, which minimized lighting-induced segmentation errors.

3.3. Real-Time Assessment Model of Greenhouse Strawberry Health Status

After the temperature correction algorithm, the samples are randomly scrambled. Use the first 70% of the data as the training set and the rest as the test set. By looking at the training data, we end up choosing logarithm as the hypothesis function for fitting. The Levenberg–Marquardt optimization algorithm was used to iterate the training process, and the maximum number of iterations was set to 400 times. At this point, the function is a one-way nonlinear regression with only three variables to update and iterate. Therefore, the number of training data sets is sufficient to carry out data fitting on the training set, and the results are shown in Figure 5c. The trend of the fitted curve accords with the ideal growth model. We also note that many of the training samples are far from the fitted curve, giving a large residual distribution in Figure 5d. The predicted values are obtained by fitting the curve in Figure 5c.

Water stress, disease conditions and maturity have a great impact on canopy temperature by regulating metabolism, as shown in Figure 5a. Water deficit stress usually leads to stomatal closure, resulting in a canopy temperature higher than air temperature [51]. In addition, damage by root pathogens, systemic infections or leaf pathogens often affects the transpiration rate and water flow of the whole plant or plant organs, which also has a certain impact on canopy temperature [52]. Canopy temperature measured by thermal sensors is affected by crop growth stage. This is the physiological stage that affects the transpiration and respiration of the crop, which in turn, affects the thermal status of the crop. Therefore, considering the above interference terms, combined with the definition of strawberry health status in Equation (8), a nonlinear regression analysis was carried out on the sample data obtained from the control group experiment, as shown in Figure 5b. R² was 0.93808 and the residual sum of squares was 0.00688. Strawberry growth conditions were selected as independent variables, and normalized canopy temperature was selected as dependent variable. The growth model established is the inverse function of Equation (7), so the growth model of strawberry plant is as follows:

$$T_{norm}=0.78+0.41\times ln\left(H-0.01\right)$$

(13)

where $T_{norm}$ is the normalized canopy temperature, and $H$ is the health that comprehensively evaluates the growth condition of strawberries in this paper, including but not limited to maturity, disease, water and fertilizer problems.

Combined with the selection of interference term evaluation model, the passed sample image is almost not affected by the interference term. Taking into account the measurement error, the calculated temperature value is allowed to vary within a small range. As shown in Figure 5e, in the residual analysis of the growth model in Figure 5d, the interval is defined as the trusted residual interval by the maximum of the residual distribution. The training samples were screened through the trusted residual interval, as shown in Figure 5f. After filtering, data beyond the trusted residual interval is deprecated. If the residual obtained from the calculated value minus the predicted value is not within the trusted residual interval, it is determined that the monitored strawberry area is unhealthy.

3.4. Evaluation Model Effectiveness Analysis

The higher the Pearson index, the higher the positive correlation between the two variables. This verifies the contribution of the temperature correction algorithm to the growth model, and further compares the effect of the temperature-health assessment model with the actual value of the overall health assessment of the comprehensive variable. As can be seen from Figure 6a,b, the Pearson index of the calculated crown leaf average temperature before correction, namely the calculated strawberry health status and the overall health evaluation value of the comprehensive variable, was 0.7591, and the semi-covariance index after correction was 0.8493. Therefore, the adjusted temperature-health assessment model can reach 84.93% of the overall health assessment effect of the integrated variables. It outperformed the model established by Nicholle Hatton et al. for soybean fungal infection status and canopy temperature by 12.48% [33].

And a comparison experiment is made between the RGB-based method and the proposed method. As shown in Figure 6c, the R² of the proposed method is higher than that of the RGB-based method. The accuracy of the leaf soft-measurement method in judging the health status of strawberries can reach 0.869, which is 30% higher than traditional methods. Its F1 value is 0.84, the precision rate is 0.82 and the recall rate is 0.86 (Bootstrap 95%, CI: [84.7%, 89.1%]). Compared with Harshana et al. [40], the average accuracy rate for strawberry classification has increased by 4.28%. This is due to the large size and high density of strawberry crown leaves. As shown in Figure 6d, tree crown leaves can cover fruits, resulting in incomplete image feature information. Most of the experiments in the papers required the help of humans or robots to expose strawberry fruit. Instead, the method can indirectly monitor the health of strawberries by studying the relationship between canopy temperature and health status.

The comparison experiments above are intended to prove that the method can measure temperature more accurately, to satisfy the constraint judgment. However, it is still worth noting that when the light intensity is strong, the gray value of the collected crown leaf image will be greater than the true value, leading to the judgment of this part of the image as soil during threshold segmentation. Therefore, the ROI region of the image becomes smaller, which will lead to inaccurate calculation of the average temperature, resulting in errors.

3.5. System Performance Superiority Analysis

Considering that the performance of the BC26 module is easily affected by the greenhouse environment, experimental tests are carried out. The image acquisition node is fixed at the edge of the ridge. The occlusion index is proposed and defined as the ratio of the part of the BC26 antenna obscured by leaves to the total part. The respective information transmission rates of the antenna under different occlusion conditions are tested. As shown in Figure 7, the larger the area of the occluded antenna, the lower the information transmission rate. When fully blocked, the transfer rate can reach 2.4 KB/s. The bit error rate of the whole transmission process is 0%. Since the BC26 module has a maximum transmission of 512 bytes at a time, single visible and thermal images are 600 KB and 120 KB, respectively, so it needs to be subcontracted. The size of a single packet is set to 512 bytes, and a CRC check method is added to verify the information. The transmission of a visible image is about 254 s, and the transmission of a thermal image is about 50 s. After testing and verification, the image information can be transmitted in the greenhouse.

After the images are uploaded, algorithms deployed on Alibaba Cloud servers process and calculate the images. The entire process took about 54 s and was accurate at 0.869. It only takes 0.12 s to send the test results to your phone or computer and display the strawberry health information.

If an RGB camera is used, 360 visible images (about 240 MB) are required to cover the entire strawberry planting area, while ensuring detection accuracy. The image processing takes 2.4 min to run and no images are screened. Encouragingly, with an infrared thermal imager, only 180 images are needed (90 visible images and 90 thermal images, stored in a special file format for infrared thermal images, with a size of about 18 MB). If all the interference indexes of the images meet the evaluation model, the image processing only takes 0.9 min. Compared to traditional RGB camera-based methods, infrared thermal imaging-based methods save nearly 90% of data traffic and reduce processing time by nearly 60%.

The research on screening conditions is shown in

Table 1. Comparison of detection accuracy of several screening items (The “√” indicates that this item has been adopted.).

Solar Zenith Angle (36~70°)	Light Intensity (30,000 ± 500 lx)	Air Temp (25 ± 3 °C)	Humidity (73~78% RH)	Accuracy
√	√	√	√	0.87 ± 0.02
√	√	√		0.71 ± 0.04
	√	√	√	0.24 ± 0.02
√		√	√	0.26 ± 0.02
√	√		√	0.33 ± 0.02
		√	√	0.08 ± 0.01

. The value range of this parameter varies in different seasons. By comparing the detection accuracy under different screening conditions, it can be found that the more screening conditions, the higher the detection accuracy. In addition, the limiting of solar zenith angle and light intensity is more conducive to improving the accuracy of the system.

The STM32 and light sensor GY-30 is always in operation to sense changes in the sun’s altitude, and when the sun’s altitude meets the requirements, the cloud wakes up the environmental information acquisition node. The variation law of the sun’s altitude angle is shown in Figure 3c. According to Figure 3c, the temperature and humidity sensor DHT11 is set to work for 4 h per day, 4 times per hour, and each time for 3 s. When the data collected by the environmental information acquisition node meets the requirements, the cloud wakes up the CMOS module ov5640 and the infrared thermal imaging module Lepton3.5 for shooting, and sends the collected data back to the cloud. The power consumption per unit area of the environment detection node and image acquisition node is only 15,271 J per day, which is only 1% of the traditional RGB method.

4. Conclusions

This paper overcomes the difficulty of light interference caused by light condition and plant morphology in strawberry health monitoring in a greenhouse, and discusses the influence of interference terms on canopy temperature calculation. In order to reduce the influence of uneven illumination on images, a temperature correction algorithm is proposed. An evaluation model of interference items is established to screen the data set. According to the relationship between strawberry metabolism and canopy temperature, a feasible and accurate strawberry temperature-health evaluation model was established. Unlike hyperspectral methods requiring 100× more data, our system achieves 86.9% accuracy using only thermal channels, reducing processing time by 60%. This trade-off favors deployability in IoT settings where power and bandwidth are constrained. The additional sunlight angle and intensity sensors were used as the baseline to reduce inference of strawberry health with environmental interference, and the image acquisition node was activated only when the environmental interference was less than the specified value. The consumption per unit area per day was only 15,271 J, which was only 1% of the traditional RGB method. This system can provide a preliminary indication of the early disease status of the plant remotely, and in the future, it can even evaluate the earliest emergence time of each strawberry disease, providing a new idea for plant disease research.