Water and fertilizer self-adaptive irrigation control method and system based on mango growth stage

Abstract

The invention belongs to the technical field of agricultural automation, and particularly relates to a water and fertilizer self-adaptive irrigation control method and system based on a mango growth stage. The method comprises the following steps: firstly, acquiring a crown image every day through an image acquisition device, and extracting a bud number Nb, a fruit number Nf and a leaf greenness index G through processing; secondly, judging a growth stage S according to conditions; calculating a target irrigation amount range [alpha s * KQ, beta s * KQ] and a fertilizer solution concentration range [gamma s * KC, delta s * KC] according to the growth stage S query coefficient table; and finally controlling the irrigation system and the fertilization device to output target values. According to the method, accurate regulation and control of water and fertilizer are achieved through image recognition and growth stage linkage decision, and the method is mainly used for water-saving and fertilizer-saving management of mango planting.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology. More specifically, this invention relates to a method and system for adaptive irrigation control of water and fertilizer based on the mango growth stage. Background Technology

[0002] Mango cultivation has long relied on manual experience or fixed phenological calendars to determine growth stages. However, actual flower and fruit development is significantly affected by varietal characteristics, regional climate, and tree vigor. For example, the timing of flower bud emergence can vary by 7-10 days within the same orchard. Traditional methods, which use a uniform timeframe for switching water and fertilizer, result in over 30% of plants having water and fertilizer supply that is misaligned with their actual physiological state. Especially during the period of coexistence of flowers and fruit (when flower buds have not yet withered and young fruit has formed), the lack of quantitative standards often leads to incorrect classification as either the flowering or fruiting stage, causing an imbalance in nutrient distribution.

[0003] Existing intelligent irrigation systems (such as single-factor soil moisture control) do not correlate with key plant physiological parameters. Typical problems include: during the transition from budding to flowering, if soil moisture is adequate but the number of flower buds is insufficient, the system still applies high water levels as during flowering, inhibiting root development; during fruit growth, fertilization based solely on soil parameters easily overlooks leaf nutrient status, leading to excessive fertilizer use. This disconnect reduces water and fertilizer utilization by 15%-20% and exacerbates the risk of soil salinization.

[0004] Fixed-threshold models struggle to adapt to interannual climate fluctuations. In colder years, the transition from flowering to fruit development is delayed by 2-3 weeks, but existing systems still follow historical average switching strategies, resulting in insufficient water supply during the young fruit stage and a 25% decrease in fruit enlargement rate. Furthermore, the lack of a mechanism linking flower and fruit quantity makes it impossible to identify the transition period; simply attributing it to the flowering stage leads to insufficient potassium fertilizer application, affecting fruit sugar accumulation.

[0005] Manual inspections suffer from time lag: the daily variation rate of flower bud count during peak bloom exceeds 20%, but fruit farmers only record it every 3-5 days. While multispectral scanning by drones can acquire data, the processing cycle exceeds 48 hours and the cost per scan exceeds 1,000 yuan, making it difficult to support daily decision-making. Leaf sensors, which have been tried before, are easily damaged by wind and sand (lifespan < 2 months), flower sound recognition is affected by environmental noise (false positive rate > 40%), and satellite remote sensing has insufficient spatial resolution (> 10 meters) to distinguish individual flowers and fruits.

[0006] Existing solutions suffer from three major constraints: first, flower and fruit morphology recognition is greatly affected by light and humidity; second, multi-source data fusion lacks a dynamic weighting mechanism; and third, low-cost equipment struggles to balance accuracy and robustness. This has made precise water and fertilizer management in orchards a long-standing bottleneck in the industry. Summary of the Invention

[0007] One object of the present invention is to provide a water and fertilizer adaptive irrigation control method based on the mango growth stage, comprising the following steps: S1. Collect images of the mango tree canopy once a day using an image acquisition device, and process the canopy images using an image processing unit to extract the number of flower buds. Nb Quantity of fruits Nf and leaf greenness index G ; S2, based on extraction Nb , Nf and G And determine the current growth stage S : like Nb ≥6 and Nf When =0, the current growth stage is determined. S Flowering season; like Nb ≤5 and Nf If ≥1, then the current growth stage is determined. S This is the fruit growth period; like Nb ≥6 and Nf ≥1, calculate the transition period score TS=0.6×( Nb / 30)+0.4×( Nf / 50), when TS≥0.5, the current growth stage is determined. S This is the transitional period between flowering and fruiting; If none of the above conditions are met and G ∈[0.4,0.7], determine the current growth stage. S Maturity stage; Otherwise, determine the current growth stage. S The budding stage; S3. Growth stage determined in step S2 S The target irrigation volume range is calculated using a water and fertilizer decision function model. Q min ( S ), Q max ( S )] and target fertilizer concentration range[ C min ( S ), C max ( S )]: Q min ( S ) =α s × K Q , Qmax ( S )=β s × K Q , C min ( S )=γ s × K C , C max ( S )=δ s × K C ,in, Q For irrigation volume, C This refers to the concentration of the fertilizer solution. K Q As the baseline irrigation amount, K C As the baseline concentration, α S , β S , c S , d S To be related to the growth stage S The relevant coefficients are obtained by looking up a preset growth coefficient table; S4. Control the irrigation system's output irrigation volume. Q ∈[ Q min , Q max Simultaneously control the output concentration of the fertilizer application device. C ∈[ C min ( S ), C max ( S )).

[0008] Preferably, the image processing unit in step S1 processes the tree canopy image by including the following steps: a) The tree canopy RGB image is divided into flower bud region mask, fruit region mask and leaf region mask by a pre-trained DeepLabv3+ segmentation network; b) Based on the bud region mask, convert the pixels within the region to the HSV color space, and apply the dual-threshold conditional hue component H. hsv Binarize the image with saturation components Sat ∈ [0°, 15°] and saturation components Sat ∈ [0.4, 1.0] to generate a primary binary image of flower buds. Perform a 3×3 elliptic kernel morphological closing operation on the primary binary image to remove noise and holes, and then generate a connected component labeling map. In the connected component labeling map, filter connected components with a pixel area > 50; the number of these components represents the number of valid flower buds. Nb ; c) Based on the fruit region masks, input the pre-trained Mask R-CNN instance segmentation network to separate connected fruits and output independent fruit masks. Count the number of masks as... Nf ; d) Based on the leaf area mask, extract a 100×100 pixel ROI region at its center. Within the ROI region, convert the RGB values ​​to CIE 1976 L according to the CIED65 standard light source model. a b * Color space, for each pixel according to the formula G = (2.5 × a ∗ - b ∗ ) / 100 to calculate the greenness value G .

[0009] Preferably, in step d), based on the leaf region mask, multiple 100×100 pixel ROI regions are randomly selected, and the greenness values ​​calculated for each ROI region are averaged to obtain the leaf greenness index. G .

[0010] Preferably, an exception handling rule is added in step S2: When all conditions are not met, the current growth stage will be... S Set to the most recent successfully determined historical growth stage; if the system has no historical records on its first run, then the default will be used. S It is the budding stage.

[0011] Preferably, a dynamic environmental compensation mechanism is added between steps S1 and S2, which includes: S1.1 Real-time acquisition of ambient light intensity L and air humidity H ; S1.2, if L <10 klux or H >85%, trigger compensation algorithm: Regression model fitted based on historical database Nb'=k 1 × Nb × e^ 0.02L and Nf'=k 2 × Nf ×(1-0.005H), correcting the number of flower buds. Nb and the number of fruits Nf ,in k 1 , k 2 Variety calibration coefficient; Correction value Nb' , Nf' instead of originalNb , Nf Input the decision process for step S2.

[0012] Preferably, a soil moisture feedback closed-loop control is inserted between steps S3 and S4, which includes: S3.1. The average volumetric water content at a depth of 20cm is obtained in real time by using a soil moisture sensor array buried in the root system of fruit trees. i ; S3.2, if i < i slow (S) ,in, i slow (S) For the current growth stage S The drought warning threshold is based on the amount of emergency irrigation. Q e =min( Q max , Q ×[1+0.2( i slow (S) -θ)]) covers the original target irrigation volume range and simultaneously shortens the irrigation interval to 6 hours.

[0013] Preferably, a multi-source data consistency fusion mechanism is added between steps S2 and S3, including: S2.1 Real-time acquisition of ambient light intensity L air humidity H and the corrected number of flower buds Nb' Quantity of fruits Nf' At the same time, soil moisture content was obtained. i And the results of the current growth stage determination S ; S2.2 Calculate the data conflict index: DCI=ω 1 × ΔN+ω 2 × Dth ,in, oh 1. oh 2 Let be the weighting coefficient, satisfying oh 1+ oh 2=1, and dynamically calibrated using historical environmental and soil datasets, Δ N This corrects for the relative difference between the number of flower buds and fruits; the calculation formula is as follows: Δ i It is the relative deviation of soil moisture content from the ideal value, and the calculation formula is: ; S2.3, if DCI ≥0.5 and Dth > 0.2 × i hist ， in, i hist Based on historical averages, generate confidence growth stages. S' : like Dth > ΔN According to i Value matching growth phase: when i << 0.15 hours S' The budding stage; when i When ∈ [0.15, 0.25] S' Flowering season; when i When ∈ [0.25, 0.35] S' This is the fruit growth period; when i >0.35 S' Maturity stage; Otherwise, maintain S' = S ; Will S' Output to step S3; S2.4, if DCI <0.5, maintaining the original growth stage S Enter step S3.

[0014] Preferably, a dynamic real-time weight calibration mechanism is added in step S2.2, including: S2.2.1 Calculate the environmental credibility score EC based on ambient light intensity L and air humidity H: Where EC∈[0,1], EC=1 indicates no environmental interference, and EC=0 indicates strong interference; S2.2.2, Based on soil moisture content i Calculate the soil credibility score (SC): ,in i avg This represents the average soil moisture content over the past 24 hours. i std Its standard deviation; if the data is less than 24 hours, then i avg = i slow (S) , i std =0.1; S2.2.3, Dynamic calibration weighting coefficients: ω1=(EC+ε) / (EC+SC+2ε)×0.7, ω2=(SC+ε) / (EC+SC+2ε)×0.3, where e =0.01, which is a constant to prevent division by zero; S2.2.4, After calibration oh 1 , oh 2 enter DCI Calculation formula.

[0015] An adaptive irrigation and fertilization control system based on the mango growth stage is provided, including an image acquisition device, an image processing unit, a control unit, an irrigation system, and a fertilization device; An image acquisition device used to acquire images of the mango tree canopy once a day; An image processing unit, connected to an image acquisition device, is configured to perform the following operations: The pre-trained DeepLabv3+ segmentation network was used to divide the RGB image of the tree canopy into masked regions of flower buds, fruit, and leaves. Based on the bud region mask, the pixels within the region are converted to the HSV color space, and the dual-threshold conditional hue component H is applied. hsv Binarizing the image with saturation components Sat ∈ [0°, 15°] and saturation components Sat ∈ [0.4, 1.0] generates a primary binary image of flower buds. After performing a 3×3 elliptic kernel morphological closing operation on the primary binary image to eliminate noise and holes, a connected component labeling map is generated. Connected components with a pixel area greater than 50 are selected from the connected component labeling map, and their number is used as the number of flower buds. Nb ; Based on the fruit region mask, a morphological opening operation was performed using a 5×5 circular kernel to separate the adhered fruit regions. The minimum circumcircle of each connected region was extracted from the separated regions. The diameter of the circumcircle was converted to a physical size according to the camera calibration parameters. Connected regions with a physical diameter greater than 5 mm were selected, and their number was used as the fruit count. Nf ; Based on the leaf area mask, multiple 100×100 pixel Regions of Interest (ROIs) are randomly selected. Within each ROI, the RGB values ​​are converted to the CIE1976Lab color space according to the CIE D65 standard light source model. For each pixel, the formula is applied... G = (2.5 × a ∗ - b ∗ The greenness value is calculated by dividing the greenness value by 100, and the average of the greenness values ​​of all ROI areas is taken as the leaf greenness index. G ; The control unit, connected to the image processing unit, is configured to perform the following operations: Based on extraction Nb , Nf and G And determine the current growth stage S : like Nb ≥6 and Nf When =0, the current growth stage is determined. S Flowering season; like Nb ≤5 and Nf If ≥1, then the current growth stage is determined. S This is the fruit growth period; like Nb ≥6 and Nf ≥1, calculate the transition period score TS=0.6×( Nb / 30)+0.4×( Nf / 50), when TS≥0.5, the current growth stage is determined. S This is the transitional period between flowering and fruiting; If none of the above conditions are met and G ∈[0.4,0.7], determine the current growth stage. S Maturity stage; Otherwise, determine the current growth stage. S The budding stage; When all conditions are not met, the current growth stage will be... S Set to the most recent successfully determined historical growth stage; if the system has no historical records on its first run, then the default will be used. S The budding stage; Based on the current growth stage S The target irrigation volume range is calculated using a water and fertilizer decision function model. Q min ( S ), Q max ( S )] and target fertilizer concentration range[ C min ( S ), C max ( S )]: Q min ( S ) =α s × K Q , Q max ( S )=β s ×K Q , C min ( S )=γ s × K C , C max ( S )=δ s × K C ,in, K Q As the baseline irrigation amount, K C As the baseline concentration, α S , β S , c S , d S To be related to the growth stage S Correlation coefficient; The irrigation system, connected to the control unit, is configured to output... Q min ( S )to Q max ( S Irrigation volume within the specified range; The fertilizer applicator, connected to the control unit, is configured to output... C min ( S )to C max ( S The concentration of fertilizer solution within the range of ).

[0016] The present invention has at least the following beneficial effects: This invention utilizes a quantitative growth stage determination mechanism (flowering-fruiting transition period score TS≥0.5) and dynamic linkage with water and fertilizer decision-making to precisely match the needs of mangoes at each growth stage. Specifically, it addresses the problem of over-irrigation during the budding stage (saving 21% of water) while simultaneously ensuring nutrient supply during the flowering and fruiting stages, avoiding the inhibition of young fruit enlargement caused by traditional fixed irrigation. Field experiments show a 14% increase in yield per plant and a 30% improvement in resource utilization.

[0017] This invention effectively eliminates false identification of flower buds caused by light interference by combining HSV dual threshold filtering (hue 0-15° + saturation 0.4-1.0) with 3×3 elliptic kernel morphological closing operation, reducing the false detection rate from 18.7% to 4.2%. The Mask R-CNN deep learning model breaks through the limitations of traditional morphology and achieves accurate segmentation of fruits with overlap >50%, reducing the false recognition rate of adhesion from 22.3% to 3.1%, providing a reliable data foundation for determining the growth stage.

[0018] This invention overcomes the interference of localized reflection / shadow on leaves by employing a multi-region random sampling strategy (3-5 100×100 pixel ROIs). Under strong light, the fluctuation range of greenness value is narrowed from 0.41-0.72 to 0.48-0.61 (normal range 0.4-0.7), and the misjudgment rate of maturity is reduced from 23.3% to 6.7%. This design particularly ensures the accurate determination of the critical state (G=0.4-0.7), avoiding mismatch of water and fertilizer strategies.

[0019] This invention maintains decision-making continuity under extreme conditions such as continuous rain through a historical data inheritance mechanism: by calling the most recent 7 days of valid growth stage records, the irrigation interruption rate during the flowering period is reduced from 61.2% to 8.3%. The initial run is designed with the germination period as the default to prevent the system from running idle, and the fluctuation range of water and fertilizer under abnormal operating conditions is controlled within ±12%.

[0020] This invention addresses optical interference specifically through an environmental compensation model: an exponential correction formula. Nb'=k 1 × Nb × e ^ 0.02L The problem of reduced petal reflection under low light conditions (<10 klux) was compensated for, and the bud recognition error was improved from -22.6% to +1.8%; linear model Nf'=k 2 × Nf ×(1-0.005H) suppresses the blurring of fruit edges caused by high humidity (>85%), with a counting deviation of ≤3.5%, ensuring data reliability under severe weather conditions.

[0021] This invention achieves rapid drought response through a closed-loop soil moisture feedback system: a 20cm depth soil sensor array monitors moisture content in real time. i ,when i Emergency irrigation amount when below the stage threshold (if the actual growth period θ < 0.22) Q e By increasing the water deficit ratio (up to 20%), the response delay was reduced from 24 hours to less than 2 hours, the fruit cracking rate decreased from 18.5% to 3.1%, and the decline in fruit enlargement rate narrowed to 8.7%.

[0022] This invention resolves data inconsistencies under environmental interference by combining a multi-source conflict detection mechanism (DCI≥0.5) with soil-dominant decision rules: in strong light / high humidity scenarios, water content is prioritized. i Matching growth stages (e.g.) i >0.25 (forced to be set to the fruit growth period), the misjudgment rate dropped from 40% to 8.5%. This mechanism is automatically activated in 63.8% of high-conflict scenarios, improving system robustness.

[0023] This invention overcomes the limitations of fixed weights through a dynamic weight calibration model: based on environmental confidence EC (light-humidity coupled assessment) and soil confidence SC (24-hour data stability), it automatically adjusts image weights during rainstorms. oh 1 The soil weight decreased from 0.7 to 0.3. oh 2 The error rate rose to 0.7. The misjudgment rate for sudden weather events improved significantly from 38.7%, and smooth decision-making transitions were achieved in 87% of sudden weather events.

[0024] This invention achieves cross-module collaboration through system integration design: the image acquisition and processing unit ensures daily dynamic monitoring of flowers and fruits; the control unit integrates growth stage determination, environmental compensation, and soil moisture feedback algorithms; and the actuator coordinates irrigation and fertilization.

[0025] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0026] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0027] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0028] This invention provides a water and fertilizer adaptive irrigation control method based on the mango growth stage, comprising the following steps: S1. Collect images of the mango tree canopy once a day using an image acquisition device, and process the canopy images using an image processing unit to extract the number of flower buds. Nb Quantity of fruits Nf and leaf greenness index G ; S2, based on extraction Nb , Nf and G And determine the current growth stage S : like Nb ≥6 and Nf When =0, the current growth stage is determined. S Flowering season; like Nb ≤5 and Nf If ≥1, then the current growth stage is determined. S This is the fruit growth period; like Nb ≥6 and Nf ≥1, calculate the transition period score TS=0.6×( Nb / 30)+0.4×( Nf / 50), when TS≥0.5, the current growth stage is determined. S This is the transitional period between flowering and fruiting; If none of the above conditions are met and G ∈[0.4,0.7], determine the current growth stage. S Maturity stage; Otherwise, determine the current growth stage. S The budding stage; S3. Growth stage determined in step S2 S The target irrigation volume range is calculated using a water and fertilizer decision function model. Q min ( S ), Q max ( S )] and target fertilizer concentration range[ C min ( S ), C max ( S )]: Q min ( S ) =α s × K Q , Q max ( S )=β s × K Q , C min ( S )=γ s × K C , C max ( S )=δ s × K C ,in, Q For irrigation volume,C This refers to the concentration of the fertilizer solution. K Q As the baseline irrigation amount, K C As the baseline concentration, α S , β S , c S , d S To be related to the growth stage S The relevant coefficients are obtained by looking up a preset growth coefficient table; S4. Control the irrigation system's output irrigation volume. Q ∈[ Q min , Q max Simultaneously control the output concentration of the fertilizer application device. C ∈[ C min ( S ), C max ( S )]; Specifically, the image acquisition device uses a regular RGB camera that can be installed 1.5 meters above the tree canopy to collect images of the mango tree canopy at regular intervals every day. The image processing unit uses a pre-trained DeepLabv3+ network to segment the regions of flower buds, fruits, and leaves using masks. Flower bud identification: Convert the flower bud mask pixels to the HSV color space, filter pixels with a hue component of 0 to 15 degrees and a saturation of 0.4 to 1.0, generate a binary image, and then perform denoising using a 3×3 elliptical kernel closure operation. Count the number of connected components with a pixel area greater than 50 as the denoising result. Nb ; Fruit counting: Perform a 5×5 circular kernel opening operation on the fruit mask to separate the adhered regions, and extract the number of circumscribed connected regions with a physical diameter greater than 5 mm as the count. Nf ; Leaf greenness: Three 100×100 pixel leaf ROI areas were randomly selected and converted from RGB to Lab color according to the CIE D65 standard. * Color space, calculate the greenness value of each pixel G =(2.5 a ∗ - b ∗ () / 100 and take the average; Based on collection Nb , Nf , G Data execution growth phase S Classification: satisfy Nb ≥6 and Nf When the value is 0, it is marked as the flowering period; when Nb ≥6 and Nf When ≥1, calculate the transition period score TS=0.6×( Nb / 30)+0.4×( Nf / 50), when TS≥0.5, it is determined to be the transition period. Among them, 30 / 50 is the experimental value of Tainong No. 1, and the actual value needs to be calibrated according to the variety. If the aforementioned conditions are not met and the leaf greenness is... G A value between 0.4 and 0.7 is considered to be in the mature stage. Otherwise, it is classified as the budding stage. In case of abnormal situations, the most recent historical growth stage will be called. The first run defaults to the budding stage. Based on the judgment result S, the target irrigation volume range is determined by consulting the preset coefficient table. α s × K Q ,β s × K Q ] and fertilizer solution concentration range [γ s × K C ,δ s × K C (Example: Flowering period coefficient) α S =0.8, β S =1.0, c S =0.6, d S =0.8); The irrigation system uses drip irrigation tape laid 20 cm deep around the roots of the fruit trees, and the fertilization device uses a Venturi fertilizer pump. When it is determined to be the flowering period (e.g.) Nb =8, Nf =0), based on the baseline irrigation amount K Q =10 liters / plant·day, baseline concentration K C Based on a concentration of 1.0 g / L, the output irrigation volume is 8 to 10 liters per plant, and the fertilizer solution concentration is 0.6 to 0.8 g / L. The control unit adjusts the drip irrigation flow rate and fertilizer injection pump proportional valve in real time to ensure that the actual irrigation volume (e.g., 9 liters) and fertilizer concentration (e.g., 0.7 g / L) fall within the target range.

[0029] Under the same orchard conditions, this proposed solution was compared with a traditional timed irrigation system. The traditional system uses a fixed irrigation mode with a daily irrigation volume of 12 liters per tree and a fertilizer concentration of 0.9 grams per liter. Experimental data showed that the total irrigation volume of this proposed solution was 850 cubic meters, significantly lower than the 1080 cubic meters of the traditional system; the total fertilizer usage was 68 kg, a substantial reduction compared to the 97 kg of the traditional system; and in terms of yield, the average yield per tree in this proposed solution reached 25.3 kg, higher than the 22.1 kg of the traditional system. The results indicate that by adaptively adjusting water and fertilizer supply during the growth stages, this proposed solution achieves resource conservation while ensuring yield.

[0030] Traditional timed and quantitative irrigation systems ignore the dynamic growth needs of fruit trees, leading to wasted water and fertilizer during the budding stage or insufficient nutrition during the flowering and fruiting stage. This solution uses image recognition to quantify the characteristics of growth stages and establish a dynamic mapping relationship between growth stages and water and fertilizer parameters, thus solving the core problem of imbalanced ratios during transitional periods (such as when flowers and fruits coexist).

[0031] In another embodiment of the present invention, step S1, in which the image processing unit processes the tree canopy image, includes the following steps: a) The tree canopy RGB image is divided into flower bud region mask, fruit region mask and leaf region mask by a pre-trained DeepLabv3+ segmentation network; b) Based on the bud region mask, convert the pixels within the region to the HSV color space, and apply the dual-threshold conditional hue component H. hsv Binarize the image with saturation components Sat ∈ [0°, 15°] and saturation components Sat ∈ [0.4, 1.0] to generate a primary binary image of flower buds. Perform a 3×3 elliptic kernel morphological closing operation on the primary binary image to remove noise and holes, and then generate a connected component labeling map. In the connected component labeling map, filter connected components with a pixel area > 50; the number of these components represents the number of valid flower buds. Nb ; c) Based on the fruit region masks, input the pre-trained Mask R-CNN instance segmentation network to separate connected fruits and output independent fruit masks. Count the number of masks as... Nf ; d) Based on the leaf area mask, extract a 100×100 pixel ROI region at its center. Within the ROI region, convert the RGB values ​​to CIE 1976 L according to the CIED65 standard light source model. a b * Color space, for each pixel according to the formula G = (2.5 × a ∗ - b ∗ ) / 100 to calculate the greenness value G ; This embodiment uses a pre-trained DeepLabv3+ segmentation network to process tree canopy RGB images, generating region masks for flower buds, fruits, and leaves. The pre-trained model is trained on 10,000 labeled images, with a segmentation accuracy of mIoU=0.89 on the test set and an inference time of <1s (NVIDIA Jetson TX2). In the flower bud recognition stage, the mask pixels are converted to the HSV color space, and binarization is performed using dual thresholds of hue component 0 to 15 degrees and saturation 0.4 to 1.0. Then, noise is eliminated by a 3×3 elliptic kernel closure operation. Finally, connected regions with a pixel area greater than 50 are selected as the number of flower buds. Nb Fruit counting uses the Mask R-CNN model to separate the adhered regions and output the number of independent masks. Nf Leaf greenness detection was performed within three randomly selected 100×100 pixel areas on the leaf. RGB values ​​were converted to the Lab* color space using a CIE D65 standard light source. The greenness value of each pixel was calculated according to the formula, and the average value was taken as the result. G ; In a comparative experiment on 100 mango trees, this scheme significantly improved the recognition accuracy compared to the traditional RGB thresholding method: the false negative rate of flower buds decreased from 18.7% to 4.2% and the false positive rate of fruit adhesion decreased from 22.3% to 3.1%; the processing time was stabilized at 0.8 seconds, which is slightly higher than the traditional method's 0.3 seconds, but the overall accuracy meets the real-time control requirements of the orchard; a typical case is that in a certain detection, 8 valid flower buds were accurately identified (3 noise points were filtered out by closing operation) and 5 independent fruits (Mask R-CNN separated 2 groups of adhesion).

[0032] Compared to existing technologies (such as single RGB threshold segmentation), this solution addresses core pain points through a three-tiered design: 1) HSV dual thresholding combined with morphological filtering effectively suppresses light interference, avoiding misjudgments caused by twilight red light; 2) The Mask R-CNN deep learning model overcomes the limitations of traditional morphology, accurately separating fruits with more than 50% overlap; 3) Multi-region random sampling eliminates local leaf shading bias, improving the representativeness of the greenness index. This design ensures reliable flower and fruit growth parameters are obtained in complex orchard environments.

[0033] In another embodiment of the present invention, in step d), based on the leaf region mask, multiple 100×100 pixel ROI regions are randomly selected, and the greenness values ​​calculated for each ROI region are averaged to obtain the leaf greenness index. G ; In this implementation, a multi-region random sampling strategy is adopted in the leaf greenness detection stage: within the leaf mask generated by image segmentation, the system automatically selects 3 to 5 rectangular regions of 100×100 pixels (3 by default) to cover different directions of the tree canopy. For each region, the RGB values ​​are converted to LabVIEW according to the CIE D65 standard light source model.* In a color space, the greenness value is calculated pixel by pixel, and the average value of that region is taken. Finally, the arithmetic mean of the results for all regions is calculated as the leaf greenness index. G The process is executed by an embedded image processing unit. The camera is vertically mounted on a bracket 1.5 meters above the tree canopy. In a typical case, during a certain test, the system collected data from the sunny side (greenness 0.58), the shady side (0.49), and the leaf edge area (0.52), and finally output a balance value of 0.53. In a comparative test on 60 mango trees, the multi-area sampling method significantly reduced the impact of environmental disturbances. Under strong light conditions, the traditional single-point sampling method resulted in a greenness fluctuation range of 0.41 to 0.72, leading to a 23.3% misjudgment of growth stages; this method reduced the fluctuation range to 0.48 to 0.61, lowering the misjudgment rate to 6.7%. It was also effective under low light conditions: single-point sampling showed a fluctuation range of 0.36 to 0.68 (misjudgment rate 18.9%), while this method showed a fluctuation range of 0.45 to 0.57 (misjudgment rate 5.1%), demonstrating that this method can stably adapt to changes in light intensity.

[0034] Existing single-point sampling techniques, which fix the central area of ​​the leaf, are easily affected by reflected light in strong light (resulting in an artificially high greenness of 0.72) and by shadows in weak light (resulting in a low greenness of 0.36), often leading to misjudgments of the budding and maturity stages. This scheme uses spatial random sampling to offset local biases, particularly ensuring the reliability of maturity stage determination (greenness threshold of 0.4 to 0.7), thus solving the core problem of traditional methods being greatly affected by changes in ambient light.

[0035] In another embodiment of the present invention, an exception handling rule is added in step S2: When all conditions are not met, the current growth stage will be... S Set to the most recent successfully determined historical growth stage; if the system has no historical records on its first run, then the default will be used. S The budding stage; This implementation method adds an anomaly handling rule to address situations where the growth stage determination fails. This rule applies when the number of flower buds, the number of fruits, and the greenness of leaves all fail to meet the conditions (such as...). Nb =3 and Nf =0 and G When the value is 0.8, the system automatically activates the exception handling mechanism: First, it retrieves the most recently successfully determined historical growth stage record (such as flowering data stored 7 days ago), which is stored in the non-volatile memory of the embedded device; if the system runs for the first time without historical data (such as when a new device is installed), the default setting is the budding stage. A typical example is when continuous rain causes the image to become blurry, the system maintains the irrigation strategy based on the flowering record from 5 days ago to avoid decision interruption.

[0036] In a test on 80 mango trees, this approach significantly improved stability compared to the traditional system without an anomaly handling mechanism. The error rate in the anomaly handling group was 8.3%, with daily irrigation fluctuations controlled within ±12%; while the error rate in the untreated control group reached 61.2%, with irrigation fluctuations as high as ±47%. Specifically, the traditional system forcibly reverted to the budding stage (low water and fertilizer) when flowering was abnormal, leading to flower wilting; while this approach, by inheriting historical stages, ensured the continuity of water and fertilizer supply during the flowering and fruiting periods.

[0037] Existing technologies use fixed default values ​​(always set to the budding period), which has significant drawbacks: interrupting irrigation during abnormal flowering periods causes flower drop, and incorrectly setting the budding period during maturity results in wasted water and fertilizer. This solution dynamically calls the most recent reliable historical data, which not only solves the problem of detection failure caused by temporary environmental interference (heavy rain, strong light) but also ensures the baseline supply for the first run, achieving a smooth transition of water and fertilizer strategies under abnormal conditions.

[0038] In another embodiment of the present invention, a dynamic environmental compensation mechanism is added between steps S1 and S2, which includes: S1.1 Real-time acquisition of ambient light intensity L and air humidity H ; S1.2, if L <10 klux or H >85%, trigger compensation algorithm: Regression model fitted based on historical database Nb'=k 1 × Nb × e^ 0.02L and Nf'=k 2 × Nf ×(1-0.005H), correcting the number of flower buds. Nb and the number of fruits Nf ,in k 1 , k 2 Variety calibration coefficient; Correction value Nb' , Nf' instead of originalNb , Nf Input the decision process for step S2; This implementation specifically includes deploying light sensors and air humidity sensors in the orchard: the light sensor can be a silicon photodiode type device, installed at a height of 1.2 meters above the tree canopy; the humidity sensor can be a capacitive probe, fixed in the shaded area of ​​the side branches of the fruit tree. When the ambient light intensity is detected to be below 10 kilolux or the air humidity is above 85%, the system automatically triggers a compensation algorithm. This algorithm is based on a regression model established from historical data, which identifies the number of flower buds in the original image. Nb and number of fruits Nf Correction was made: the flower bud correction adopted an exponential model. Nb'=k 1 × Nb × e^ 0.02L Fruit correction uses a linear model Nf'=k 2 × Nf ×(1-0.005H), where the variety calibration coefficient is... k 1 , k 2 Through localized trials (such as Tainong No. 1) k 1 =1.05), corrected Nb' , Nf' Replace the original value input with the subsequent growth stage determination process.

[0039] A typical case study is a period of continuous rainy weather (light intensity 8 kilolux, humidity 92%): Original image identification of flower bud count Nb =12, number of fruits Nf =4; Environmental parameters trigger a compensation mechanism to calculate correction values: Nb =1.05×12× e 0.02 × 8 ≈ 14.8 Nf =0.98×4×(1−0.005×92)≈3.2; Input the correction value into the growth stage determination module.

[0040] To verify the effect, a comparative test was conducted on 80 mango trees: under simulated continuous rainy weather, the uncompensated group had a flower bud recognition value of 9.3 (deviation -22.6%) and a fruit recognition value of 3.2 (deviation -18.9%) due to image blurring, resulting in 6 misjudgments; the compensated group controlled the flower bud recognition deviation to +1.8% and the fruit deviation to -3.5% through environmental correction, with only 1 misjudgment.

[0041] Traditional methods interrupt decision-making during periods of continuous rain, leading to a lack of water and fertilizer supply during the flowering period. This solution addresses two core issues by establishing quantitative compensation models based on light intensity, bud index, and humidity, and the number of fruits: firstly, the missed detection of buds due to reduced petal reflection under low light conditions; and secondly, the blurring of fruit edges caused by high humidity. This mechanism ensures data reliability while maintaining the continuity of water and fertilizer decisions, making it particularly suitable for the management needs of orchards in rainy areas.

[0042] In another embodiment of the present invention, a soil moisture feedback closed-loop control is inserted between steps S3 and S4, which includes: S3.1. The average volumetric water content at a depth of 20cm is obtained in real time by using a soil moisture sensor array buried in the root system of fruit trees. i ; S3.2, if i < i slow (S) ,in, i slow (S) For the current growth stage S The drought warning threshold is based on the amount of emergency irrigation. Q e =min(Q max ,Q×[1+0.2( i slow (S) -θ)]) covers the original target irrigation volume range and simultaneously shortens the irrigation interval to 6 hours; The soil moisture feedback mechanism in this embodiment is implemented by burying a capacitive soil moisture sensor array 20 cm deep in the root system of fruit trees. The sensors are arranged radially with a spacing of 50 cm, and the data is transmitted to the control unit via LoRa. The system presets drought thresholds for each growth stage: 0.12 for budding, 0.18 for flowering, 0.22 for fruit development, and 0.15 for ripening. The real-time average moisture content... i When the water level falls below the threshold, an emergency irrigation calculation is triggered: the minimum value between the target upper limit and the original irrigation amount increased proportionally to the water deficit is taken (e.g., during the fruit growth period). i When the value is 0.19, calculate the incremental increase of 20% × 0.03 / 0.22, and at the same time shorten the irrigation interval to 6 hours.

[0043] During the peak summer fruit growing season, a continuous drought occurred, and the system monitored a moisture content drop to 0.19 (below the threshold of 0.22). An emergency irrigation of 12.72 liters per plant was immediately administered (the original maximum was 12 liters), followed by irrigation every 6 hours. Comparative experiments showed that the soil moisture feedback group had a response delay of less than 2 hours under drought conditions, with a fruit cracking rate of only 3.1% and a decrease in fruit enlargement rate of 8.7%. In contrast, the no-feedback group had a delay of 24 hours, with a fruit cracking rate rising to 18.5% and a decrease in fruit enlargement rate of 34.2%.

[0044] Traditional irrigation systems continue to irrigate at fixed intervals during drought periods, leading to persistent soil water shortages during the fruit's growth period, which in turn causes fruit cracking when sudden rainfall occurs later. This solution addresses two major issues through real-time soil moisture monitoring and dynamic irrigation adjustments: firstly, it provides a rapid response to soil moisture deficits, preventing prolonged drought stress; secondly, it prevents drastic changes in moisture levels through high-frequency, low-volume irrigation, ensuring stable fruit development.

[0045] In another embodiment of the present invention, a multi-source data consistency fusion mechanism is added between steps S2 and S3, including: S2.1 Real-time acquisition of ambient light intensity L air humidity H and the corrected number of flower buds Nb' Quantity of fruits Nf' At the same time, soil moisture content was obtained. i And the results of the current growth stage determination S ; S2.2 Calculate the data conflict index: DCI=ω 1 × ΔN+ω 2 × Dth ,in, oh 1. oh 2 Let be the weighting coefficient, satisfying oh 1+ oh 2=1, and dynamically calibrated using historical environmental and soil datasets, Δ N This corrects for the relative difference between the number of flower buds and fruits; the calculation formula is as follows: Δ i It is the relative deviation of soil moisture content from the ideal value, and the calculation formula is: ; S2.3, if DCI ≥0.5 and Dth > 0.2 × i hist ， in, i hist Based on historical averages, generate confidence growth stages. S' : like Dth > ΔN According to i Value matching growth phase: when i << 0.15 hours S' The budding stage; when i When ∈ [0.15, 0.25] S' Flowering season; when i When ∈ [0.25, 0.35] S' This is the fruit growth period; when i >0.35 S' Maturity stage; Otherwise, maintain S' = S ; WillS' Output to step S3; S2.4, if DCI <0.5, maintaining the original growth stage S Enter step S3; The source data consistency fusion mechanism in this embodiment is implemented through a systematic architecture. Regarding equipment configuration, the ambient light sensor can be a silicon photodiode type illuminance meter, installed on a stainless steel bracket 1.5 meters above the tree canopy, with a measurement range covering 0 to 100 kilolux; the air humidity sensor can be a capacitive probe with an accuracy of ±2%, fixed on the shady side of the main branch of the fruit tree at a height of 1.2 meters above the ground; the soil moisture sensor array consists of four sets of capacitive probes, buried in a cross shape within the drip line of the fruit tree, with a depth strictly controlled at 20 ± 0.5 cm, and connected to the control unit via an RS485 bus. At the data processing level, a conflict detection model is established: dynamically calculating the relative difference in the number of flowers and fruits ΔN (the absolute difference between the corrected value and the original value divided by the maximum of the two) and the relative deviation of soil moisture content. Dth (The absolute difference between the real-time value and the current growth stage threshold divided by the threshold) is weighted and fused into the Data Conflict Index (DCI) (initial weights). oh 1 = 0.7 oh 2 =0.3). The threshold for triggering re-judgment is set to DCI≥0.5 and Dth >0.2 times the historical average, if the soil deviation is at this point Dth If the difference in the number of flowers and fruits is greater than ΔN, then the moisture content should be used directly. i Rematching growth stages: when the moisture content is below 0.15, it is set as the budding stage (such as in sandy soil under continuous drought conditions); 0.15 to 0.25 is the flowering stage (the best flowering conditions for loam); 0.25 to 0.35 is the fruit growth stage (for clay soil with good water retention conditions); and above 0.35 is the ripening stage (after rain when the soil is saturated).

[0046] The typical working process under continuous rainy conditions is as follows: First, the image system was affected by rain and fog, resulting in a reduced number of flower buds after correction. Nb' Only 4 were identified (there should actually be 12), fruit count Nf' The value is 0, indicating the initial assessment is the germination stage; simultaneously, the soil sensor reports an average moisture content of 0. i =0.28, significantly higher than the germination stage threshold of 0.15; The system automatically performs multi-source fusion: calculates the difference in the number of flowers and fruits ΔN=|(4+0)-(11+3)| / max(14,1)=0.71 (based on historical database correction value), and soil moisture deviation. Dth=|0.28-0.15| / 0.15=0.87; Combining the environmental confidence level EC=0.3 (due to light intensity of 6klux and humidity of 93% significantly deviating from the ideal range) and the soil confidence level SC=0.9 (24-hour data standard deviation of 0.02), the dynamic calibration weight is... oh 1 = 0.25 oh 2 =0.75, the final DCI = 0.25 × 0.71 + 0.75 × 0.87 = 0.82. This is because the re-judgment condition was triggered and... Dth >ΔN, system based i =0.28 corrects the growth stage to the fruit growth period. In a rainy season test on 200 Tainong mango trees in the Sanya monsoon region of Hainan (average daily rainfall of 15 mm), the traditional image recognition system had an average misjudgment rate of 46.2% due to continuous rain. This solution, through a fusion mechanism, initiated soil data-driven correction in 63.8% of high-conflict scenarios, reducing the overall misjudgment rate to 13.5%, with the misjudgment rate during the critical flowering and fruiting period improving from 51.7% to 9.3%, while simultaneously achieving water savings of 23%.

[0047] This mechanism overcomes three technical bottlenecks: First, traditional systems suffer from flower and fruit recognition errors exceeding 40% under strong light and high humidity conditions. This solution addresses image feature distortion caused by continuous rain by using a re-judgment rule dominated by soil moisture content (e.g., forcibly setting θ>0.25 as the fruit growth period). Second, while existing technologies employ fixed weights or simple majority voting, this mechanism pioneers a dynamic weight allocation model based on environmental credibility EC (dual-parameter evaluation of light and humidity) and soil credibility SC (historical data stability analysis), automatically reducing image weights to 0.25 under high humidity conditions. Third, it adds threshold gradients for different soil types (e.g., reducing the threshold for the germination period in sandy soil from 0.15 to 0.10) to eliminate decision-making errors caused by mismatches in water and soil conditions. Practical application shows that this mechanism intercepted 127 abnormal irrigation operations during a three-month rainy season test, preventing flower and fruit drop losses due to misjudgment, and is particularly suitable for the complex planting environment of tropical rainy regions.

[0048] In another embodiment of the present invention, a dynamic real-time weight calibration mechanism is added in step S2.2, including: S2.2.1 Calculate the environmental credibility score EC based on ambient light intensity L and air humidity H: Where EC∈[0,1], EC=1 indicates no environmental interference, and EC=0 indicates strong interference; S2.2.2, Based on soil moisture content i Calculate the soil credibility score (SC): ,in iavg This represents the average soil moisture content over the past 24 hours. i std Its standard deviation; if the data is less than 24 hours, then i avg = i slow (S) , i std =0.1; S2.2.3, Dynamic calibration weighting coefficients: ω1=(EC+ε) / (EC+SC+2ε)×0.7, ω2=(SC+ε) / (EC+SC+2ε)×0.3, where e =0.01, which is a constant to prevent division by zero; S2.2.4, After calibration oh 1 , oh 2 enter DCI Calculation formula; The dynamic weight calibration mechanism in this embodiment is achieved by evaluating data reliability in real time. The environmental reliability EC is calculated based on light intensity L and air humidity H: when the light intensity is between 5 and 70 klux and the humidity is between 40% and 85%, the EC score is 1.0; otherwise, it is calculated using the formula EC = 0.5 × (1 - max(|L - 37.5|, |H - 62.5|) / 50) (e.g., EC = 0.32 when the light intensity is 90 klux and the humidity is 30%). The soil reliability SC is based on the soil moisture content. i Stability: If the current i Value and 24-hour average i avg The deviation does not exceed the standard deviation i std (Default when data is insufficient) i std If θ = 0.1), then SC = 1.0; otherwise, SC = exp(-|θ- i avg | / (2 i std )) Calculate (e.g., θ = 0.22, i avg =0.25, i std When SC=0.67 (when SC=0.03), the dynamic weighting coefficient is calculated according to... oh 1 =(EC+0.01) / (EC+SC+0.02)×0.7, oh 2=(SC+0.01) / (EC+SC+0.02)×0.3 for calibration, with zero constant ε=0.01. The calibration algorithm runs on the main processor of the control unit. Typical sudden weather scenario (sunny to rainy): Initial environment: light intensity 65 klux (ideal), humidity 50%, EC=1.0; Soil data: θ=0.24, 24-hour mean i avg =0.23, standard deviation i std =0.02, SC=0.88; Initial weights ω1 = 0.70 × 0.7 = 0.49, ω2 = 0.88 × 0.3 = 0.26 (after normalization ω1 = 0.65, ω2 = 0.35); Sudden downpour: Sunlight drops to 8 klux, humidity rises to 92%, EC = 0.5 × (1 - max(|8 - 37.5|, |92 - 62.5|) / 50) = 0.15, soil θ = 0.29 (delayed rainwater infiltration), SC remains at 0.85; Dynamic weight calibration: ω1=(0.15+0.01) / (0.15+0.85+0.02)×0.7≈0.11, ω2=(0.85+0.01) / 1.02×0.3≈0.25 (after normalization ω1=0.31, ω2=0.69); During the typhoon season test in Hainan, the traditional fixed-weight scheme (ω1=0.7, ω2=0.3) had a misjudgment rate of 38.7% under sudden weather changes. This scheme significantly reduced the misjudgment rate through dynamic calibration and achieved smooth weight transition in 87% of weather change events, avoiding decision jumps caused by drastic environmental changes.

[0049] Existing technologies suffer from two drawbacks: first, they maintain a high environmental weight (70%) during the initial stages of heavy rainfall when soil data has not yet responded, leading to misjudging high humidity as the flowering period; second, soil data remains heavily weighted after the rain stops and the weather clears, causing misjudgment of the maturity period. This solution overcomes these two technical bottlenecks through a dual reliability assessment and dynamic recalibration mechanism: first, it establishes a light-humidity coupled environmental reliability model to accurately quantify the real-time reliability of sensors; second, it designs an exponential decay assessment based on soil data stability to address data distortion caused by infiltration delay. Practical application shows that this mechanism improves the continuity of system decision-making during sudden weather changes, avoiding drastic fluctuations in irrigation strategies.

[0050] In another embodiment of the present invention, a water and fertilizer adaptive irrigation control system based on the mango growth stage is provided, including an image acquisition device, an image processing unit, a control unit, an irrigation system, and a fertilizer application device. An image acquisition device used to acquire images of the mango tree canopy once a day; An image processing unit, connected to an image acquisition device, is configured to perform the following operations: The pre-trained DeepLabv3+ segmentation network was used to divide the RGB image of the tree canopy into masked regions of flower buds, fruit, and leaves. Based on the bud region mask, the pixels within the region are converted to the HSV color space, and the dual-threshold conditional hue component H is applied. hsv Binarizing the image with saturation components Sat ∈ [0°, 15°] and saturation components Sat ∈ [0.4, 1.0] generates a primary binary image of flower buds. After performing a 3×3 elliptic kernel morphological closing operation on the primary binary image to eliminate noise and holes, a connected component labeling map is generated. Connected components with a pixel area greater than 50 are selected from the connected component labeling map, and their number is used as the number of flower buds. Nb ; Based on the fruit region mask, a morphological opening operation was performed using a 5×5 circular kernel to separate the adhered fruit regions. The minimum circumcircle of each connected region was extracted from the separated regions. The diameter of the circumcircle was converted to a physical size according to the camera calibration parameters. Connected regions with a physical diameter greater than 5 mm were selected, and their number was used as the fruit count. Nf ; Based on the leaf area mask, multiple 100×100 pixel Regions of Interest (ROIs) are randomly selected. Within each ROI, the RGB values ​​are converted to the CIE1976Lab color space according to the CIE D65 standard light source model. For each pixel, the formula is applied... G = (2.5 × a ∗ - b ∗ The greenness value is calculated by dividing the greenness value by 100, and the average of the greenness values ​​of all ROI areas is taken as the leaf greenness index. G ; The control unit, connected to the image processing unit, is configured to perform the following operations: Based on extraction Nb , Nf and G And determine the current growth stage S : like Nb ≥6 and Nf When =0, the current growth stage is determined. S Flowering season; like Nb ≤5 and Nf If ≥1, then the current growth stage is determined. S This is the fruit growth period; like Nb ≥6 and Nf ≥1, calculate the transition period score TS=0.6×( Nb / 30)+0.4×( Nf / 50), when TS≥0.5, the current growth stage is determined. S This is the transitional period between flowering and fruiting; If none of the above conditions are met and G ∈[0.4,0.7], determine the current growth stage. S Maturity stage; Otherwise, determine the current growth stage. S The budding stage; When all conditions are not met, the current growth stage will be... S Set to the most recent successfully determined historical growth stage; if the system has no historical records on its first run, then the default will be used. S The budding stage; Based on the current growth stage S The target irrigation volume range is calculated using a water and fertilizer decision function model. Q min ( S ), Q max ( S )] and target fertilizer concentration range[ C min ( S ), C max ( S )]: Q min ( S ) =α s × K Q , Q max ( S )=β s × K Q , C min ( S )=γ s × K C , C max ( S )=δ s × K C ,in, K Q As the baseline irrigation amount, K C As the baseline concentration, α S , β S , cS , d S To be related to the growth stage S Correlation coefficient; The irrigation system, connected to the control unit, is configured to output... Q min ( S )to Q max ( S Irrigation volume within the specified range; The fertilizer applicator, connected to the control unit, is configured to output... C min ( S )to C max ( S Fertilizer solution concentration within the range of ) The water and fertilizer adaptive irrigation control system of this embodiment consists of five parts: the image acquisition device uses a 20-megapixel global shutter RGB camera, which is vertically fixed to an aluminum alloy bracket 1.5 meters above the tree canopy. It automatically takes images of the tree canopy at 6:00 AM every day. The image processing unit is equipped with a pre-trained DeepLabv3+ segmentation network and performs a hierarchical processing flow: first, it divides the regions of flower buds, fruits, and leaves into masks; in the flower bud recognition stage, the mask pixels are converted to the HSV color space, and dual threshold binarization with hue components of 0 to 15 degrees and saturation of 0.4 to 1.0 is applied. After noise is eliminated by 3×3 elliptic kernel closure operation, connected regions with a pixel area greater than 50 are selected as the number of flower buds. Nb Fruit counting employed the Mask R-CNN model to separate adhered regions and counted the number of independent masks with a physical diameter exceeding 5 mm. Nf Leaf greenness detection randomly selected three 100×100 pixel leaf areas, and the RGB values ​​were converted to Lab color using a CIE D65 standard light source. * The greenness index of the leaves is finally output by calculating the greenness value pixel by pixel in the color space. G The control unit uses industrial-grade edge computing equipment, installed in a waterproof control box, and has built-in growth stage determination logic (flowering period threshold Nb≥6 and...). Nf =0, Flowering and Fruiting Transition Period Score TS ≥0.5, etc.) and water and fertilizer decision-making models (benchmark irrigation amount) K Q =10 liters / plant·day, baseline concentration K C =1.0 g / L). The irrigation system uses pressure-compensated drip irrigation tape, which is laid in the soil layer 20 cm deep around the roots of the fruit trees; the fertilization device uses a Venturi fertilizer pump, which precisely controls the fertilizer concentration through a proportional valve, with an error range of ±0.1 g / L.

[0051] The typical workflow of the system is as follows: After acquiring images in the early morning, the processing unit identifies the number of flower buds. Nb =8 (2 noise points were removed by morphological filtering), number of fruits Nf =3 (Mask R-CNN separates 1 group of adherent fruits), leaf greenness G =0.52 (average of 3 areas); the control unit calculates the flowering and fruiting transition period score TS=0.6×(8 / 30)+0.4×(3 / 50)=0.52≥0.5, and determines it to be the flowering and fruiting transition period; the irrigation range is obtained by consulting the preset coefficient table [ α s × K Q ,β s × K Q =[0.9×10,1.1×10]=[9,12] liters / plant, fertilizer concentration range [γ s × K C ,δ s × K C =[0.7×1.0, 0.9×1.0]=[0.7,0.9] g / L; the drip irrigation system outputs 10 liters / plant of water-fertilizer mixture, and the fertilizer pump adjusts the concentration to 0.8 g / L. A comparative experiment was conducted in the same orchard over a complete 120-day growing season: the total irrigation volume of the system of this invention was 850 cubic meters, the fertilizer usage was 68 kg, and the average yield per plant was 25.3 kg; the corresponding data for the traditional timed irrigation system (fixed daily 12 liters / plant, 0.9 g / L) were 1080 cubic meters, 97 kg, and 22.1 kg. In fault-tolerant scenarios, when continuous rain causes image blurring, the system calls the previous day's valid historical data (flowering period records) to maintain an irrigation volume of 8-10 liters / plant, avoiding the water shortage during the flowering period caused by the traditional system downgrading to the budding stage mode (only 5-7 liters).

[0052] This system addresses core industry pain points through three key innovations: First, traditional soil moisture control systems continue to irrigate at high rates during the flowering and fruiting period, which can lead to excessive vegetative growth and inhibit fruit enlargement. This system, however, dynamically adjusts water and fertilizer parameters using a transition period scoring model (TS≥0.5) to precisely match the needs of the transition period. Second, while existing technologies directly switch to a timed mode when the image system fails, this system employs a historical data inheritance mechanism (such as calling the most recent 7 days of valid records) to ensure the continuity of decisions during critical growth periods. Third, it integrates multi-source data for collaborative resource optimization, such as using coefficients during the budding stage. α s =0.5 Reduce irrigation to 50% of the baseline value, and adjust the amount based on leaf greenness at maturity. GThe system reduces fertilizer concentration within the range of [0.4, 0.7], a synergistic mechanism lacking in traditional methods. Practical application shows that this system achieves 21% water savings (compared to traditional methods) at a planting base in Sanya, Hainan, while simultaneously increasing yield per plant, effectively solving the industry problem of water and fertilizer imbalance during the flowering and fruiting transition period.

[0053] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A water and fertilizer adaptive irrigation control method based on mango growth stages, characterized in that, Includes the following steps: S1. Collect images of the mango tree canopy once a day using an image acquisition device, and process the canopy images using an image processing unit to extract the number of flower buds. Nb Quantity of fruits Nf and leaf greenness index G ; S2, based on extraction Nb , Nf and G And determine the current growth stage S : like Nb ≥6 and Nf When =0, the current growth stage is determined. S Flowering season; like Nb ≤5 and Nf If ≥1, then the current growth stage is determined. S This is the fruit growth period; like Nb ≥6 and Nf ≥1, calculate the transition period score TS=0.6×( Nb / 30)+0.4×( Nf / 50), when TS≥0.5, the current growth stage is determined. S This is the transitional period between flowering and fruiting; If none of the above conditions are met and G ∈[0.4,0.7], determine the current growth stage. S Maturity stage; Otherwise, determine the current growth stage. S The budding stage; S3. Growth stage determined in step S2 S The target irrigation volume range is calculated using a water and fertilizer decision function model. Q min ( S ), Q max ( S )] and target fertilizer concentration range[ C min ( S ), C max ( S )]: Q min ( S ) =α s × K Q , Q max ( S )=β s × K Q , C min ( S )=γ s × K C , C max ( S )=δ s × K C ,in, Q For irrigation volume, C This refers to the concentration of the fertilizer solution. K Q As the baseline irrigation amount, K C As the baseline concentration, α S , β S , γ S , δ S To be related to the growth stage S The relevant coefficients are obtained by looking up a preset growth coefficient table; S4. Control the irrigation system's output irrigation volume. Q ∈[ Q min , Q max Simultaneously control the output concentration of the fertilizer application device. C ∈[ C min ( S ), C max ( S )).

2. The control method as described in claim 1, characterized in that, Step S1, in which the image processing unit processes the tree canopy image, includes the following steps: a) The tree canopy RGB image is divided into flower bud region mask, fruit region mask and leaf region mask by a pre-trained DeepLabv3+ segmentation network; b) Based on the bud region mask, convert the pixels within the region to the HSV color space, and apply the dual-threshold conditional hue component H. hsv Binarize the image with saturation components Sat ∈ [0°, 15°] and saturation components Sat ∈ [0.4, 1.0] to generate a primary binary image of flower buds. Perform a 3×3 elliptic kernel morphological closing operation on the primary binary image to remove noise and holes, and then generate a connected component labeling map. In the connected component labeling map, filter connected components with a pixel area > 50; the number of these components represents the number of valid flower buds. Nb ; c) Based on the fruit region masks, input the pre-trained Mask R-CNN instance segmentation network to separate connected fruits and output independent fruit masks. Count the number of masks as... Nf ; d) Based on the leaf area mask, extract a 100×100 pixel ROI region at its center. Within the ROI region, convert the RGB values ​​to CIE 1976 L according to the CIE D65 standard light source model. a b * Color space, for each pixel according to the formula G = (2.5 × a ∗ - b ∗ ) / 100 to calculate the greenness value G .

3. The control method as described in claim 2, characterized in that, In step d), based on the leaf region mask, multiple 100×100 pixel ROI regions are randomly selected, and the greenness values ​​calculated for each ROI region are averaged to obtain the leaf greenness index. G .

4. The control method as described in claim 1, characterized in that, In step S2, add exception handling rules: When all conditions are not met, the current growth stage will be... S Set to the most recent successfully determined historical growth stage; if the system has no historical records on its first run, then the default will be used. S It is the budding stage.

5. The control method as described in claim 1, characterized in that, A dynamic environmental compensation mechanism is added between steps S1 and S2, which includes: S1.1 Real-time acquisition of ambient light intensity L and air humidity H ; S1.2, if L <10 klux or H >85%, trigger compensation algorithm: Regression model fitted based on historical database Nb'=k 1 × Nb × e^ 0.02L and Nf'=k 2 × Nf ×(1-0.005H), correcting the number of flower buds. Nb and the number of fruits Nf ,in k 1 , k 2 Variety calibration coefficient; Correction value Nb' , Nf' replaces the original Nb , Nf Input the decision process for step S2.

6. The control method as described in claim 5, characterized in that, A soil moisture feedback closed-loop control is inserted between steps S3 and S4, which includes: S3.

1. The average volumetric water content at a depth of 20cm is obtained in real time by using a soil moisture sensor array buried in the root system of fruit trees. θ ; S3.2, if θ < θ slow (S) ,in, θ slow (S) For the current growth stage S The drought warning threshold is based on the amount of emergency irrigation. Q e =min( Q max , Q ×[1+0.2( θ slow (S) -θ)]) covers the original target irrigation volume range and simultaneously shortens the irrigation interval to 6 hours.

7. The control method as described in claim 6, characterized in that, A multi-source data consistency fusion mechanism is added between steps S2 and S3, including: S2.1 Real-time acquisition of ambient light intensity L air humidity H and the corrected number of flower buds Nb' Quantity of fruits Nf' At the same time, soil moisture content was obtained. θ And the results of the current growth stage determination S ; S2.2 Calculate the data conflict index: DCI=ω 1 × ΔN+ω 2 × Δθ ,in, ω 1. ω 2 Let be the weighting coefficient, satisfying ω 1+ ω 2=1, and dynamically calibrated using historical environmental and soil datasets, Δ N This corrects for the relative difference between the number of flower buds and fruits; the calculation formula is as follows: Δ θ It is the relative deviation of soil moisture content from the ideal value, and the calculation formula is: ; S2.3, if DCI ≥0.5 and Δθ > 0.2 × θ hist ， in, θ hist Based on historical averages, generate confidence growth stages. S' : like Δθ > ΔN According to θ Value matching growth phase: when θ << 0.15 hours S' The budding stage; when θ When ∈ [0.15, 0.25] S' Flowering season; when θ When ∈ [0.25, 0.35] S' This is the fruit growth period; when θ >0.35 S' Maturity stage; Otherwise, maintain S' = S ; Will S' Output to step S3; S2.4, if DCI <0.5, maintaining the original growth stage S Enter step S3.

8. The control method as described in claim 7, characterized in that, In step S2.2, a dynamic real-time weight calibration mechanism is added, including: S2.2.1 Calculate the environmental credibility score EC based on ambient light intensity L and air humidity H: Where EC∈[0,1], EC=1 indicates no environmental interference, and EC=0 indicates strong interference; S2.2.2, Based on soil moisture content θ Calculate the soil credibility score (SC): ,in θ avg This represents the average soil moisture content over the past 24 hours. θ std Its standard deviation; if the data is less than 24 hours, then θ avg = θ slow (S) , θ std =0.1; S2.2.3, Dynamic calibration weighting coefficients: ω1=(EC+ε) / (EC+SC+2ε)×0.7, ω2=(SC+ε) / (EC+SC+2ε)×0.3, where ε =0.01, which is a constant to prevent division by zero; S2.2.4, After calibration ω 1 , ω 2 enter DCI Calculation formula.

9. A water and fertilizer adaptive irrigation control system based on mango growth stages, characterized in that, It includes an image acquisition device, an image processing unit, a control unit, an irrigation system, and a fertilization device; An image acquisition device used to acquire images of the mango tree canopy once a day; An image processing unit, connected to an image acquisition device, is configured to perform the following operations: The pre-trained DeepLabv3+ segmentation network was used to divide the RGB image of the tree canopy into masked regions of flower buds, fruit, and leaves. Based on the bud region mask, the pixels within the region are converted to the HSV color space, and the dual-threshold conditional hue component H is applied. hsv Binarizing the image with saturation components Sat ∈ [0°, 15°] and saturation components Sat ∈ [0.4, 1.0] generates a primary binary image of flower buds. After performing a 3×3 elliptic kernel morphological closing operation on the primary binary image to eliminate noise and holes, a connected component labeling map is generated. Connected components with a pixel area greater than 50 are selected from the connected component labeling map, and their number is used as the number of flower buds. Nb ; Based on the fruit region mask, a morphological opening operation was performed using a 5×5 circular kernel to separate the adhered fruit regions. The minimum circumcircle of each connected region was extracted from the separated regions. The diameter of the circumcircle was converted to a physical size according to the camera calibration parameters. Connected regions with a physical diameter greater than 5 mm were selected, and their number was used as the fruit count. Nf ; Based on the leaf area mask, multiple 100×100 pixel ROI regions are randomly selected. Within each ROI region, RGB values ​​are converted to the CIE1976Lab color space according to the CIED65 standard light source model. For each pixel, the formula is applied... G = (2.5 × a ∗ - b ∗ The greenness value is calculated by dividing the greenness value by 100, and the average of the greenness values ​​of all ROI areas is taken as the leaf greenness index. G ; The control unit, connected to the image processing unit, is configured to perform the following operations: Based on extraction Nb , Nf and G And determine the current growth stage S : like Nb ≥6 and Nf When =0, the current growth stage is determined. S Flowering season; like Nb ≤5 and Nf If ≥1, then the current growth stage is determined. S This is the fruit growth period; like Nb ≥6 and Nf ≥1, calculate the transition period score TS=0.6×( Nb / 30)+0.4×( Nf / 50), when TS≥0.5, the current growth stage is determined. S This is the transitional period between flowering and fruiting; If none of the above conditions are met and G ∈[0.4,0.7], determine the current growth stage. S Maturity stage; Otherwise, determine the current growth stage. S The budding stage; When all conditions are not met, the current growth stage will be... S Set to the most recent successfully determined historical growth stage; if the system has no historical records on its first run, then the default will be used. S The budding stage; Based on the current growth stage S The target irrigation volume range is calculated using a water and fertilizer decision function model. Q min ( S ), Q max ( S )] and target fertilizer concentration range[ C min ( S ), C max ( S )]: Q min ( S ) =α s × K Q , Q max ( S )=β s × K Q , C min ( S )=γ s × K C , C max ( S )=δ s × K C ,in, K Q As the baseline irrigation amount, K C As the baseline concentration, α S , β S , γ S , δ S To be related to the growth stage S Correlation coefficient; The irrigation system, connected to the control unit, is configured to output... Q min ( S )to Q max ( S Irrigation volume within the specified range; The fertilizer applicator, connected to the control unit, is configured to output... C min ( S )to C max ( S The concentration of fertilizer solution within the range of ).

Images