Precise fertilization intelligent decision-making method and system based on multi-source remote sensing

Abstract

The invention relates to the technical field of intelligent decision making, in particular to an accurate fertilization intelligent decision making method and system based on multi-source remote sensing, and the method comprises the steps: obtaining a multi-source remote sensing image which comprises an unmanned aerial vehicle image and a stationary satellite image; acquiring soil nutrient data and fertility indexes of the sample field parcels; acquiring meteorological data and soil data; carrying out image preprocessing and feature extraction; constructing a crop identification and classification model; carrying out crop type identification and block division on the multi-source remote sensing image; extracting crop growth features based on the multi-source remote sensing image and the result of the crop identification and classification model; based on the crop growth characteristics and the soil nutrient data, analyzing the relationship between the crop growth stage and the nutrient demand; constructing a nutrient utilization efficiency maximization model; based on a nutrient utilization efficiency maximization model, generating a precise fertilization decision scheme; and outputting a precise fertilization decision scheme, and carrying out precise analysis and intelligent decision making on the growth state and nutrient demand of farmland crops.

Description

Technical Field

[0001] The present invention relates to the field of intelligent decision-making technology, and more specifically, to a precision fertilization intelligent decision-making method and system based on multi-source remote sensing. Background Art

[0002] With the continuous advancement of agricultural modernization, precision fertilization technology, as an important means to improve agricultural production efficiency and reduce environmental pollution, has become a research hotspot in the field of agricultural science and technology. Traditional fertilization methods often rely on experience and extensive soil testing, which is difficult to meet the needs of modern agriculture for precise and personalized fertilization. In recent years, with the rapid development of remote sensing technology and artificial intelligence, precision fertilization decision-making methods based on remote sensing data have gradually become the mainstream direction of research.

[0003] At present, the closest existing technologies mainly focus on using a single remote sensing data source for crop growth status monitoring and fertilization decision-making. These methods usually use satellite remote sensing or UAV remote sensing data to estimate crop nutrient status by calculating vegetation indices or establishing simple regression models. However, this single data source approach has many limitations. First, it is difficult for a single remote sensing platform to take into account both spatial resolution and temporal resolution at the same time, and it often cannot meet the needs of precision fertilization for high spatiotemporal resolution data. Secondly, the information provided by a single data source is limited, and it is difficult to fully reflect the complex changes in crop growth status and soil nutrients. In addition, existing methods mostly use traditional statistical methods or simple machine learning algorithms in data processing and analysis, which makes it difficult to fully explore the deep information in multi-source heterogeneous data.

[0004] On the other hand, most existing precision fertilization decision-making systems are static, rule-based systems that lack the ability to adapt to dynamic changes in crop growth. These systems usually make decisions based on preset fertilization rules and are difficult to flexibly adjust according to actual conditions. At the same time, existing systems often ignore the importance of agronomic expert knowledge and fail to effectively integrate expert experience with data-driven methods.

[0005] In addition, existing technologies also have deficiencies in evaluating fertilization effects and feedback mechanisms. Most systems lack real-time monitoring and evaluation of crop nutrient absorption after fertilization, and are unable to adjust fertilization strategies in a timely manner, resulting in low fertilizer utilization efficiency and even environmental pollution. Summary of the invention

[0006] In view of the problems existing in the above-mentioned prior art, the present invention proposes an intelligent decision-making method and system for precise fertilization based on multi-source remote sensing. The present invention aims to solve the following technical problems: how to effectively integrate multi-source remote sensing data to comprehensively and accurately monitor the growth status and nutrient requirements of crops; how to construct a dynamic and adaptive fertilization decision-making model to achieve precise and personalized fertilization plans; how to organically combine agronomic expert knowledge with data-driven methods to improve the scientificity and reliability of decision-making; and how to establish a real-time evaluation and feedback mechanism to continuously optimize fertilization strategies.

[0007] The present invention provides an intelligent decision-making method for precise fertilization based on multi-source remote sensing, including:

[0008] An acquisition step, including:

[0009] Acquire multi-source remote sensing images, where the multi-source remote sensing images include unmanned aerial vehicle images and geostationary satellite images;

[0010] Acquire soil nutrient data and fertility indicators of sample plots;

[0011] Acquire meteorological data and soil data;

[0012] A processing step, including:

[0013] Based on the multi-source remote sensing images, perform image preprocessing and feature extraction;

[0014] Based on the soil nutrient data and fertility indicators, construct a crop recognition and classification model;

[0015] Based on the crop recognition and classification model, identify crop types and divide blocks for the multi-source remote sensing images;

[0016] Based on the multi-source remote sensing images and the results of the crop recognition and classification model, extract crop growth characteristics;

[0017] Based on the crop growth characteristics and the soil nutrient data, analyze the relationship between crop growth stages and nutrient requirements;

[0018] Based on the relationship between crop growth stages and nutrient requirements, construct a model for maximizing nutrient use efficiency;

[0019] An output step, including:

[0020] Based on the model for maximizing nutrient use efficiency, generate a precise fertilization decision-making plan;

[0021] Output the precise fertilization decision-making plan.

[0022] Preferably, the image preprocessing specifically includes:

[0023] Perform orthorectification, filtering and denoising, radiometric calibration, resampling, atmospheric correction, geometric correction, image registration and fusion on the UAV images;

[0024] Perform atmospheric correction, geometric correction and image fusion on the geostationary satellite images.

[0025] Preferably, the feature extraction specifically includes:

[0026] Extract terrain slope and elevation data based on the UAV images;

[0027] Perform spectral inversion processing on the geostationary satellite images to obtain vegetation quantitative remote sensing parameters;

[0028] Construct time series data of vegetation quantitative remote sensing parameters for UAV images and time series data of vegetation quantitative remote sensing parameters for geostationary satellite images based on the vegetation quantitative remote sensing parameters.

[0029] Preferably, the construction of the crop recognition and classification model specifically includes:

[0030] Obtain the original image data containing different crops in the high-resolution satellite images;

[0031] Annotate the original image data to form a sample set;

[0032] Divide the sample set into a training set and a validation set;

[0033] Construct an initial classification model using a convolutional neural network;

[0034] Use the backpropagation algorithm and the stochastic gradient descent function to minimize the average cross-entropy loss function and train the parameters of the initial classification model;

[0035] Evaluate the model accuracy using the validation set, and when the model reaches the preset accuracy threshold, obtain the crop recognition and classification model.

[0036] Preferably, the extraction of the crop growth characteristics specifically includes:

[0037] Use a UAV equipped with a multispectral digital camera to obtain near-surface images of farmland;

[0038] Use a spectral analyzer to obtain spectral curves of different crops in different bands;

[0039] Adopt an improved U-Net network to extract the spatial features of different crop species from the near-surface images to obtain a spatial feature distribution map of different crops;

[0040] Construct a crop chlorophyll content analysis model using a convolutional neural network;

[0041] Based on the crop chlorophyll content analysis model, the chlorophyll content indexes of crop leaves at each growth stage are calculated.

[0042] Preferably, the analysis of the relationship between the crop growth stage and nutrient requirements specifically includes:

[0043] Define the growth stage of the crop and the corresponding objective function;

[0044] Combined with the mapping relationship between chlorophyll content and spectral characteristics, design an objective function solving model for different growth stages of crops based on transfer learning and convolutional neural network;

[0045] Construct a stage feature extraction model, and according to the result of crop recognition, input the objective function of different growth stages of the crop to calculate the crop feature distribution in different growth stages.

[0046] Preferably, the construction of the nutrient use efficiency maximization model specifically includes:

[0047] Combined with multiple crop objective functions, extract the nutritional characteristics of crops at different stages to determine the crop nutrient requirements;

[0048] Adopt an improved U-Net network to convert the spectral curve features of the crop into crop health degree image features;

[0049] Introduce a temporal attention mechanism and adopt a reinforcement learning framework to realize nutrient decision-making for crops at different stages;

[0050] Construct a precise fertilization plan library based on historical data;

[0051] According to the knowledge of different agronomists, extract the impacts of different fertilization plans on soil and climate, and construct an agronomist knowledge distillation model.

[0052] Preferably, it further includes an evaluation step:

[0053] During the growth and development of the crop, monitor the nutrient utilization rates of nitrogen, phosphorus, and potassium respectively;

[0054] When the nutrient utilization rate is lower than 80%, fertilize according to the fertilization rules;

[0055] Monitor the change of nutrient utilization rate in real time to judge whether fertilization is completed.

[0056] Preferably, the meteorological data includes wind speed and direction, light intensity, temperature, and humidity; the soil data includes soil water content and soil pH value, where the soil pH value is a parameter with 5 dimensions, representing acidity coefficient, alkalinity coefficient, salinity coefficient, hardness coefficient, and alkalinity index respectively.

[0057] An intelligent decision-making system for precise fertilization based on multi-source remote sensing that executes the described method, comprising:

[0058] A data acquisition unit for collecting multi-source remote sensing images, soil nutrient data, fertility indicators, meteorological data, and soil data;

[0059] A data preprocessing unit for preprocessing and feature extraction of the multi-source remote sensing images;

[0060] A crop recognition and classification unit for constructing a crop recognition and classification model and performing crop species recognition and block division on the preprocessed multi-source remote sensing images;

[0061] A crop growth analysis unit for extracting crop growth characteristics and analyzing the relationship between crop growth stages and nutrient requirements;

[0062] A nutrient use efficiency modeling unit for constructing a model to maximize nutrient use efficiency;

[0063] A decision generation unit for generating a precise fertilization decision-making plan based on the model to maximize nutrient use efficiency;

[0064] An evaluation and feedback unit for monitoring nutrient utilization rate and adjusting fertilization decisions according to the monitoring results;

[0065] Among them, the system realizes intelligent decision-making for precise fertilization based on multi-source remote sensing through the collaborative work of the above-mentioned units.

[0066] The present invention realizes precise analysis and intelligent decision-making on the growth status and nutrient requirements of farmland crops by innovatively combining advanced technologies such as multi-source remote sensing data, deep learning, and reinforcement learning, and integrating agronomic expert knowledge. Specifically, the present invention has the following remarkable beneficial effects:

[0067] 1. Multi-source data fusion effect: The present invention realizes the organic combination of high spatial resolution and high temporal resolution by fusing UAV images and geostationary satellite images, greatly improving the accuracy and timeliness of monitoring crop growth status. The synergistic effect of multi-source data not only makes up for the deficiencies of a single data source but also provides a more comprehensive and reliable information basis through complementary enhancement.

[0068] 2. Deep learning-enabled feature extraction: Using deep learning models such as the improved U-Net network for image processing and feature extraction significantly improves the accuracy of crop recognition and growth status analysis. This method can automatically learn and extract complex features, avoiding the limitations of manually designed features in traditional methods and providing richer and more effective information input for precise fertilization decision-making.

[0069] 3. Dynamic Adaptive Decision-making Mechanism: By introducing a temporal attention mechanism and a reinforcement learning framework, the present invention achieves precise capture of the dynamic changes in crop growth and adaptive decision-making. This dynamic decision-making mechanism can continuously adjust the fertilization strategy according to real-time monitoring data, effectively cope with the complex changes in the farmland environment, and improve the adaptability and accuracy of fertilization decisions.

[0070] 4. Organic Combination of Expert Knowledge and Data-driven: The present invention innovatively proposes an agronomic expert knowledge distillation model, which skillfully integrates expert experience into the data-driven decision-making model. This method not only retains the advantages of data-driven methods but also fully utilizes the valuable experience of agronomic experts, significantly improving the scientificity and interpretability of decisions.

[0071] 5. Closed-loop Optimization Evaluation and Feedback Mechanism: By real-time monitoring the nutrient utilization rate and timely adjusting the fertilization strategy, the present invention establishes a complete closed-loop optimization system. This mechanism can continuously evaluate the fertilization effect, timely discover and correct problems, and continuously optimize the fertilization plan, thereby achieving the maximization of fertilizer utilization efficiency and the minimization of environmental impact.

[0072] 6. System Integration and Practicality Enhancement: The precise fertilization intelligent decision-making system proposed by the present invention adopts a modular and scalable architecture design, with good adaptability and scalability. The system integrates the entire process from data collection, processing, analysis to decision generation, and provides a user-friendly interface, greatly improving the practicality and promotion value of the system.

[0073] In summary, through the organic combination of innovative technologies such as multi-source data fusion, deep learning, and reinforcement learning, the present invention realizes the intelligence and personalization of precise fertilization decision-making. This not only significantly improves fertilizer utilization efficiency and reduces environmental pollution but also provides scientific and reliable fertilization guidance for farmers. The methods and systems of the present invention have important theoretical significance and practical application value for promoting agricultural modernization and sustainable development. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is the flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0075] Please refer to Figure 1 , the present invention provides a precise fertilization intelligent decision-making method and system based on multi-source remote sensing. This method makes full use of multi-source remote sensing data and combines advanced artificial intelligence technologies to achieve precise analysis of the growth status and nutrient requirements of farmland crops, thereby providing intelligent decision-making support for precise fertilization. The technical solutions of the present invention will be described in detail below.

[0076] First, the method of the present invention includes an acquisition step, a processing step, and an output step. In the acquisition step, three types of data are mainly acquired: multi-source remote sensing images, soil nutrient data and fertility indicators, and meteorological and soil data. Among them, the multi-source remote sensing images include unmanned aerial vehicle images and geostationary satellite images. The acquisition of such multi-source data lays a solid data foundation for subsequent precise analysis.

[0077] In the processing step, the method of the present invention first preprocesses and extracts features from the acquired multi-source remote sensing images. The preprocessing process includes operations such as image correction, denoising, and calibration to improve data quality. Feature extraction mainly extracts valuable information from the preprocessed images, such as vegetation indices, terrain features, etc.

[0078] Next, based on the soil nutrient data and fertility indicators, the method constructs a crop recognition and classification model. This model uses convolutional neural network (CNN) technology and can effectively identify different types of crops. In the model construction process, first, high-resolution satellite image samples containing different crops are prepared, and after being labeled, they are divided into a training set and a validation set. Then, the model parameters are optimized through the backpropagation algorithm and the stochastic gradient descent function to minimize the average cross-entropy loss function. The model training process can be expressed as:

[0079]

[0080] where L is the loss function, N is the number of samples, C is the number of crop categories, y ij is the true label, and p ij is the predicted probability.

[0081] Preferably, the present invention uses ResNet50 as the basic network structure and performs fine-tuning and optimization on this basis. During model training, the initial learning rate is set to 0.001 and decays by 10 times every 50 epochs. When the accuracy on the validation set reaches 98% or there is no obvious improvement for 10 consecutive epochs, the training stops. Such settings can avoid overfitting problems while ensuring the model performance.

[0082] After obtaining the crop recognition and classification model, the method applies it to the multi-source remote sensing images to realize the recognition and block division of farmland crop types. This step provides important spatial reference information for subsequent precise analysis.

[0083] The method of the present invention further includes extracting crop growth characteristics and analyzing the relationship between crop growth stages and nutrient requirements. Specifically, a multi-spectral camera carried by a drone is used to obtain near-surface images of farmland, and a spectral analyzer is used to obtain spectral curves of different crops in different bands. Then, an improved U-Net network is used to extract the spatial characteristics of crops, obtaining a crop spatial characteristics distribution map. The improvement of the U-Net network is mainly reflected in the following aspects:

[0084] 1. Introduce residual connections to alleviate the problem of gradient disappearance;

[0085] 2. Use a spatial attention mechanism to enhance the model's perception ability of key regions;

[0086] 3. Adopt depthwise separable convolutions to reduce the number of parameters and improve computational efficiency.

[0087] Based on the extracted spatial characteristics, this method further constructs a crop chlorophyll content analysis model. This model uses convolutional neural network technology to combine spectral characteristics and spatial characteristics to achieve accurate estimation of chlorophyll content. The loss function of the model uses the root mean square error (RMSE), expressed as:

[0088]

[0089] where n is the number of samples, y i is the true chlorophyll content, is the predicted value. In practical applications, the present invention also considers the different requirements for chlorophyll content at different growth stages. Therefore, the spectral curve characteristics of crop leaves are decomposed to obtain multiple linear regression equations for different stages. For example, for rice, the growth period can be divided into four stages: tillering stage, jointing stage, heading stage, and filling stage. The chlorophyll content estimation equation for each stage can be expressed as:

[0090] Chl i =a i ·NDVI+b i ·EVI+c i ·OSAVI+d i ,

[0091] where Chl i is the chlorophyll content in the i-th growth stage, NDVI, EVI, and OSAVI are the normalized difference vegetation index, enhanced vegetation index, and optimized soil-adjusted vegetation index respectively, and a i , b i , c i and d i are regression coefficients.

[0092] Through the above steps, the method of the present invention realizes a comprehensive analysis of the crop growth status. Next, based on these analysis results, a model for maximizing nutrient use efficiency is constructed. This model comprehensively considers the crop growth stage, nutrient requirements, soil conditions, and meteorological factors, aiming to achieve precise fertilization and improve fertilizer use efficiency.

[0093] During the model construction process, the present invention introduces a temporal attention mechanism and a reinforcement learning framework. The temporal attention mechanism can capture the dynamic changes in the crop growth process, while the reinforcement learning framework can continuously optimize the fertilization decision according to the actual situation. Specifically, the deep Q-network (DQN) is used as the core algorithm of reinforcement learning, and its Q-value update formula is:

[0094]

[0095] where Q(s t ,a t ) represents the Q-value of taking action a t in state s t , r t is the immediate reward, γ is the discount factor, and α is the learning rate. The method of the present invention also constructs a precise fertilization plan library and introduces an agronomic expert knowledge distillation model. This approach can organically combine expert experience with data-driven methods to further improve the scientificity and reliability of decision-making. The knowledge distillation process can be expressed as:

[0096]

[0097] where L KD is the knowledge distillation loss function, T is the temperature parameter, KL is the KL divergence, σ is the softmax function, z t and z s are the output logits of the teacher model and the student model respectively, y is the true label, CE is the cross-entropy loss, and α is the balance parameter.

[0098] Finally, in the output step, the method of the present invention generates a precise fertilization decision plan based on the model for maximizing nutrient use efficiency and outputs this plan. This plan not only considers the real-time growth status and nutrient requirements of the crop, but also combines soil conditions, meteorological factors, and expert experience, and can provide scientific and precise fertilization guidance for farmers.

[0099] The method of the present invention also includes an evaluation step of monitoring the nutrient utilization rates of nitrogen, phosphorus, and potassium during the growth and development of crops. When the nutrient utilization rate is lower than 80%, the system will apply fertilizers according to the preset fertilization rules. The selection of this threshold is based on a large amount of experimental data and expert experience, which can not only ensure that the crops obtain sufficient nutrients but also avoid waste and environmental pollution caused by over-fertilization. At the same time, the system will monitor the changes in the nutrient utilization rate in real time and timely judge whether the fertilization is completed, so as to achieve precise control.

[0100] In practical applications, the method of the present invention also fully considers the influence of meteorological and soil factors. Meteorological data includes key parameters such as wind speed and direction, light intensity, temperature, and humidity. Soil data includes soil water content and pH value, where the pH value is subdivided into 5 dimensions: acidity coefficient, alkalinity coefficient, salinity coefficient, hardness coefficient, and alkalinity index. This multi-dimensional method for characterizing soil pH value can more comprehensively reflect the physical and chemical properties of the soil and provide a more reliable basis for precise fertilization decision-making.

[0101] Generally speaking, the intelligent decision-making method for precise fertilization based on multi-source remote sensing proposed by the present invention realizes precise analysis and intelligent decision-making of the growth status and nutrient requirements of farmland crops through advanced technologies such as multi-source data fusion, deep learning, and reinforcement learning. This method can not only improve the fertilizer utilization efficiency and reduce environmental pollution but also provide scientific and reliable fertilization guidance for farmers, having important practical value and promotion significance. The improved U-Net network structure maintains the overall encoder-decoder structure of U-Net, but has made several improvements on this basis. First, residual connections are introduced into each convolutional block, which helps to alleviate the vanishing gradient problem in deep networks and enables information to be transmitted more directly between the shallow and deep layers. Second, spatial attention modules are introduced at key positions in the encoder and decoder, which enables the network to better focus on important regions in the image and improve the efficiency of feature extraction. Finally, the present invention uses depthwise separable convolutions to replace standard convolutions, and this improvement significantly reduces the number of network parameters while maintaining good feature extraction capabilities.

[0102] The method of the present invention adopts different processing procedures for UAV images and geostationary satellite images. For UAV images, orthorectification is first performed to eliminate image distortion caused by the shooting angle. This step is usually implemented using the collinearity equation and digital elevation model (DEM). The orthorectified image can more accurately reflect the actual position and shape of the ground objects.

[0103] Next, the method filters and denoises the UAV images. Preferably, the bilateral filtering algorithm is used, which can remove noise while maintaining the sharpness of the image edges. The mathematical expression of bilateral filtering is as follows:

[0104]

[0105] Among them, I is the input image, p and q are pixel positions, and are Gaussian functions in the spatial domain and value domain respectively, and W p is the normalization factor. In the embodiments of the present invention, the values of σ s and σ r range from 3 - 5 and 0.1 - 0.3 respectively. These parameter values are obtained through a large number of experiments and can achieve a good balance between denoising effect and detail preservation.

[0106] Subsequently, the method performs radiometric calibration and resampling on the UAV images. Radiometric calibration aims to convert the digital number (DN value) of the image into a physical quantity, such as radiance or reflectance. Resampling is to unify the image resolutions obtained by different sensors. In the preferred embodiment of the present invention, bicubic interpolation is used for resampling, and this method can effectively reduce the error introduced by resampling while maintaining the image quality.

[0107] For geostationary satellite images, the method of the present invention mainly performs atmospheric correction and geometric correction. Atmospheric correction uses the FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hyper cubes) model, which is based on the MODTRAN radiative transfer code and can effectively remove the influence of atmospheric scattering and absorption. Geometric correction is achieved using a polynomial model and ground control points (GCPs). Usually, 10 - 15 evenly distributed GCPs are selected, and the correction accuracy is controlled within 0.5 pixels.

[0108] Finally, the method fuses the processed UAV images and geostationary satellite images. Preferably, a high-pass filtering (HPF) fusion algorithm is used, which can improve the spatial resolution while maintaining the spectral characteristics of the multispectral image. The mathematical expression of HPF fusion is as follows:

[0109] MS fus = MS + (PAN HP × gain),

[0110] where MS fus is the fused multispectral image, MS is the original multispectral image, PAN HP is the result of the panchromatic image after high-pass filtering, and gain is the gain coefficient. In the embodiments of the present invention, the value range of gain is 0.25 - 0.35, and this value can achieve a good balance between spectral fidelity and spatial detail enhancement.

[0111] Feature extraction step. The method of the present invention first extracts terrain slope and elevation data based on UAV images. This step is implemented using a digital elevation model (DEM). Specifically, the Horn algorithm is used to calculate the slope, which takes into account the elevation values of 8 adjacent pixels around the central pixel. The calculation formula is as follows:

[0112]

[0113] where and are the elevation gradients in the x and y directions respectively.

[0114] For geostationary satellite images, the method performs spectral inversion processing to obtain vegetation quantitative remote sensing parameters. Preferably, the PROSAIL model is used for inversion. This model combines the leaf optical property model PROSPECT and the canopy bidirectional reflectance model SAIL, and can effectively simulate the bidirectional reflectance characteristics of the vegetation canopy. The input parameters of the PROSAIL model include chlorophyll content, leaf area index, average leaf inclination angle, etc., and the output is the reflectance in different bands. By minimizing the difference between the simulated reflectance and the measured reflectance, the vegetation parameters can be inverted.

[0115] Based on the obtained vegetation quantitative remote sensing parameters, the method further constructs the time series data of the vegetation quantitative remote sensing parameters of UAV images and geostationary satellite images. These time series data are of great significance for analyzing crop growth status and predicting yields. When constructing the time series data, the present invention uses the Savitzky-Golay filtering algorithm to smooth the original data to eliminate the influence of noise and outliers. The mathematical expression of Savitzky-Golay filtering is as follows:

[0116]

[0117] where g i is the filtered value, f i is the original value, c n is the convolution coefficient, and m is half of the filtering window size. In the embodiments of the present invention, the filtering window size is selected to be 5 - 9, and the polynomial order is 2 - 3. These parameter settings can effectively remove the influence of short-term fluctuations while maintaining the trend of the time series data.

[0118] Through the above steps, the method of the present invention realizes the comprehensive preprocessing and feature extraction of multi-source remote sensing images. These processed data provide a reliable basis for subsequent crop identification, growth status analysis and fertilization decision-making. It should be noted that each processing step of the method has been carefully parameter-tuned and verified to ensure the best effect in practical applications.

[0119] In practical applications, the method of the present invention also takes into account the particularities of different types of crops and regions. For example, for crops such as rice planted in paddy fields, the influence of water bodies needs to be particularly considered when extracting terrain features. In this case, the method uses an improved DEM extraction algorithm, which combines multi-temporal images and water indices (such as NDWI) to accurately describe the terrain of paddy fields.

[0120] In addition, the method of the present invention also has good scalability and adaptability. With the continuous development of remote sensing technology, new sensors and platforms are emerging continuously. This method can flexibly integrate new data sources and processing algorithms according to needs. For example, for hyperspectral remote sensing data, feature selection and dimensionality reduction algorithms such as principal component analysis (PCA) or independent component analysis (ICA) can be introduced to effectively process high-dimensional data.

[0121] Generally speaking, the multi-source remote sensing image preprocessing and feature extraction method proposed by the present invention realizes the comprehensive and efficient processing of remote sensing data through a series of carefully designed algorithms and processes. This not only provides reliable data support for subsequent precise fertilization decision-making, but also lays a solid foundation for other agricultural remote sensing applications. After the preprocessing and feature extraction of multi-source remote sensing images are completed by the method of the present invention, a crop recognition and classification model is further constructed, which is crucial for precise fertilization decision-making.

[0122] For the construction of the crop recognition and classification model, the method first obtains the original image data containing different crops in high-resolution satellite images. Preferably, satellite images with a spatial resolution of 0.5 - 2 meters are used, such as WorldView-3 or GaoFen-2, etc. Such high-resolution images can clearly show the texture and structural features of crops, which is beneficial to improving the classification accuracy.

[0123] After obtaining the original image data, the method annotates these data to form a sample set. During the annotation process, professional remote sensing image interpretation software such as ENVI or eCognition is used, combined with ground survey data and expert knowledge for accurate annotation. To ensure the representativeness and balance of the samples, in the embodiments of the present invention, no less than 1000 samples are selected for each type of crop, and the samples are ensured to be evenly distributed in space and time.

[0124] Next, the method divides the annotated sample set into a training set and a validation set. Preferably, the stratified random sampling strategy is adopted, and the training set and the validation set are divided according to a ratio of 7:3. This division method can ensure that the proportion of each type of crop in the training set and the validation set is the same, avoiding model bias caused by sample imbalance.

[0125] In terms of model construction, the present invention uses a convolutional neural network (CNN) as the basic architecture. Specifically, ResNet50 is selected as the backbone network and targeted improvements are made. The improvements mainly include the following aspects:

[0126] 1. Introduce a spatial attention mechanism to enhance the model's perception ability of key features.

[0127] 2. Adopt the FPN (Feature Pyramid Network) structure to improve the extraction ability of multi-scale features.

[0128] 3. Use Focal Loss as the loss function to alleviate the problem of class imbalance.

[0129] During the model training process, this method uses the backpropagation algorithm and the stochastic gradient descent function to optimize the model parameters. The loss function uses Focal Loss, and its mathematical expression is as follows:

[0130] FL(p t )=-α t (1-p t ) γ log(p t ),

[0131] where p t is the predicted probability, α t is the balance factor, and γ is the focusing parameter. In the preferred embodiment of the present invention, α t is set to 0.25 and γ is set to 2. These parameter values are obtained through a large number of experiments, which can effectively alleviate the problem of class imbalance and improve the classification accuracy of small-class samples.

[0132] Preferably, this method uses the Adam optimizer, the initial learning rate is set to 0.001, and the cosine annealing strategy is used to dynamically adjust the learning rate. During the training process, the performance of the validation set is evaluated every 50 epochs. When the validation set accuracy reaches 98% or there is no obvious improvement for 10 consecutive epochs, the training stops. This training strategy can effectively prevent overfitting while ensuring the model performance.

[0133] For crop growth feature extraction, the method of the present invention first uses a drone equipped with a multi-spectral digital camera to obtain near-ground images of farmland. Preferably, a 6-channel multi-spectral camera is used, including blue, green, red, red edge, near-infrared, and thermal infrared bands. This multi-spectral combination can comprehensively capture the spectral characteristics of crops, which is beneficial for subsequent analysis.

[0134] Meanwhile, this method uses a spectral analyzer to obtain spectral curves of different crops in different bands. Preferably, a hyperspectral analyzer with a wavelength range of 350 - 2500 nm is adopted, and the spectral resolution is better than 3 nm. Such high-precision spectral data provides a reliable basis for establishing the relationship between crop physiological characteristics and spectral features.

[0135] After obtaining multi-source data, this method uses an improved U-Net network to extract the spatial features of different crop species from near-surface images, and obtains spatial feature distribution maps of different crops. The structure of the improved U-Net network is as described above, and the mathematical expression of the spatial attention mechanism is as follows:

[0136] M(F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)]))

[0137] where F is the input feature map, and f 7×7 represents a 7×7 convolution operation, and σ is the sigmoid activation function. This attention mechanism can help the model better focus on important spatial regions and improve the efficiency of feature extraction. Next, this method uses a convolutional neural network to construct a crop chlorophyll content analysis model. This model takes the extracted spatial features and spectral features as inputs and outputs the estimated chlorophyll content. The loss function of the model uses the weighted sum of the root mean square error (RMSE) and the relative error (RE), and the mathematical expression is as follows:

[0138] L = α·RMSE + (1 - α)·RE

[0139] where α is the weight coefficient, which takes the value of 0.7 in the embodiments of the present invention. This loss function design can consider both the absolute error and the relative error simultaneously, and improve the estimation accuracy of the model in different chlorophyll content ranges.

[0140] Based on the constructed chlorophyll content analysis model, this method further considers the influence of different growth stages on chlorophyll content. Specifically, the crop growth period is divided into several key stages, such as the tillering stage, jointing stage, heading stage, and filling stage of rice. For each stage, an independent multiple linear regression equation is established to estimate the chlorophyll content. The general form of the regression equation is as follows:

[0141] Chl i = a i ·NDVI + b i ·EVI + c i ·OSAVI + d i ·LAI + e i

[0142] where Chl iis the chlorophyll content at the i-th growth stage, NDVI, EVI, and OSAVI are the normalized difference vegetation index, enhanced vegetation index, and optimized soil-adjusted vegetation index respectively, LAI is the leaf area index, a i , b i , c i , d i and e i are regression coefficients. This phased estimation method can more accurately reflect the change of chlorophyll content of crops at different growth stages.

[0143] Analysis of the relationship between crop growth stages and nutrient requirements. The method of the present invention first defines the growth stages of crops and the corresponding objective functions. Taking rice as an example, its growth cycle is divided into five stages: seedling stage, tillering stage, jointing-booting stage, heading stage, and filling-maturing stage. The objective function of each stage considers the key physiological indexes of this stage, such as the number of tillers, the number of grains per panicle, and the 1000-grain weight.

[0144] On this basis, this method combines the mapping relationship between chlorophyll content and spectral characteristics, and designs a solution model for the objective functions of different growth stages of crops based on transfer learning and convolutional neural network. The core idea of this model is to use a pre-trained deep learning model (such as VGG16 or ResNet) as a feature extractor, and then add an adaptation layer and an output layer on this basis to adapt to the solution of the objective function of a specific crop and growth stage. The loss function of the model is designed as follows:

[0145] L total = L task + λ·L transfer ,

[0146] where, L t ask is the loss of a specific task (such as mean square error), L transfer is the regularization term of transfer learning (such as L2 norm), and λ is a balance coefficient. In the embodiments of the present invention, the value range of λ is 0.01 - 0.1, and the specific value is adjusted according to the size of the data set and the complexity of the task.

[0147] Finally, this method constructs a stage feature extraction model. According to the results of crop recognition, the objective functions of different growth stages of crops are input, and the crop feature distribution in different growth stages is calculated. This model adopts a recurrent neural network (RNN) structure, especially a long short-term memory network (LSTM), to capture the temporal characteristics of crop growth.

[0148] Through the above model, the method of the present invention can comprehensively analyze the feature distribution of crops at different growth stages, providing an important basis for subsequent nutrient requirement analysis and fertilization decision-making.

[0149] Construction of the nutrient utilization efficiency maximization model. The method of the present invention first combines multiple crop objective functions, extracts the nutritional characteristics of crops at different stages, and determines the nutrient requirements of crops. This step is based on the above-mentioned analysis results of crop growth stages, and comprehensively considers the physiological characteristics of crops and environmental factors.

[0150] Next, this method uses an improved U-Net network to convert the spectral curve characteristics of crops into crop health degree image characteristics. The improvements are mainly reflected in the following aspects:

[0151] 1. Introduce residual connections to alleviate the problem of gradient disappearance;

[0152] 2. Adopt spatial and channel attention mechanisms to enhance the model's perception ability of key features;

[0153] 3. Use dilated convolution to expand the receptive field and capture context information in a larger range.

[0154] While maintaining the original structural advantages, the improved U-Net network significantly improves the efficiency of feature extraction and conversion.

[0155] During the model training process, the present invention introduces a temporal attention mechanism to capture the dynamic changes in the crop growth process. The mathematical expression of the temporal attention mechanism is as follows:

[0156]

[0157] Among them, h t is the hidden state at the current moment, is the average value of the hidden states at all moments, and score is the similarity calculation function. This mechanism can adaptively focus on the important features in different time periods and improve the model's sensitivity to temporal changes. To optimize the nutrient decision-making of crops at different stages, this method adopts a reinforcement learning framework.

[0158] Specifically, the deep Q-network (DQN) is used as the core algorithm. The state space of DQN includes factors such as crop growth stage, soil nutrient content, and meteorological conditions; the action space is the fertilizer application schemes of different types and dosages; the reward function considers multiple aspects such as crop yield, fertilizer utilization efficiency, and environmental impact. The Q-value update formula is as follows:

[0159]

[0160] Among them, Q(s t ,a t ) represents the Q-value of taking action a t in state s t , r tδ is the immediate reward, γ is the discount factor, and α is the learning rate. In the embodiments of the present invention, γ takes a value of 0.95, the initial value of α is 0.001, and a learning rate decay strategy is adopted.

[0161] In addition, the method of the present invention also constructs a precise fertilization plan library based on historical data. This plan library not only contains successful fertilization cases in history but also integrates the knowledge and experience of agronomy experts. To effectively utilize these valuable experience and knowledge, this method constructs an agronomy expert knowledge distillation model. This model takes the fertilization decision of agronomy experts as the teacher model and the fertilization decision system based on deep learning as the student model. Through the way of knowledge distillation, the expert experience is transformed into a form that can be understood and applied by the deep learning model.

[0162] The loss function of knowledge distillation is designed as follows:

[0163]

[0164] Among them, L KD is the knowledge distillation loss function, T is the temperature parameter, KL is the KL divergence, σ is the softmax function, z t and z s are the output logits of the teacher model and the student model respectively, y is the true label, CE is the cross-entropy loss, and α is the balance parameter.

[0165] After constructing the model for maximizing nutrient use efficiency, the method of the present invention further introduces an evaluation and feedback mechanism to ensure the accuracy and effectiveness of fertilization decisions.

[0166] During the growth and development period of crops, this method monitors the utilization rates of the three main nutrients, nitrogen, phosphorus, and potassium, respectively. The calculation of nutrient utilization rate adopts an improved apparent recovery rate method, and its mathematical expression is as follows:

[0167]

[0168] Among them, NUE i is the utilization rate of the i-th nutrient, U i is the absorption amount of the i-th nutrient by the crop under fertilization conditions, U 0 is the absorption amount of the i-th nutrient by the crop under non-fertilization conditions, and F i is the application amount of the i-th nutrient. In practical applications, this method uses near-ground hyperspectral remote sensing technology combined with machine learning algorithms to achieve rapid and non-destructive estimation of crop nutrient absorption amounts.

[0169] Preferably, the present invention sets a threshold value of the nutrient utilization rate at 80%. When the utilization rate of a certain nutrient is monitored to be lower than this threshold, the system will trigger a fertilization decision. The selection of this threshold is based on a large amount of field test data and expert experience, which can not only ensure that crops obtain sufficient nutrient supply but also minimize nutrient loss and environmental pollution.

[0170] When making a fertilization decision, this method will perform intelligent matching according to a pre-established fertilization rule base. The construction of the fertilization rule base comprehensively considers multiple factors such as crop variety, growth stage, soil type, and climate conditions, and uses a method combining fuzzy logic and decision tree to achieve precise decision-making under complex conditions.

[0171] After fertilization, this method will continuously monitor the change of the nutrient utilization rate to determine whether fertilization is completed. This real-time monitoring and feedback mechanism can effectively prevent over-fertilization and improve fertilizer utilization efficiency. During the monitoring process, the exponentially weighted moving average (EWMA) algorithm is used to smooth the nutrient utilization rate data to eliminate the influence of short-term fluctuations. The calculation formula of EWMA is as follows:

[0172] S t =αY t +(1-α)S t-1 ,

[0173] where, S t is the smoothed value at time t, Y t is the observed value at time t, and α is the smoothing coefficient. In the embodiment of the present invention, α takes a value of 0.3, and this value can achieve a better balance between sensitivity and stability.

[0174] Regarding meteorological and soil data, the method of the present invention fully considers the influence of these key environmental factors on precise fertilization. Meteorological data includes wind speed and direction, light intensity, temperature, and humidity. The acquisition of these data mainly adopts two methods: one is to collect in real time using a field meteorological station, and the other is to perform data assimilation and short-term forecasting by combining a mesoscale numerical weather prediction model (such as the WRF model).

[0175] Soil data includes soil moisture content and pH value. Among them, the soil pH value is subdivided into 5 dimensions: acidity coefficient, alkalinity coefficient, salinity coefficient, hardness coefficient, and alkalinity index. This multi-dimensional method for characterizing the pH value stems from the research accumulation of the research team of the present invention and can more comprehensively and accurately reflect the physical and chemical properties of the soil. The calculation methods of each dimension coefficient are as follows:

[0176] 1. Acidity coefficient (AC): AC=-log 10 [H + , when pH<7;

[0177] 2. Alkalinity Coefficient (BC): BC = -log 10 [OH - , when pH > 7;

[0178] 3. Salinity Coefficient (SC): SC = EC × 640, where EC is the electrical conductivity;

[0179] 4. Hardness Coefficient (HC):

[0180] 5. Alkalinity Index (AI):

[0181] These refined soil parameters provide a more reliable and comprehensive basis for precise fertilization decision-making.

[0182] Finally, based on the multi-source remote sensing-based precise fertilization intelligent decision-making system, the present invention proposes a modular and scalable system architecture. The system mainly includes the following core units:

[0183] Data acquisition unit 1, responsible for collecting multi-source remote sensing images, soil nutrient data, fertility indicators, meteorological data, and soil data. This unit adopts a distributed data acquisition architecture and supports seamless access of multiple sensors and data sources.

[0184] Data preprocessing unit 2, preprocesses and extracts features from the collected multi-source remote sensing images. This unit adopts a pipeline processing architecture and can efficiently process large-scale remote sensing data.

[0185] Crop recognition and classification unit 3, constructs a crop recognition and classification model, and performs crop species recognition and block division on the preprocessed multi-source remote sensing images. This unit is implemented based on a deep learning framework (such as TensorFlow or PyTorch) and supports online update and transfer learning of the model.

[0186] Crop growth analysis unit 4, extracts crop growth characteristics and analyzes the relationship between crop growth stages and nutrient requirements. This unit integrates a variety of machine learning algorithms and physiological and ecological models and can adapt to the characteristics of different crops and regions.

[0187] Nutrient use efficiency modeling unit 5, constructs a model for maximizing nutrient use efficiency. This unit adopts a reinforcement learning framework and combines an expert knowledge base to achieve dynamic optimization of fertilization decisions.

[0188] Decision generation unit 6, generates a precise fertilization decision-making plan based on the model for maximizing nutrient use efficiency. This unit adopts a multi-criteria decision-making method and comprehensively considers economic benefits, environmental impacts, and operational feasibility.

[0189] The evaluation and feedback unit 7 monitors the nutrient utilization rate and adjusts the fertilization decision according to the monitoring results. This unit adopts a closed-loop control strategy to ensure the continuous optimization of the fertilization decision.

[0190] These units interact through standardized interfaces and data formats, forming a complete intelligent decision-making closed loop. The system adopts a microservices architecture, and each unit can be independently deployed and scaled to adapt to application scenarios of different scales and complexities.

[0191] In practical applications, this system is also equipped with a visualization interface and a mobile APP, which facilitate farmers and agricultural technicians to view the field conditions in real time, receive fertilization suggestions, and provide operation feedback. The system also supports direct docking with intelligent fertilization equipment to achieve a fully automated process of "decision-making - execution - feedback".

[0192] Generally speaking, the precise fertilization intelligent decision-making method and system based on multi-source remote sensing proposed by the present invention achieve precise analysis and intelligent decision-making on the growth status and nutrient requirements of farmland crops through advanced technologies such as multi-source data fusion, deep learning, and reinforcement learning. This method and system can not only significantly improve the fertilizer utilization efficiency and reduce environmental pollution, but also provide scientific and reliable fertilization guidance for farmers, which is of great significance for promoting agricultural modernization and sustainable development.

[0193] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. Intelligent decision-making method for precision fertilization based on multi-source remote sensing, characterized by ,include: The acquisition steps include: Acquiring multi-source remote sensing images, wherein the multi-source remote sensing images include drone images and stationary satellite images; Obtain soil nutrient data and fertility indicators of sample plots; Obtain meteorological data and soil data; The processing steps include: Based on the multi-source remote sensing images, image preprocessing and feature extraction are performed; Based on the soil nutrient data and fertility index, a crop identification and classification model is constructed; Based on the crop identification and classification model, the multi-source remote sensing image is subjected to crop type identification and block division; Extracting crop growth characteristics based on the multi-source remote sensing images and the results of the crop identification and classification model; Analyzing the relationship between crop growth stage and nutrient demand based on the crop growth characteristics and the soil nutrient data; Based on the relationship between the crop growth stage and nutrient demand, a nutrient utilization efficiency maximization model is constructed; The output steps include: Based on the nutrient utilization efficiency maximization model, a precise fertilization decision-making plan is generated; Output the precise fertilization decision plan.

2. The method according to claim 1, characterized in that , the image preprocessing specifically includes: Performing orthorectification, filtering and denoising, radiometric calibration, resampling, atmospheric correction, geometric correction, image registration and fusion on the drone image; Performing atmospheric correction, geometric correction and image fusion on the geostationary satellite image.

3. The method according to claim 1, characterized in that , the feature extraction specifically includes: Based on the drone image, extract terrain slope and elevation data; Based on the stationary satellite image, spectral inversion processing is performed to obtain quantitative remote sensing parameters of vegetation; Based on the vegetation quantitative remote sensing parameters, time series data of vegetation quantitative remote sensing parameters of unmanned aerial vehicle images and time series data of vegetation quantitative remote sensing parameters of geostationary satellite images are constructed.

4. The method according to claim 1, characterized in that ,The construction of the crop identification and classification model specifically includes: Obtain raw image data of different crops in high-resolution satellite images; Annotating the original image data to form a sample set; Dividing the sample set into a training set and a validation set; Convolutional neural network is used to build the initial classification model; Using a back propagation algorithm and a stochastic gradient descent function to minimize the average cross entropy loss function, the parameters of the initial classification model are trained; The validation set is used to evaluate the model accuracy, and when the model reaches a preset accuracy threshold, the crop recognition and classification model is obtained.

5. The method according to claim 1, characterized in that , the extraction of crop growth characteristics specifically includes: Use drones equipped with multispectral digital cameras to obtain near-ground images of farmland; Use a spectrum analyzer to obtain spectrum curves of different crops in different bands; Using an improved U-Net network, the spatial characteristics of different crop types are extracted from the near-ground image to obtain a distribution map of spatial characteristics of different crops; Convolutional neural network is used to build a crop chlorophyll content analysis model; Based on the crop chlorophyll content analysis model, the chlorophyll content index of the crop leaves at each growth stage is calculated.

6. The method according to claim 1, characterized in that , the analysis of the relationship between crop growth stage and nutrient demand specifically includes: Define the crop growth stages and corresponding objective functions; Combining the mapping relationship between chlorophyll content and spectral characteristics, a model for solving the objective function of crops at different growth stages based on transfer learning and convolutional neural network is designed; The stage feature extraction model is constructed. According to the results of crop identification, the objective function of crops in different growth stages is input to calculate the distribution of crop characteristics in different growth stages.

7. The method according to claim 1, characterized in that The construction of the nutrient utilization efficiency maximization model specifically includes: Combine multiple crop objective functions to extract the nutritional characteristics of crops at different stages and determine the nutrient requirements of crops; The improved U-Net network is used to convert the spectral curve characteristics of crops into image features of crop health status; Introducing the temporal attention mechanism and using the reinforcement learning framework to achieve crop nutrient decision-making at different stages; Build a library of precision fertilization solutions based on historical data; According to the knowledge of different agronomic experts, the impact of different fertilization schemes on soil and climate is extracted, and an agronomic expert knowledge distillation model is constructed.

8. The method according to claim 1, characterized in that , also includes the evaluation step: During crop growth and development, the nutrient utilization rates of nitrogen, phosphorus and potassium are monitored separately; When the nutrient utilization rate is lower than 80%, fertilize according to the fertilization rules; Monitor the changes in nutrient utilization rate in real time to determine whether fertilization has ended.

9. The method according to claim 1, characterized in that The meteorological data include wind speed and direction, light intensity, temperature, and humidity; the soil data include soil moisture content and soil pH value, wherein the soil pH value is a parameter of five dimensions, representing acidity coefficient, alkalinity coefficient, salinity coefficient, hardness coefficient, and alkalinity index, respectively.

10. An intelligent decision-making system for precise fertilization based on multi-source remote sensing that implements the method described in any one of claims 1 to 9, characterized in that ,include: Data acquisition unit, used to collect multi-source remote sensing images, soil nutrient data, fertility indicators, meteorological data and soil data; A data preprocessing unit, used for preprocessing and feature extraction of the multi-source remote sensing images; The crop identification and classification unit is used to build a crop identification and classification model, and to identify crop species and divide the preprocessed multi-source remote sensing images into blocks; Crop growth analysis unit, used to extract crop growth characteristics and analyze the relationship between crop growth stage and nutrient demand; Nutrient use efficiency modeling unit, used to build a nutrient use efficiency maximization model; A decision generation unit, used for generating a precise fertilization decision plan based on the nutrient utilization efficiency maximization model; Evaluation feedback unit, used to monitor nutrient utilization and adjust fertilization decisions based on monitoring results; Among them, the system realizes intelligent decision-making on precise fertilization based on multi-source remote sensing through the collaborative work of the above-mentioned units.

Images