Cotton field moisture monitoring model architecture method and system based on unmanned aerial vehicle remote sensing

Abstract

The invention discloses a cotton field moisture monitoring model architecture method and system based on unmanned aerial vehicle remote sensing, and belongs to the technical field of cotton moisture monitoring. The method comprises the following steps: step 1, obtaining an unmanned aerial vehicle image; step 2, batch extraction of vegetation indexes and texture features is carried out; step 3, building a double-loop cognitive learning network; and step 4, inputting the obtained set into a network for monitoring. The invention further provides a cotton field moisture monitoring model architecture system based on unmanned aerial vehicle remote sensing. According to the method, a double-circulation structure of interpretation and refining is constructed, closed-loop optimization of the moisture monitoring model from feature screening to structure compression to performance enhancement is achieved, the method is suitable for a cotton field area with a large area and high data isomerism, and a new thought and reliable support are provided for precise regulation and intelligent management of agricultural water resources.

Description

Technical Field

[0001] This invention belongs to the field of cotton moisture monitoring technology, and relates to a cotton field moisture monitoring model architecture method and system based on UAV remote sensing. Background Technology

[0002] Cotton is one of the world's most important economic crops, possessing both oilseed and fiber value. Particularly in China's Xinjiang Uyghur Autonomous Region, its planting area and yield rank first in the country, making it the region's largest pillar industry. To improve cotton production efficiency and resource allocation, and to achieve precise monitoring and management of its cultivation, advanced information acquisition and intelligent processing methods are urgently needed. Among these, the integration of remote sensing imagery and artificial intelligence technology is becoming a key direction for modern agricultural development.

[0003] In recent years, with the popularization of high-resolution satellite imagery and UAV remote sensing technology, combining deep learning models to classify crops and monitor moisture in remote sensing images has become a research hotspot. Satellite or UAV imagery provides rich spectral and spatial information (such as RGB and NIR bands), while deep learning models can automatically mine semantic features in images to assist in agricultural information extraction tasks such as planting area division and moisture estimation. However, existing technologies still face major challenges in practical large-scale farmland monitoring, especially for single crops (such as cotton): the integration of multi-resolution features is difficult; images acquired by UAVs at different flight altitudes and platforms have inconsistent resolutions; existing models generally lack unified modeling capabilities for cross-scale features, requiring manual batch training or independent design of network structures, which reduces the adaptability and stability of the algorithms. The feature selection process relies on experience and lacks interpretability. Most remote sensing analysis models directly use raw vegetation indices (such as NDVI and GNDVI) or texture indices (such as GLCM) as input, heavily relying on human experience in feature selection, lacking a systematic screening mechanism, and also lacking quantitative evaluation of the actual contribution of various features, making it difficult to meet the needs of agricultural experts for "interpretable decision-making." The traditional classification model has poor transferability and weak generalization performance. During the training process, it does not fully consider the differences in image performance caused by regional differences, year changes and growth period differences, which leads to a significant drop in the accuracy of the model in cross-regional or cross-year applications, thus limiting its practicality.

[0004] To address the above problems, some improvement ideas have been attempted in existing technologies, such as:

[0005] (1) A method and system for controlling cotton moisture monitoring and drip irrigation based on soil water potential, relating to the field of agricultural planting technology, the key points of which are: establishing a daily distribution curve of soil water potential; establishing a daily distribution curve of temperature; calculating the daily consumption distribution curve of the previous day; establishing a correlation curve of unit temperature consumption; calculating the actual daily temperature difference curve; matching historical daily temperature difference segments from historical daily temperature difference curves, and extracting the correlation segment of unit temperature consumption, and recombining the correlation segment of unit temperature consumption to establish a correlation curve of predicted consumption; calculating the predicted daily consumption curve for the current day, and controlling the drip irrigation amount to the monitored object in real time according to the predicted daily consumption curve. This invention predicts the daily consumption of the current day based on historical data, without the need to calculate the drip irrigation amount of the monitored object in real time, with low overall complexity and low network requirements, and can adapt to planting areas with poor network conditions.

[0006] (2) A method for monitoring water content in drip irrigation cotton based on soil water potential aims to use a TRS-II soil water potential meter to dynamically monitor soil water potential, combine it with the water demand pattern of cotton, and adopt a combination of mathematical model and monitoring methods. By periodically measuring the soil water potential value and water demand of cotton in the field, an early warning model for monitoring cotton water demand and irrigation time based on soil water potential is established. The complementary advantages of the early warning model and real-time monitoring technology are brought into play. On this basis, an early warning system for drip irrigation cotton based on soil water potential is established, thereby improving the accuracy and flexibility of farmland irrigation water scheduling, making the recommended results more practical and applicable to a wider range.

[0007] In summary, existing methods have significant shortcomings in cross-resolution feature fusion, feature interpretation and screening, model generalization and structural innovation, making it difficult to meet the needs of large-scale cotton-producing areas such as Xinjiang Uygur Autonomous Region for high-precision, cross-year, and interpretable extraction of cotton planting areas and moisture monitoring.

[0008] Current cotton moisture monitoring technologies primarily rely on soil water potential meters and drip irrigation control systems. These systems monitor changes in physical water potential in the field and combine this data with historical meteorological data or mathematical models to predict and regulate irrigation volume. For example, existing technologies establish correlation models between daily temperature and soil water potential distribution curves and unit temperature consumption, thereby generating drip irrigation prediction curves to guide actual irrigation operations. While these methods have relatively low overall complexity and low network requirements, their core reliance on field sensors places high demands on deployment conditions, equipment stability, and regional coverage. Their promotion and maintenance, especially in vast cotton-growing areas like Xinjiang Uygur Autonomous Region, face practical difficulties. Furthermore, the models themselves are built based on empirical formulas, making them ill-suited to handle sudden climate changes or dynamic variations under different geographical conditions, lacking sufficient flexibility and generalization capabilities.

[0009] On the other hand, current moisture prediction methods do not effectively integrate spatial data such as remote sensing images, resulting in a lack of spatial distribution and visualization capabilities in the monitoring results. They can only achieve point-level data analysis, making it difficult to support accurate decision-making in large-scale planting areas. In addition, although some solutions introduce the linkage between mathematical modeling and monitoring systems, they generally lack intelligent learning capabilities in their model structure and have not established an interpretable evaluation mechanism for the contribution of input variables. This causes the models to still rely on human experience and frequent parameter tuning when transferred or reused, reducing system efficiency and intelligence.

[0010] There is an urgent need for a cotton field moisture monitoring model architecture method and system based on UAV remote sensing. Summary of the Invention

[0011] The purpose of this invention is to solve the aforementioned technical problems and provide a cotton field moisture monitoring model architecture method and system based on UAV remote sensing. It proposes a dual-loop cognitive learning framework to construct a non-contact moisture monitoring system based on remote sensing image data, eliminating reliance on soil sensors and improving the wide-area adaptability and deployment efficiency of cotton field moisture monitoring. The system integrates multi-resolution vegetation indices and texture features. In the first stage, it uses a random forest model and the SHAP method to perform interpretability analysis on feature importance and selects the 10 most informative features through recursive feature elimination (RFE). In the second stage, the selection results are fed back into a lightweight pre-trained model to construct a knowledge distillation mechanism, resulting in the final cross-scale prediction model. This model not only improves prediction accuracy and model interpretability but also significantly enhances its transferability across different flight platforms, imaging scales, and growth stages, solving the problem of strong dependence on manual feature engineering in traditional models.

[0012] This invention constructs a dual-loop structure of "interpretation + refinement," achieving closed-loop optimization of the water monitoring model from feature selection to structural compression and performance enhancement. It is applicable to large-scale cotton fields with highly heterogeneous data, providing new ideas and reliable support for the precise regulation and intelligent management of agricultural water resources.

[0013] The purpose of this invention is to solve the following technical problems:

[0014] 1. Insufficient ability to integrate multi-source features: Traditional algorithms have difficulty in effectively integrating vegetation indices and texture features from different spatial resolutions and time scales, resulting in local biases or poor overall generalization ability in monitoring results;

[0015] 2. Lack of interpretability in the feature selection process: Most current models adopt an end-to-end training approach, which lacks quantitative analysis of the contribution of each input feature, making it difficult for the model to be transferred and applied in different regions and growth stages;

[0016] 3. Feature engineering relies on manual intervention: In traditional methods, feature selection and dimensionality reduction often require human experience and judgment, which lacks systematicity and universality, affecting the automation and adaptability of the model;

[0017] 4. The model is difficult to adapt to multi-resolution datasets: When faced with multi-resolution remote sensing image data, existing monitoring models often need to adjust parameters or structures individually, and cannot form a unified and robust monitoring mechanism.

[0018] Therefore, the purpose of this invention is to propose a dual-loop cognitive learning framework (DCCL), which introduces an interpretable two-stage feature extraction mechanism to achieve automatic screening and reconstruction of key features in UAV images of different resolutions and multiple time periods, thereby improving the accuracy, generalization ability and interpretability of the moisture monitoring model, reducing human intervention, and having good cross-regional and cross-growth period transfer performance.

[0019] To achieve the above objectives, the present invention mainly provides the following technical solutions:

[0020] First, this invention provides a cotton field moisture monitoring model architecture method based on UAV remote sensing, the method comprising the following steps:

[0021] Step 1: Acquire drone imagery;

[0022] Step 2: Perform batch extraction of vegetation index and texture features;

[0023] Step 3: Build a dual-loop cognitive learning network;

[0024] Step 4: Input the obtained set into the network for monitoring.

[0025] Furthermore, in step 1, color images and individual band images were acquired using a DJI Mavic 3 multispectral drone, and the acquired images were processed using DJI Plot and ENVI 5.6. Specifically, the drone flew at ground altitudes of 20m, 40m, and 80m.

[0026] Preferably, in step 1, the UAV flies at a ground altitude of 40m and 80m.

[0027] Furthermore, in step 2, specifically: based on the ENVI 5.6 platform, complete the batch automatic extraction of vegetation index and texture features. Utilize ENVI's image processing capabilities and the automation of IDL scripts, combined with the spectral and spatial information required for cotton moisture monitoring, perform combined calculations and grayscale analysis on the four key bands of green, red, near-infrared and red edge to generate a comprehensive feature input set.

[0028] Furthermore, in step 3, interpretability learning and feature compression optimization are integrated to form two nested learning mechanisms.

[0029] Further, in step 4, (1) the feature subset F2 after SHAP interpretation and RFE screening is used as the final input set, uniformly formatted into a standard feature matrix, and input into the random forest regression model that has been trained in the second stage. For each remote sensing sample pixel or sample point, the model will automatically output its corresponding target prediction value, forming a continuous moisture estimation map of the entire cotton field area.

[0030] (2) The prediction results are spatially mapped in the geographic information system or ENVI platform to display the moisture status of each area in the form of raster images, and can be displayed intuitively in the form of color scale maps and heat maps.

[0031] Second, this embodiment of the invention also provides a cotton field moisture monitoring model architecture system based on UAV remote sensing. The system includes: one or more processors; a memory for storing one or more programs; the processors are configured to execute program instructions stored in the memory, and the program instructions execute the above-described cotton field moisture monitoring model architecture method based on UAV remote sensing when they are executed.

[0032] Third, this embodiment of the invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by one or more processors, implements the above-described cotton field moisture monitoring model architecture method based on UAV remote sensing.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] Traditional cotton moisture monitoring methods generally rely on soil sensors and empirical models, which are not only costly to deploy and have limited coverage, but also poorly adaptable to sudden weather changes and complex terrains. This results in low monitoring accuracy and poor automation in large cotton fields. Even with the introduction of remote sensing images and deep learning methods, existing models often face limitations such as feature selection relying on experience, redundant model structure, poor interpretability, and weak transferability.

[0035] This invention introduces a "dual-loop cognitive learning" mechanism, utilizing a closed-loop structure of interpretation-filtering-refinement composed of random forest, SHAP, and RFE, to achieve automatic integration of multi-resolution remote sensing features, accurate identification of key features, and lightweight optimization of the model structure. Furthermore, by incorporating a knowledge distillation strategy, it effectively enhances the model's robustness and generalizability in cross-regional, cross-temporal, and cross-platform applications. Therefore, this invention significantly improves the model's interpretability, automation level, and field deployment capabilities while maintaining prediction accuracy, making it an ideal solution for cotton field moisture monitoring that combines "accuracy, efficiency, and reusability." Attached Figure Description

[0036] Figure 1 Drones capture images of the land parcels;

[0037] Figure 2 Vegetation index and texture feature acquisition process;

[0038] Figure 3 Network structure diagram (process one);

[0039] Figure 4 Network structure diagram (process two);

[0040] Figure 5 Network monitoring process. Detailed Implementation

[0041] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Embodiments of the present invention provide a model architecture method for cotton field moisture monitoring based on UAV remote sensing, the specific steps of which are as follows:

[0043] The first step is to acquire drone imagery.

[0044] This invention utilizes a DJI Mavic 3 multispectral drone to acquire color images and individual band images, and processes the acquired images using DJI Plot and ENVI 5.6. Specifically, the drone flew at ground altitudes of 20m, 40m, and 80m. In practical applications, we recommend flight altitudes of 40m and 80m for data source fusion. Figure 1 Images of land parcels captured by a drone.

[0045] The second step is to extract vegetation indices and texture features in batches.

[0046] After completing the region cropping of the remote sensing image, this invention enters the feature construction stage, which involves the batch automatic extraction of vegetation indices and texture features based on the ENVI 5.6 platform. This stage fully utilizes ENVI's image processing capabilities and the automation advantages of IDL scripts, combining the spectral and spatial information required for cotton moisture monitoring, and performing combined operations and grayscale analysis on four key bands—green (G), red (R), near-infrared (NIR), and red edge (RE)—to generate a comprehensive feature input set.

[0047] First, relying on the four bands G, R, NIR, and RE, a total of 35 vegetation indices were calculated using ENVI's built-in band calculation tools and a custom script. These indices cover standard vegetation indices (such as NDVI, GNDVI, EVI, SAVI, MSAVI, etc.), red-edge enhancement indices (such as NDRE, RECI), and green enhancement indices (such as GCI, TGI), as well as a series of ratio and difference-based vegetation indices. These indices can effectively reflect the spectral response, water content, and chlorophyll activity of plants, constituting the core spectral characteristics reflecting the health status of cotton.

[0048] Secondly, in terms of texture feature extraction, the system converts the aforementioned bands into grayscale images and applies a 3×3 grayscale co-occurrence matrix (GLCM) window in four directions (0°, 45°, 90°, and 135°) to extract eight types of texture parameters, including contrast, homogeneity, entropy, energy (ASM), correlation, dissimilarity, mean, and variance. The extraction of each texture parameter in the four directions generates a total of 32 texture feature layers, meticulously depicting the spatial structure and vegetation texture features in the cotton field imagery.

[0049] To improve processing efficiency and avoid human error, this invention employs an IDL script to drive the ENVI tool, automating the entire batch processing and feature extraction process for images. The script automatically identifies the input image path and naming rules, batch reads multispectral images and extracts the required bands, sequentially executes preset vegetation index formulas and texture calculation commands, and finally outputs all generated feature layers to a specified path with standardized naming for subsequent model aggregation and reuse. This automation process ensures high consistency in image processing across different flight batches and time points, providing stable, standardized, and structured feature inputs for model training.

[0050] Using the above method, the system ultimately constructs an initial high-dimensional feature set consisting of 35 vegetation indices and 32 texture features, and outputs it in tabular (CSV) or multi-band image stack (HDR Stack) format, serving as the basic data support for subsequent random forest training and feature refinement mechanisms. The process for obtaining vegetation indices and texture features is as follows: Figure 2 As shown.

[0051] The third step is to build a dual-loop cognitive learning network.

[0052] After extracting and normalizing high-dimensional features, this invention enters the modeling stage, namely, constructing a "dual-loop cognitive learning network" to achieve accurate prediction of cotton moisture status. This model framework organically integrates "interpretable learning" and "feature compression optimization," forming two nested learning mechanisms that correspond to the two key processes of feature contribution evaluation and model performance retraining, respectively. With decision transparency, lightweight structure, and strong cross-scale transferability as its core objectives, this network effectively solves the problems of poor interpretability and weak generalization ability of traditional "black box models" in agricultural remote sensing monitoring.

[0053] This invention employs a Random Forest (RF) model as the base learner and introduces the SHAP (SHapley Additive exPlanations) algorithm to analyze the feature importance of the model output. In the first stage, all extracted vegetation indices and texture features are uniformly input into the initial RF model for the first fitting training, used to predict target variables (such as cotton leaf water content or cotton plant moisture index). After the model training is completed, the marginal contribution of the input features is quantified using the SHAP algorithm to obtain the importance score of each feature in the prediction process. The system automatically selects the top 20 features with the highest contribution to form the first feature subset F1.

[0054] In the second stage, the model initiates a recursive feature elimination (RFE) mechanism around the feature subset F1. In this stage, the system stratifies the features within F1 according to SHAP weights and dynamically calculates the overall performance of the feature combination based on the prediction error output by the model. In each iteration, the system removes one or more features with the lowest current weights and re-evaluates the prediction performance of the remaining feature subset. This process is repeated until the optimal feature subset F2 (typically containing about 10 key features) is selected, which maintains prediction accuracy under the condition of dimensionality simplification. This process achieves dynamic dual optimization of model structure and feature dimensionality through a cognitive cycle of "compression-feedback-retraining".

[0055] Finally, the system re-inputs the feature subset F2 into the final constructed random forest model for a second round of training, generating a deployable target prediction model. This model is not only lightweight and has a short training time, but also possesses high interpretability and transferability. In particular, when faced with image data of different resolutions, different growth stages, or images spanning multiple years, the model can still maintain stable performance output, greatly improving its adaptability in real agricultural scenarios.

[0056] By constructing this dual-loop cognitive network, this invention not only achieves closed-loop self-optimization from original features to the final model, but also makes innovative breakthroughs in several aspects such as interpretive learning, feature compression, and cross-scale fusion, providing an efficient, accurate, and reusable solution for large-scale, heterogeneous remote sensing data-driven crop water monitoring. The network structure diagram is shown below. Figure 3 , Figure 4 As shown.

[0057] The fourth step is to input the obtained set into the network for monitoring.

[0058] After constructing and optimizing the dual-loop cognitive learning network, this invention enters the final application stage. The optimal feature set obtained through feature refinement is input into the final model to accurately monitor and output the moisture status of cotton. This stage realizes the actual deployment and monitoring functions of the model and is a crucial link in the entire system's transition from algorithm research to application scenarios.

[0059] First, the feature subset F2, interpreted by SHAP and filtered by RFE, is used as the final input set, uniformly formatted into a standard feature matrix, and input into the random forest regression model trained in the second stage. The model structure has been adapted and compressed according to the optimal feature set, resulting in shorter training and inference times and enabling efficient operation under ordinary computing resources. For each remote sensing sample pixel or sample point, the model will automatically output its corresponding target prediction value (such as the equivalent water thickness of the cotton canopy), forming a continuous moisture estimation map of the entire cotton field area.

[0060] Secondly, the prediction results can be spatially mapped on a Geographic Information System (GIS) or ENVI platform to display the moisture status of each area in the form of raster images, supporting intuitive display in the form of color scale maps, heat maps, etc. This visualization output method not only provides agricultural managers with intuitive decision-making references, but also truly transforms remote sensing data into "interpretable, operable, and applicable" intelligent agricultural information.

[0061] Furthermore, because the final model's input is an automatically selected, multi-scale adapted fusion feature set, the system possesses excellent cross-regional and cross-temporal transfer capabilities. Without retraining, the model can be directly applied to cotton remote sensing images from other flight batches, different plots, and even different years; "ready-to-use" prediction deployment can be achieved simply by completing the feature extraction process. This feature significantly reduces model deployment and maintenance costs, enhancing its practicality and promotional value.

[0062] In summary, this monitoring phase completes a closed loop from raw image acquisition, feature extraction, feature refinement, model training to spatial prediction result output, demonstrating the high automation, strong interpretability, and wide adaptability of the system design of this invention, and providing an efficient and reliable technical means for large-scale, precise monitoring of cotton field moisture dynamics.

[0063] This invention also provides a cotton field moisture monitoring model architecture system based on UAV remote sensing, including:

[0064] It includes: one or more processors; a memory for storing one or more programs; the processors are configured to execute program instructions stored in the memory, which, when executed, perform the above-described cotton field moisture monitoring model architecture method based on UAV remote sensing.

[0065] This invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by one or more processors, implements the above-described cotton field moisture monitoring model architecture method based on UAV remote sensing.

[0066] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0067] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0068] In addition, the memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0069] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0070] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0071] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0072] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0073] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0074] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0075] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0076] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0077] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0078] Key technical points include, but are not limited to, the following aspects:

[0079] 1. Feature fusion and extraction mechanism

[0080] A multi-index vegetation index and texture feature fusion method based on four key bands (G, R, NIR, RE) and a batch processing technique using the ENVI platform to automatically extract 35 vegetation indices and 32 texture features. This feature fusion process achieves joint representation of spectral information and spatial structure, providing a rich input foundation for subsequent models.

[0081] 2. Two-stage interpretable learning and feature compression mechanism

[0082] This paper innovatively introduces a dual-loop cognitive structure consisting of the SHAP feature interpretation method and the RFE feature elimination algorithm. The first stage of model training obtains feature importance, and then recursively filters and optimizes based on this, effectively reducing model dimensionality and improving prediction performance and interpretability.

[0083] 3. Knowledge distillation-driven model retraining scheme

[0084] The second round of random forest model was retrained using a refined subset of features, resulting in a lightweight and stable final moisture monitoring model. Compared to the initial model, this model has better generalization ability, transferability, and deployment flexibility, making it suitable for multi-scenario and multi-resolution remote sensing data analysis.

[0085] 4. Adaptive design across scales and time

[0086] The model does not need to be remodeled for different remote sensing platforms or different growth stages. It can quickly complete the monitoring deployment by simply re-inputting images and performing standard feature extraction procedures, which significantly reduces the cost of manual intervention and training.

[0087] 5. Fully automated deployment path

[0088] This invention automates the entire process from data preprocessing, feature extraction, feature selection to model output. It is suitable for rapid processing and model reuse of large batches of remote sensing images and has high practical value in scenarios such as dynamic monitoring of farmland moisture and intelligent irrigation control.

[0089] In summary, this invention takes the "feature interpretation + compression + distillation" three-in-one cognitive learning structure as its core, integrating remote sensing information processing, machine learning model training and interpretability mechanisms, breaking through the bottlenecks of traditional models that are "uninterpretable, difficult to transfer, and heavily reliant on human intervention". It has significant technological innovation and engineering promotion value, and the above key contents should be the core technical solution protected by this patent.

[0090] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Furthermore, although some specific terms are used in this specification, these terms are merely for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. A cotton field moisture monitoring model architecture method based on UAV remote sensing, characterized in that, The method includes the following steps: Step 1: Acquire drone imagery; Step 2: Perform batch extraction of vegetation index and texture features; Step 3: Build a dual-loop cognitive learning network; Step 4: Input the obtained set into the network for monitoring.

2. The method according to claim 1, characterized in that, In step 1, color images and individual band images were acquired using a DJI Mavic 3 multispectral drone, and the acquired images were processed using DJI Plot and ENVI 5.

6. The specific drone flight altitudes above the ground were 20m, 40m, and 80m.

3. The method according to claim 2, characterized in that, In step 1, the drone flies at altitudes of 40m and 80m above the ground.

4. The method according to claim 1, characterized in that, In step 2, specifically: based on the ENVI 5.6 platform, the batch automatic extraction of vegetation index and texture features is completed. Utilizing ENVI's image processing capabilities and the automation of IDL scripts, combined with the spectral and spatial information required for cotton moisture monitoring, the four key bands of green, red, near-infrared and red edge are combined and analyzed in grayscale to generate a comprehensive feature input set.

5. The method according to claim 1, characterized in that, In step 3, interpretability learning and feature compression optimization are integrated to form two nested learning mechanisms.

6. The method according to claim 1, characterized in that, Step 4 specifically includes the following steps: (1) The feature subset F2 after SHAP interpretation and RFE screening is used as the final input set, uniformly formatted into a standard feature matrix, and input into the random forest regression model that has been trained in the second stage. For each remote sensing sample pixel or sample point, the model will automatically output its corresponding target prediction value to form a continuous moisture estimation map of the entire cotton field area. (2) The prediction results are spatially mapped in the geographic information system or ENVI platform to display the moisture status of each area in the form of raster images, and can be displayed intuitively in the form of color scale maps and heat maps.

7. A cotton field moisture monitoring model architecture system based on UAV remote sensing, characterized in that, include: It includes: one or more processors; a memory for storing one or more programs; the processors are configured to execute program instructions stored in the memory, which, when executed, perform the cotton field moisture monitoring model architecture method based on UAV remote sensing as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program, when executed by one or more processors, implements the cotton field moisture monitoring model architecture method based on UAV remote sensing as described in any one of claims 1 to 6.

Images