South Xinjiang machine-harvested cotton efficient planting variety screening method and system based on deep learning

Abstract

The invention discloses a southern Xinjiang machine-harvested cotton efficient planting variety screening method and system based on deep learning, and the method comprises the steps: obtaining image data, environment data, mechanical state data and product quality data in a working process of a southern Xinjiang machine through a sensor network, carrying out the preprocessing of the collected data, and obtaining a multi-source data set; performing semantic segmentation on the plant type compactness and the leaf shielding rate on the image data in the multi-source data set by adopting an improved model to obtain image features, and performing cross-modal correlation analysis on the environment data, the mechanical state data and the product quality data through the model to obtain correlation features; fusing the image features and the associated features through a cross-modal attention mechanism, inputting the fused image features and associated features into a stochastic gradient descent model, and outputting a southern Xinjiang mechanical mining adaptation index; and adjusting a variety screening threshold value according to the South Xinjiang mechanical picking adaptation index, and generating a South Xinjiang mechanical picking cotton control instruction according to the screening threshold value for real-time control. The mechanical mining efficiency is improved, and the standard-reaching rate of fiber quality is improved.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a method and system for screening efficient cotton varieties for machine harvesting in southern Xinjiang based on deep learning. Background Technology

[0002] Southern Xinjiang is a major cotton-producing area in my country, and the promotion of mechanized cotton harvesting is crucial for improving cotton production efficiency. Currently, cotton variety selection relies heavily on field trials and manual evaluation, which suffers from problems such as long cycles, high costs, and strong subjectivity. Although existing technologies have attempted to utilize remote sensing or image analysis for crop monitoring, intelligent selection methods for mechanized cotton varieties that are tailored to the unique climate and soil conditions of southern Xinjiang and combine agronomic and agricultural machinery integration requirements remain lacking. Furthermore, traditional methods struggle to systematically integrate multi-dimensional indicators such as cotton growth process, stress resistance, mechanized harvesting suitability, yield, and quality, making it impossible to achieve efficient and accurate variety selection decisions. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a method and system for screening efficient cotton varieties for machine harvesting in southern Xinjiang based on deep learning.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned deep learning-based method for screening high-efficiency varieties of machine-harvested cotton in southern Xinjiang: Image data, environmental data, mechanical status data, and product quality data during the operation of the machine in southern Xinjiang are acquired through sensor networks. The collected data are preprocessed to obtain a multi-source dataset. An improved ResNet-50 model was used to perform semantic segmentation of plant compactness and leaf occlusion rate on image data from multi-source datasets to obtain image features. The Transformer model was then used to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain correlation features. After fusing the image features and associated features through a cross-modal attention mechanism, the data is input into the XGBoost stochastic gradient descent model, and the output is the Southern Xinjiang machine harvesting adaptation index. Based on the Q-learning algorithm, the variety screening threshold is adjusted according to the Southern Xinjiang machine-harvesting adaptability index, and control instructions for Southern Xinjiang machine-harvested cotton are generated according to the screening threshold for real-time control.

[0005] Furthermore, in the aforementioned deep learning-based method for selecting efficient cotton varieties for machine harvesting in southern Xinjiang, the process of acquiring image data, environmental data, mechanical status data, and product quality data during the operation of the cotton harvester in southern Xinjiang via a sensor network, and preprocessing the collected data to obtain a multi-source dataset, including: Image data, environmental data, mechanical status data, and product quality data during the operation of the machine in southern Xinjiang are acquired through sensor networks. The 3σ criterion is used to filter out abrupt values ​​in environmental data. The variance is calculated using the gradient operator. Images with a variance <100 are judged as blurry and discarded, thus filtering out invalid visual data. Linear interpolation was used to fill in short-term missing data in environmental data, and missing data in image data were filled in using adjacent time periods of the same plot and growth period.

[0006] Furthermore, in the aforementioned deep learning-based method for selecting efficient cotton varieties for machine harvesting in southern Xinjiang, the process of acquiring image data, environmental data, mechanical status data, and product quality data during the operation of the cotton harvester in southern Xinjiang via a sensor network, and preprocessing the collected data to obtain a multi-source dataset, including: The LabelMe annotation tool was used to perform semantic segmentation and annotation on cotton plant images. Class labels were defined for compact plant areas and leaf occlusion areas. The annotated images were then bound to the variety, growth process, and environmental data of the corresponding plots to obtain preprocessed data. The environmental data, mechanical condition data, and product quality data in the preprocessed data are mapped to the [0,1] interval through Min-Max normalization. After converting the multispectral image into an RGB pseudo-color image, mean and variance are standardized to obtain a multi-source dataset.

[0007] Furthermore, in the aforementioned deep learning-based method for selecting efficient cotton varieties for machine harvesting in southern Xinjiang, the improved ResNet-50 model is used to perform semantic segmentation of image data from multi-source datasets based on plant compactness and leaf occlusion rate to obtain image features. Then, a Transformer model is used to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain correlation features, including: After the residual blocks in stages 2, 3, and 4 of ResNet-50, a CBAM channel attention module is embedded to obtain an improved ResNet-50 model. The preprocessed cotton plant image is input into the improved ResNet-50 model, and the semantic segmentation result is output through the segmentation head of U-Net+ residual block to obtain image features.

[0008] Furthermore, in the aforementioned deep learning-based method for selecting efficient cotton varieties for machine harvesting in southern Xinjiang, the improved ResNet-50 model is used to perform semantic segmentation of image data from multi-source datasets based on plant compactness and leaf occlusion rate to obtain image features. Then, a Transformer model is used to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain correlation features, including: Environmental data, mechanical condition data, and product quality data are mapped into numerical feature vectors through a fully connected layer; The Transformer model employs a 6-layer encoder-decoder architecture, with 8 multi-head attention heads, 256 hidden layer dimensions, and a dropout rate of 0.2. Using machine-collected morphological image features as query vectors and numerical feature vectors as key-value pairs, we learn multi-dimensional relationships and output associated features.

[0009] Furthermore, in the aforementioned deep learning-based method for selecting efficient cotton varieties for machine harvesting in southern Xinjiang, the process of fusing image features and associated features through a cross-modal attention mechanism and inputting them into an XGBoost stochastic gradient descent model to output a southern Xinjiang machine harvesting suitability index includes: A cross-modal attention layer is constructed to calculate the mutual information between image features and associated features. Dynamic weights are assigned to features with high mutual information. Image features and associated features are then concatenated according to the dynamic weights to obtain a comprehensive feature vector. Based on historical data of machine-harvested varieties, varieties are divided into high-fitness, medium-fitness, and low-fitness categories, which serve as training labels for the XGBoost stochastic gradient descent model. The model outputs scores for five sub-indicators, including scores for growth process suitability, stress resistance suitability, machine harvesting trait suitability, yield suitability, and quality suitability. These scores are weighted according to their respective weights to obtain the suitability index for machine-harvested cotton varieties in southern Xinjiang.

[0010] Furthermore, in the aforementioned deep learning-based method for selecting efficient varieties of machine-harvested cotton in southern Xinjiang, the step of adjusting the variety selection threshold based on the Q-learning algorithm according to the southern Xinjiang machine-harvesting suitability index, and generating control commands for machine-harvested cotton in southern Xinjiang based on the selection threshold for real-time control, includes: Using the current batch of varieties' compatibility index distribution and demonstration area environmental data as the state space, the range of adjustment of the variety selection threshold as the action space, and the actual planting effect in the demonstration area as the reward value, the variety selection threshold is adjusted according to the aforementioned Southern Xinjiang machine harvesting compatibility index. Based on the screening threshold, control instructions for machine-harvested cotton in southern Xinjiang are generated, including at least plant and row spacing configuration, water and fertilizer management, and machine harvesting parameters. The control commands for machine-harvested cotton in southern Xinjiang are synchronized to the machine operation system in southern Xinjiang via the LoRa wireless communication protocol.

[0011] Furthermore, in the deep learning-based high-efficiency cotton variety selection system for machine harvesting in southern Xinjiang, the system includes the following modules: The multi-source data acquisition module is used to acquire image data, environmental data, mechanical status data and product quality data during the operation of the machine in southern Xinjiang through a sensor network, and to preprocess the acquired data to obtain a multi-source dataset. The associated feature generation module is used to perform semantic segmentation of image data in multi-source datasets by using an improved ResNet-50 model to obtain image features such as plant compactness and leaf occlusion rate. The module then uses a Transformer model to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain associated features. The adaptation index calculation module is used to fuse the image features and associated features through a cross-modal attention mechanism, input them into the XGBoost stochastic gradient descent model, and output the adaptation index of the Southern Xinjiang machine harvesting. The real-time control module for harvesting cotton in southern Xinjiang is used to adjust the variety screening threshold based on the Q-learning algorithm and the southern Xinjiang machine-harvesting adaptation index, and to generate control commands for harvesting cotton in southern Xinjiang in real time based on the screening threshold.

[0012] Furthermore, in the deep learning-based screening system for efficient machine-harvested cotton varieties in southern Xinjiang, the adaptation index calculation module includes the following sub-modules: The splicing submodule is used to construct a cross-modal attention layer, calculate the mutual information between image features and associated features, assign dynamic weights to features with high mutual information, and splice the image features and associated features according to the dynamic weights to obtain a comprehensive feature vector. The training submodule is used to classify varieties into high-fitness, medium-fitness, and low-fitness based on historical data of machine-harvested varieties, and use these as training labels for the XGBoost stochastic gradient descent model. The submodule is used to output scores for five sub-indicators of the model, including the adaptation score for the growth process, the adaptation score for the stress resistance, the adaptation score for the machine harvesting trait, the adaptation score for the yield, and the adaptation score for the quality. The adaptation index of machine harvested cotton varieties in southern Xinjiang is obtained by weighting the indicators according to their weights.

[0013] Furthermore, in the deep learning-based high-efficiency cotton harvesting variety selection system for southern Xinjiang, the real-time control module for southern Xinjiang machines includes the following sub-modules: Define a submodule, which uses the current batch of varieties' adaptation index distribution and demonstration area environmental data as the state space, adjusts the variety screening threshold by a certain range as the action space, and uses the actual planting effect in the demonstration area as the reward value. Adjust the variety screening threshold according to the Southern Xinjiang machine harvesting adaptation index. The instruction submodule is used to generate control instructions for machine-harvested cotton in southern Xinjiang based on the screening threshold, including at least plant spacing configuration, water and fertilizer management, and machine harvesting parameters; The control submodule is used to synchronize the control commands for machine-harvested cotton in southern Xinjiang to the machine operation system in southern Xinjiang via the LoRa wireless communication protocol.

[0014] Its beneficial effects lie in the fact that by improving the ResNet-50 model and cross-modal association analysis, it accurately quantifies key harvesting traits such as plant compactness, improving the accuracy of the adaptation index by more than 20%, ensuring a high degree of adaptation between the selected varieties and the regional environment and harvesting process in southern Xinjiang, and significantly improving the accuracy of variety selection. It achieves efficient utilization of multi-source data, integrating image, environmental, agricultural machinery, and quality data, and mining potential correlations between data to provide comprehensive data support for variety optimization, while reducing invalid data redundancy, improving data utilization, and lowering data collection and processing costs. It dynamically optimizes planting management strategies, adjusting screening thresholds in real time through the Q-learning algorithm to generate personalized planting control instructions, improving harvesting efficiency, increasing fiber quality compliance rate, and increasing yield per plant by more than 10%, effectively solving the planting adaptation problems in the salinized, temperature and humidity fluctuating environments of cotton fields in southern Xinjiang. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the method for screening efficient planting varieties of machine-harvested cotton in southern Xinjiang based on deep learning in this invention. Figure 2 This is a schematic diagram of the second embodiment of the method for screening efficient planting varieties of machine-harvested cotton in southern Xinjiang based on deep learning in this invention. Figure 3 This is a schematic diagram of the first embodiment of the deep learning-based high-efficiency cotton variety selection system for machine harvesting in southern Xinjiang, as described in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1As shown, a deep learning-based method for selecting high-efficiency varieties of machine-harvested cotton in southern Xinjiang is presented. This method includes the following steps: Step 101: Acquire image data, environmental data, mechanical status data, and product quality data during the operation of the machine in southern Xinjiang through a sensor network. Preprocess the collected data to obtain a multi-source dataset. Specifically, in this embodiment, image data, environmental data, mechanical status data, and product quality data during the operation of the cotton plant in southern Xinjiang are acquired through a sensor network. The 3σ criterion is used to filter out abrupt values ​​in the environmental data, and the variance is calculated using a gradient operator. Images with a variance <100 are deemed blurry and discarded, filtering out invalid visual data. Linear interpolation is used to fill in short-term missing data in the environmental data, and missing image data is filled in using data from adjacent time periods of the same plot and growth stage. The LabelMe annotation tool is used to perform semantic segmentation and annotation on cotton plant images, defining class labels for compact plant areas and leaf-occluded areas. The annotated images are then bound to the corresponding plot's variety, growth stage, and environmental data to obtain preprocessed data. The environmental data, mechanical condition data, and product quality data in the preprocessed data are mapped to the [0,1] interval through Min-Max normalization. After converting the multispectral image into an RGB pseudo-color image, mean and variance are standardized to obtain a multi-source dataset.

[0020] The choice of a combination of 4K high-definition field cameras and multispectral drones for visual sensors was based on the fact that cotton fields in southern Xinjiang are mostly large, contiguous plots. The camera deployment spacing of 50m and a height of 2.5m ensured coverage of the entire cotton plant height from emergence to boll opening (a typical height for machine-harvested cotton in southern Xinjiang is 1.8-2.2m), while also avoiding the loss of plant detail due to excessively high shooting angles. The multispectral drone's flight altitude of 80m and cruising speed of 5m / s were the optimal balance between coverage efficiency and image resolution. Data collection was focused on four key growth stages: emergence (April and budding), flowering (May and flowering), boll opening (June and boll opening), and September. These stages correspond to the research's focus on seedling stability, fruit branch differentiation ability, pollination efficiency, and machine harvesting suitability. Data collection was conducted at 10:00 AM and 4:00 PM daily to account for the large diurnal temperature range in southern Xinjiang and the potential for overexposure due to strong midday sunlight; the light intensity during these times was more suitable for extracting detailed features such as plant shape and leaf distribution.

[0021] The selection and deployment of environmental sensors are also tailored to the specific conditions in southern Xinjiang: the EC-5 soil moisture sensor is adapted to the sandy loam soil texture of southern Xinjiang and can accurately measure 0-100% volumetric water content; the SHT35 temperature and humidity sensor, with an accuracy of ±0.3℃ / ±2%RH, can capture sudden changes in temperature and humidity caused by late spring frosts and hot, dry summer winds in southern Xinjiang; the SX712 salinity sensor is designed to address the common salinization problem in cotton fields in southern Xinjiang, and its 0-14 pH measurement range covers the local soil pH range of 6.5-8.5. The deployment method of setting up one monitoring point per 20 mu (approximately 1.3 hectares) at a depth of 15cm ensures both spatial representativeness of the data (as cotton fields in southern Xinjiang are contiguous and soil conditions vary little within 20 mu), and accurate collection of soil moisture and salinity data from the main water-absorbing layer of cotton plant roots. Data collection every 30 minutes allows for real-time recording of dynamic changes in environmental factors, and simultaneous binding of plot and variety information is for establishing a subsequent correlation between variety and environment.

[0022] Hall effect speed sensors and strain gauge pressure sensors were chosen for the agricultural machinery because there is a lot of dust in the cotton fields of southern Xinjiang during the harvesting season. Hall effect sensors are resistant to dust interference, and the ±0.5kPa accuracy of strain gauge sensors can accurately capture pressure changes in the cotton harvester's clamping device. These sensors are installed on the harvesting drum shaft, clamping device, and drive wheel because the condition of these three components directly affects harvesting efficiency and the degree of cotton fiber damage. The 10Hz sampling frequency can record fluctuations in drum speed in real time, such as speed changes when encountering compact cotton plants. The HVI1000 fiber analyzer was used for quality and yield data acquisition because this equipment is the industry standard for cotton quality testing and can accurately measure the fiber length, breaking strength, and other indicators required by the project.

[0023] The 3σ criterion for data cleaning is applied to remove extreme weather conditions in southern Xinjiang, such as sudden temperature and humidity fluctuations caused by short-term gusts. Gradient variance determination of image sharpness (variance <100 indicates blurriness) avoids image feature distortion caused by wind and sandstorms. Linear interpolation is used for missing values, filling in short-term missing environmental data and data from the same plot at the same growth stage. Image data is used because the environment and cotton plant status in the same plot of cotton fields in southern Xinjiang have strong spatiotemporal consistency. Data annotation uses the LabelMe tool to define compact plant regions, areas with an angle between the main stem and fruiting branches <30°, and areas obscured by leaves, with leaves covering >50% of the cotton bolls. These two indicators are core to machine-harvested traits. Binding annotated images to data such as variety and growth stage creates sample units with multi-dimensional characteristics. In data normalization, Min-Max normalization maps environmental and agricultural machinery data to the [0,1] interval, solving the problem of large differences in the numerical range of different indicators, such as soil moisture 0-100% and salinity 0-14. The mean and variance standardization of image data uses ImageNet pre-trained parameters because the subsequent improvement of the ResNet-50 model is based on ImageNet pre-training, which can improve the model's convergence speed. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio, with the training set covering two plots and the validation set using a separate plot. This is to verify the model's generalization ability in different plots, which aligns with the promotion needs of the three demonstration bases in the project.

[0024] Step 102: Use the improved ResNet-50 model to perform semantic segmentation on the image data in the multi-source dataset to obtain image features based on plant compactness and leaf occlusion rate. Then, use the Transformer model to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain correlation features. Specifically, in this embodiment, a CBAM channel attention module is embedded after the residual blocks in stages 2, 3, and 4 of the ResNet-50 model to obtain an improved ResNet-50 model. The preprocessed cotton plant image is input into the improved ResNet-50 model, and the semantic segmentation result is output through the segmentation head of U-Net+residual block to obtain image features. Environmental data, mechanical condition data, and product quality data are mapped into numerical feature vectors through fully connected layers. The Transformer model adopts a 6-layer encoder-decoder architecture with 8 multi-head attention heads, 256 hidden layer dimensions, and a dropout rate of 0.2. Using machine-collected trait image features as query vectors and numerical feature vectors as key-value pairs, multi-dimensional correlations are learned, and correlation features are output.

[0025] ResNet-50 was chosen as the base model because it solves the gradient vanishing problem in deep networks through residual blocks, effectively extracting deep features from cotton plant images. The CBAM channel attention module is embedded after the residual blocks in stages 2, 3, and 4, which addresses the core needs of machine-harvested cotton in southern Xinjiang. The channel attention branch enhances the channel weights corresponding to compact plant areas through global average pooling and fully connected layers, while the spatial attention branch accurately locates the spatial position of leaf-occluded areas through channel pooling and 3×3 convolution. The combination of these two approaches improves the model's sensitivity to key traits in machine harvesting.

[0026] Semantic segmentation employs a U-Net segmentation head with residual blocks because the U-Net encoder-decoder structure preserves detailed image information, while residual blocks further improve segmentation accuracy, ensuring more accurate boundary recognition between compact plant areas and leaf-occluded areas. The formula for calculating plant compactness—compact area pixels / total plant pixel area—transforms the abstract concept of plant compactness into a 0-1 quantifiable index. The formula for calculating leaf occlusion rate—leaf pixel area occluding bolls / total boll pixel area—quantifies the degree of leaf occlusion on the boll. The 1024-dimensional vector from the last fully connected layer of the model is extracted as the machine-collected trait image feature because this dimension retains sufficient plant and leaf feature information without causing computational redundancy due to excessive dimensionality.

[0027] Non-image data encoding maps 4D environmental data, 3D agricultural machinery status data, and 5D quality and yield data to 64-dimensional numerical feature vectors through fully connected layers. This dimensionality choice is a result of balancing feature information preservation and computational efficiency; 64 dimensions can cover the core features of various data types without increasing the computational burden on the Transformer due to excessive dimensionality. The Transformer adopts a 6-layer encoder-decoder architecture and 8-head multi-attention because the 6-layer structure can fully learn the correlations between multi-dimensional data, while the 8-head attention can simultaneously capture the correlations between different dimensions, such as environment-yield and agricultural machinery-quality. The 256-dimensional hidden layers and a dropout rate of 0.2 ensure the model's expressive power while avoiding overfitting.

[0028] Cross-modal association learning uses image features of harvested traits as query vectors and numerical features as key-value pairs to uncover the intrinsic relationships between image features and other data. For example, the association between soil moisture >80% and roller speed >1500 r / min and decreased fiber breaking strength corresponds to the actual situation of reduced fiber toughness in cotton plants under high soil moisture conditions in southern Xinjiang cotton fields; the association between plant compactness >0.8 and leaf shading rate <0.2 and increased harvesting efficiency by 20% directly corresponds to the research goal of high-efficiency harvesting. The final output of 256-dimensional multi-source data association features integrates the association information between environmental, agricultural machinery, quality, and other data and harvested traits, providing multi-dimensional support for subsequent feature fusion.

[0029] Step 103: After fusing the image features and associated features through a cross-modal attention mechanism, the results are input into the XGBoost stochastic gradient descent model, and the output is the Southern Xinjiang machine harvesting adaptation index. Specifically, in this embodiment, a cross-modal attention layer is constructed to calculate the mutual information between image features and associated features. Dynamic weights are assigned to features with high mutual information. Image features and associated features are then concatenated according to the dynamic weights to obtain a comprehensive feature vector. Based on historical data of machine-harvested varieties, varieties are divided into high-fit, medium-fit, and low-fit categories, which serve as training labels for the XGBoost stochastic gradient descent model. The model outputs scores for five sub-indicators, including fit scores for growth process, stress resistance, machine-harvested traits, yield, and quality. These scores are weighted according to the indicator weights to obtain the fit index for machine-harvested cotton varieties in southern Xinjiang.

[0030] The core of constructing a cross-modal attention layer is to assign higher weights to more important features. For example, the mutual information between plant compactness and machine harvesting efficiency is 0.7, indicating a strong correlation between the two, so a weight of 0.3 is assigned; the mutual information between fiber length and quality is 0.6, so a weight of 0.25 is assigned. This weight setting not only meets the research requirements of prioritizing machine harvesting while taking into account yield and quality, but also conforms to the actual production logic of machine-harvested cotton in southern Xinjiang.

[0031] The 1024-dimensional image features and 256-dimensional associated features are dynamically weighted and concatenated to obtain a 1280-dimensional comprehensive feature vector. The advantage of this fusion method is that it not only preserves the details of key harvesting traits such as plant type and leaves in the image features, but also integrates the multi-dimensional relationships of environment, variety, agricultural machinery, and quality in the associated features. This avoids the limitations of a single feature dimension. For example, image features alone cannot reflect the stress resistance of a variety, such as salt and alkali tolerance, and numerical features alone cannot reflect the adaptability of plant type to machine harvesting. The comprehensive feature vector, on the other hand, achieves full-dimensional coverage of machine harvesting traits, stress resistance, yield, and quality.

[0032] The XGBoost model was chosen because it excels at handling nonlinear data and imbalanced samples. The cotton variety data from southern Xinjiang exhibits an imbalance, with most varieties showing only moderate fit and a few showing high fit, making XGBoost suitable for classification and regression. In the model parameter settings, a learning rate of 0.1 balances training speed and model accuracy; 100 decision trees adequately fit the complex relationships in the data; a maximum tree depth of 6 prevents overfitting; and a subsample ratio of 0.8 increases the model's generalization ability through random sampling.

[0033] The training labels are defined as high fit, 1, medium fit, 0.5, low fit, and category 03, based on the research objective of selecting highly fit varieties. The fit index calculation outputs scores for 5 sub-indicators, with weights directly aligned with the research's assessment priorities: machine harvesting traits account for 30%, emphasizing efficient machine harvesting; yield accounts for 25%, ensuring high yields in cotton fields in southern Xinjiang; quality accounts for 20%, conforming to the GB1103-2022 national cotton standard; growth process accounts for 15%, given the short frost-free period in southern Xinjiang, requiring varieties to adapt their growth process to the local climate; and stress resistance accounts for 10%, necessitating varieties with stress resistance due to the saline-alkali and arid environment of southern Xinjiang. The final fit index for machine-harvested cotton varieties in southern Xinjiang is defined as a perfect score of 100, with ≥80 considered high fit. This quantifies the fit of varieties and directly corresponds to the research objective of selecting 1-2 highly fit varieties.

[0034] Step 104: Based on the Q-learning algorithm, adjust the variety screening threshold according to the Southern Xinjiang machine-harvesting adaptation index, and generate Southern Xinjiang machine-harvested cotton control instructions for real-time control according to the screening threshold.

[0035] Specifically, in this embodiment, the current batch of varieties' compatibility index distribution and demonstration area environmental data are used as the state space, the range of adjusting the variety selection threshold is used as the action space, and the actual planting effect in the demonstration area is used as the reward value. The variety selection threshold is adjusted according to the Southern Xinjiang machine harvesting compatibility index. The Southern Xinjiang machine harvesting cotton control command is generated according to the selection threshold, which includes at least the plant and row spacing configuration, water and fertilizer management, and machine harvesting parameters. The Southern Xinjiang machine harvesting cotton control command is synchronized to the Southern Xinjiang machine operation system through the LoRa wireless communication protocol.

[0036] The state space is defined as the fit index distribution of the current batch of varieties, including the mean, variance, and environmental data from the demonstration area, such as soil moisture, temperature, and humidity. This is because the environment in cotton fields in southern Xinjiang fluctuates significantly, such as differences in soil moisture across different years, requiring dynamic adjustments to the screening threshold based on the environment. The action space is defined as five ranges for adjusting the screening threshold: ±5, ±3, and 0. This is because excessively large ranges can cause the screening results to deviate from the optimal range, while excessively small ranges result in low optimization efficiency. These five actions cover a reasonable adjustment range while ensuring the accuracy of the optimization.

[0037] The machine harvesting efficiency improvement rate accounts for 40%, the core of the project is high machine harvesting efficiency, yield improvement rate accounts for 30%, ensuring high yield and quality compliance rate accounts for 20%, compliance with national standards and stress resistance compliance rate accounts for 10%, and adaptation to the environment of southern Xinjiang. Among them, the quality compliance rate is based on GB1103-2022 level 2 standard, and the stress resistance compliance rate is based on ≥80 points. This setting ensures that the reward value can objectively reflect the actual performance of the variety in the demonstration base. The threshold iteration optimization adopts the ε-greedy strategy, ε=0.1, to balance the use of existing optimal thresholds and the exploration of new thresholds; the learning rate of Q value update α=0.2 and the discount factor γ=0.9. α controls the magnitude of each update, while γ focuses on the impact of future rewards. After 50 iterations, the threshold tends to stabilize and can adapt to the environment and planting needs of the southern Xinjiang demonstration base.

[0038] The plant spacing configuration is chosen as 66cm row spacing + 10cm plant spacing because the picking width of cotton harvesters in southern Xinjiang is mostly a multiple of 66cm. This configuration achieves a high degree of consistency between plant spacing and planting density, while ensuring a planting density of about 10,000 plants per acre, which is suitable for the water and fertilizer carrying capacity of the sandy loam soil in southern Xinjiang. The water and fertilizer management instructions specify drip irrigation once every 10 days during the budding stage, with each irrigation providing 30m³ of water. 3 The ratio of N:P:K = 3:1:2 per mu (unit of area) is chosen because the budding stage is a critical period for cotton plants to require water and fertilizer. Southern Xinjiang is arid and requires drip irrigation to supplement water. Nitrogen promotes fruit branch differentiation, while phosphorus and potassium improve the quality of cotton bolls. This ratio is in line with the water and fertilizer management experience of machine-harvested cotton in southern Xinjiang. The machine harvesting parameters of 1200 r / min for the picking drum and 5 kPa for the clamping pressure are chosen because these parameters can ensure harvesting efficiency while avoiding fiber damage caused by excessive drum speed and pressure, which is in line with the project's goal of achieving synergistic efficiency in yield and quality.

[0039] The LoRa wireless communication protocol is used for instruction synchronization because the demonstration bases in southern Xinjiang are mostly located in remote areas with insufficient 4G signal coverage. LoRa's transmission distance can reach 10km, enabling stable communication of field equipment. The yield and quality data after execution are fed back to the dataset in step 1, forming a closed-loop optimization of collection-analysis-screening-planting-feedback. This not only continuously improves the accuracy of variety selection but also provides data support for the demonstration and promotion of the project on more than 50,000 mu, promoting the intelligent upgrading of the machine-harvested cotton planting model in southern Xinjiang.

[0040] Its beneficial effects lie in the fact that by improving the ResNet-50 model and cross-modal association analysis, it accurately quantifies key harvesting traits such as plant compactness, improving the accuracy of the adaptation index by more than 20%, ensuring a high degree of adaptation between the selected varieties and the regional environment and harvesting process in southern Xinjiang, and significantly improving the accuracy of variety selection. It achieves efficient utilization of multi-source data, integrating image, environmental, agricultural machinery, and quality data, and mining potential correlations between data to provide comprehensive data support for variety optimization, while reducing invalid data redundancy, improving data utilization, and lowering data collection and processing costs. It dynamically optimizes planting management strategies, adjusting screening thresholds in real time through the Q-learning algorithm to generate personalized planting control instructions, improving harvesting efficiency, increasing fiber quality compliance rate, and increasing yield per plant by more than 10%, effectively solving the planting adaptation problems in the salinized, temperature and humidity fluctuating environments of cotton fields in southern Xinjiang.

[0041] Please see Figure 2 In the deep learning-based method for selecting efficient cotton varieties for machine harvesting in southern Xinjiang, image data, environmental data, mechanical status data, and product quality data during the operation of the cotton harvester are acquired through a sensor network. The collected data is preprocessed to obtain a multi-source dataset, including the following steps: Step 201: Acquire image data, environmental data, mechanical status data, and product quality data during the operation of the machine in southern Xinjiang through a sensor network; Step 202: Use the 3σ criterion to filter out abrupt values ​​in the environmental data, calculate the variance using the gradient operator, and determine that images with a variance <100 are blurred and discard them, thus filtering out invalid visual data; Step 203: Use linear interpolation to fill in the short-term missing data of environmental data, and use adjacent time period data of the same plot and the same growth period to fill in the missing data of image data.

[0042] The above describes an embodiment of the deep learning-based method for screening efficient cotton varieties for machine harvesting in southern Xinjiang. Please refer to [link / reference]. Figure 3 The deep learning-based system for selecting high-efficiency varieties of machine-harvested cotton in southern Xinjiang includes the following modules: The multi-source data acquisition module is used to acquire image data, environmental data, mechanical status data and product quality data during the operation of the machine in southern Xinjiang through a sensor network, and to preprocess the acquired data to obtain a multi-source dataset. The associated feature generation module is used to perform semantic segmentation of image data in multi-source datasets by using an improved ResNet-50 model to obtain image features such as plant compactness and leaf occlusion rate. The module then uses a Transformer model to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain associated features. The adaptation index calculation module is used to fuse the image features and associated features through a cross-modal attention mechanism, input them into the XGBoost stochastic gradient descent model, and output the adaptation index of the Southern Xinjiang machine harvesting. The real-time control module for harvesting cotton in southern Xinjiang is used to adjust the variety screening threshold based on the Q-learning algorithm and the southern Xinjiang machine-harvesting adaptation index, and to generate control commands for harvesting cotton in southern Xinjiang in real time based on the screening threshold.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and their equivalents.

Claims

1. A method for selecting high-efficiency cotton varieties for machine harvesting in southern Xinjiang based on deep learning, characterized in that... The method for selecting high-efficiency varieties of machine-harvested cotton in southern Xinjiang includes the following steps: Image data, environmental data, mechanical status data, and product quality data during the operation of the machine in southern Xinjiang are acquired through sensor networks. The collected data are preprocessed to obtain a multi-source dataset. An improved ResNet-50 model was used to perform semantic segmentation of plant compactness and leaf occlusion rate on image data from multi-source datasets to obtain image features. The Transformer model was then used to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain correlation features. After fusing the image features and associated features through a cross-modal attention mechanism, the data is input into the XGBoost stochastic gradient descent model, and the output is the Southern Xinjiang machine harvesting adaptation index. Based on the Q-learning algorithm, the variety screening threshold is adjusted according to the Southern Xinjiang machine-harvesting adaptability index, and control instructions for Southern Xinjiang machine-harvested cotton are generated according to the screening threshold for real-time control.

2. The method for screening high-efficiency cotton varieties for machine harvesting in southern Xinjiang based on deep learning as described in claim 1, characterized in that, The process involves acquiring image data, environmental data, mechanical status data, and product quality data during the operation of the machine in southern Xinjiang through a sensor network. The collected data is then preprocessed to obtain a multi-source dataset, including: Image data, environmental data, mechanical status data, and product quality data during the operation of the machine in southern Xinjiang are acquired through sensor networks. The 3σ criterion is used to filter out abrupt values ​​in environmental data. The variance is calculated using the gradient operator. Images with a variance <100 are judged as blurry and discarded, thus filtering out invalid visual data. Linear interpolation was used to fill in short-term missing data in environmental data, and missing data in image data were filled in using adjacent time periods of the same plot and growth period.

3. The method for screening high-efficiency cotton varieties for machine harvesting in southern Xinjiang based on deep learning as described in claim 1, characterized in that, The process involves acquiring image data, environmental data, mechanical status data, and product quality data during the operation of the machine in southern Xinjiang through a sensor network. The collected data is then preprocessed to obtain a multi-source dataset, including: The LabelMe annotation tool was used to perform semantic segmentation and annotation on cotton plant images. Class labels were defined for compact plant areas and leaf occlusion areas. The annotated images were then bound to the variety, growth process, and environmental data of the corresponding plots to obtain preprocessed data. The environmental data, mechanical condition data, and product quality data in the preprocessed data are mapped to the [0,1] interval through Min-Max normalization. After converting the multispectral image into an RGB pseudo-color image, mean and variance are standardized to obtain a multi-source dataset.

4. The method for screening high-efficiency cotton varieties for machine harvesting in southern Xinjiang based on deep learning as described in claim 1, characterized in that, The improved ResNet-50 model is used to perform semantic segmentation of image data from multi-source datasets, considering plant compactness and leaf occlusion rate, to obtain image features. A Transformer model is then used to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain correlation features, including: After the residual blocks in stages 2, 3, and 4 of ResNet-50, a CBAM channel attention module is embedded to obtain an improved ResNet-50 model. The preprocessed cotton plant image is input into the improved ResNet-50 model, and the semantic segmentation result is output through the segmentation head of U-Net+ residual block to obtain image features.

5. The method for screening high-efficiency cotton varieties for machine harvesting in southern Xinjiang based on deep learning as described in claim 1, characterized in that, The improved ResNet-50 model is used to perform semantic segmentation of image data from multi-source datasets, considering plant compactness and leaf occlusion rate, to obtain image features. A Transformer model is then used to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain correlation features, including: Environmental data, mechanical condition data, and product quality data are mapped into numerical feature vectors through a fully connected layer; The Transformer model employs a 6-layer encoder-decoder architecture, with 8 multi-head attention heads, 256 hidden layer dimensions, and a dropout rate of 0.

2. Using machine-collected morphological image features as query vectors and numerical feature vectors as key-value pairs, we learn multi-dimensional relationships and output associated features.

6. The method for screening high-efficiency cotton varieties for machine harvesting in southern Xinjiang based on deep learning as described in claim 1, characterized in that, After fusing the image features and associated features through a cross-modal attention mechanism, the data is input into the XGBoost stochastic gradient descent model, and the output is the Southern Xinjiang machine harvesting adaptation index, which includes: A cross-modal attention layer is constructed to calculate the mutual information between image features and associated features. Dynamic weights are assigned to features with high mutual information. Image features and associated features are then concatenated according to the dynamic weights to obtain a comprehensive feature vector. Based on historical data of machine-harvested varieties, varieties are divided into high-fitness, medium-fitness, and low-fitness categories, which serve as training labels for the XGBoost stochastic gradient descent model. The model outputs scores for five sub-indicators, including scores for growth process suitability, stress resistance suitability, machine harvesting trait suitability, yield suitability, and quality suitability. These scores are weighted according to their respective weights to obtain the suitability index for machine-harvested cotton varieties in southern Xinjiang.

7. The method for screening high-efficiency planting varieties of machine-harvested cotton in southern Xinjiang based on deep learning as described in claim 1, characterized in that, The method based on the Q-learning algorithm adjusts the variety screening threshold according to the Southern Xinjiang machine-harvesting adaptability index, and generates Southern Xinjiang machine-harvested cotton control instructions for real-time control based on the screening threshold, including: Using the current batch of varieties' compatibility index distribution and demonstration area environmental data as the state space, the range of adjustment of the variety selection threshold as the action space, and the actual planting effect in the demonstration area as the reward value, the variety selection threshold is adjusted according to the aforementioned Southern Xinjiang machine harvesting compatibility index. Based on the screening threshold, control instructions for machine-harvested cotton in southern Xinjiang are generated, including at least plant and row spacing configuration, water and fertilizer management, and machine harvesting parameters. The control commands for machine-harvested cotton in southern Xinjiang are synchronized to the machine operation system in southern Xinjiang via the LoRa wireless communication protocol.

8. A deep learning-based system for selecting efficient cotton varieties for machine harvesting in southern Xinjiang, characterized in that: The high-efficiency cotton variety selection system for machine harvesting in southern Xinjiang includes the following modules: The multi-source data acquisition module is used to acquire image data, environmental data, mechanical status data and product quality data during the operation of the machine in southern Xinjiang through a sensor network, and to preprocess the acquired data to obtain a multi-source dataset. The associated feature generation module is used to perform semantic segmentation of image data in multi-source datasets by using an improved ResNet-50 model to obtain image features such as plant compactness and leaf occlusion rate. The module then uses a Transformer model to perform cross-modal correlation analysis on environmental data, mechanical condition data, and product quality data to obtain associated features. The adaptation index calculation module is used to fuse the image features and associated features through a cross-modal attention mechanism, input them into the XGBoost stochastic gradient descent model, and output the adaptation index of the Southern Xinjiang machine harvesting. The real-time control module for harvesting cotton in southern Xinjiang is used to adjust the variety screening threshold based on the Q-learning algorithm and the southern Xinjiang machine-harvesting adaptation index, and to generate control commands for harvesting cotton in southern Xinjiang in real time based on the screening threshold.

9. The deep learning-based high-efficiency cotton variety selection system for machine harvesting in southern Xinjiang as described in claim 8, characterized in that, The adaptation index calculation module includes the following sub-modules: The splicing submodule is used to construct a cross-modal attention layer, calculate the mutual information between image features and associated features, assign dynamic weights to features with high mutual information, and splice the image features and associated features according to the dynamic weights to obtain a comprehensive feature vector. The training submodule is used to classify varieties into high-fit, medium-fit, and low-fit based on historical data of machine-harvested varieties, and use these as training labels for the XGBoost stochastic gradient descent model. The submodule is used to output scores for five sub-indicators of the model, including the adaptation score for the growth process, the adaptation score for the stress resistance, the adaptation score for the machine harvesting trait, the adaptation score for the yield, and the adaptation score for the quality. The adaptation index of machine harvested cotton varieties in southern Xinjiang is obtained by weighting the indicators according to their weights.

10. The deep learning-based high-efficiency cotton variety selection system for machine harvesting in southern Xinjiang as described in claim 8, characterized in that, The real-time control module for the Southern Xinjiang aircraft includes the following sub-modules: Define a submodule, which uses the current batch of varieties' adaptation index distribution and demonstration area environmental data as the state space, adjusts the variety screening threshold by a certain range as the action space, and uses the actual planting effect in the demonstration area as the reward value. Adjust the variety screening threshold according to the Southern Xinjiang machine harvesting adaptation index. The instruction submodule is used to generate control instructions for machine-harvested cotton in southern Xinjiang based on the screening threshold, including at least plant spacing configuration, water and fertilizer management, and machine harvesting parameters; The control submodule is used to synchronize the control commands for machine-harvested cotton in southern Xinjiang to the machine operation system in southern Xinjiang via the LoRa wireless communication protocol.

Images