Machine learning-based method, system and computer-readable storage medium for predicting potential yield of non-food crops on marginal land

Abstract

The present application belongs to the field of agricultural engineering and computer technology, and specifically relates to a potential yield prediction method, system and computer readable storage medium for non-food crops on marginal land based on machine learning. The method comprises the following steps: step 1: obtaining measured planting data of non-food crops and geographic coordinate information of planting locations; step 2: constructing a feature data set; step 3: constructing multiple machine learning prediction models, and training and performance evaluating; step 4: optimizing the prediction accuracy of the multiple machine learning prediction models; step 5: screening to obtain an optimal machine learning prediction model; and step 6: predicting the yield of non-food crops in the marginal land range and generating a spatial distribution result of the yield of non-food crops. The method of the present application can realize high-precision prediction of the yield of non-food crops under different climate, soil and management conditions.

Description

Technical Field

[0001] This invention belongs to the fields of agricultural engineering and computer technology, and specifically relates to a method, system, and computer-readable storage medium for predicting the potential yield of non-grain crops on marginal land based on machine learning. Background Technology

[0002] Non-grain crops, as renewable biomass resources that do not compete with grain crops for land, have significant application potential in areas such as replacing fossil fuels and producing bio-based materials. Sweet sorghum can be converted into fuel ethanol, Jerusalem artichokes can be used to extract fructooligosaccharides, and hemp can produce green fibers, providing new pathways for optimizing agricultural structure and alleviating energy pressure. However, large-scale cultivation of these crops has yet to be achieved, with the core bottleneck being the lack of reliable methods for predicting planting feasibility and yield levels under different regions and land conditions. my country spans a vast territory with diverse land conditions, ranging from the black soil regions of Northeast China to the saline-alkali lands of Northwest China and the red soil hills of South China. The growth of non-grain crops is influenced by multiple factors, including climate, soil, and cultivation management.

[0003] Existing methods for assessing non-grain crop yields primarily rely on crop growth process models or extensions of traditional crop models. Their limitations have become a core obstacle to improving accuracy. First, they are highly parameter-dependent: model operation requires precise physiological parameters such as photosynthetic rate and transpiration coefficient. However, basic research on most non-grain crops is weak, and observational data for these parameters is scarce. Some parameters require high-precision laboratory measurements, which are costly and time-consuming. Furthermore, parameters vary significantly among different varieties, resulting in poor adaptability. Second, they have limited regional adaptability: traditional models are mostly built based on experimental data from specific regions and cannot adapt to cross-regional climate and soil differences. For example, when a sweet sorghum model from Inner Mongolia's saline-alkali land is used in coastal saline-alkali land in Shandong, the prediction error is unacceptably high due to differences in precipitation rhythms and salinity composition. Finally, they struggle to characterize complex nonlinear relationships: the interaction between extreme high temperatures, precipitation fluctuations, soil fertility, and fertilization timing is highly nonlinear. Traditional models often rely on linear assumptions and cannot simulate this multi-factor coupling process. This leads to insufficient accuracy and stability in large-scale spatial predictions and cross-regional applications, resulting in inadequate prediction accuracy at the provincial and national levels, and failing to support cross-regional industrial layout decisions.

[0004] In contrast, machine learning has demonstrated high accuracy and generalization capabilities in yield prediction for grain crops such as wheat and rice, effectively addressing multi-factor interactions by uncovering hidden patterns in the data. However, its systematic application in non-grain crops remains a blank. On one hand, non-grain crop data is scarce: unlike the decades of continuous observation data for grain crops, non-grain crops are planted in a dispersed manner with short experimental cycles, lacking standardized data covering multiple regions and growth stages, making it difficult to support model training and validation. On the other hand, the diversity of crop types makes model adaptation difficult: from energy-producing Jatropha curcas to forage crops like sugar beets, the growth characteristics of various crops differ greatly, making it impossible for general models to adapt to all categories. Furthermore, the "black box" nature of machine learning reduces trust: agricultural technicians and farmers are more concerned with yield-influencing factors and control measures, but most machine learning models cannot intuitively explain the causes of predictions, hindering their implementation and promotion in non-grain crop fields. Summary of the Invention

[0005] Based on the above situation, it is necessary to develop a machine learning-based method for predicting non-grain crop yields. By integrating validated global non-grain crop measured data, multiple prediction models can be constructed, the impact mechanism of environmental and management factors on yields can be systematically analyzed, and the optimal model can be selected for predicting non-grain yields on marginal land. This will provide reliable technical means and data support for assessing non-grain crop resource quantity and optimizing the spatial layout of raw material supply.

[0006] Therefore, this invention proposes a machine learning-based method for predicting the potential yield of non-grain crops on marginal land, comprising the following steps: Step 1: Obtaining management parameters and geographic coordinates of the planting location of non-grain crops, wherein the management parameters are measured planting data; Step 2: Based on the geographic coordinates, extracting corresponding climate parameters, soil parameters, and topographic parameters from climate parameter sources, soil parameter sources, and topographic data sources respectively, and constructing a feature dataset for non-grain crop yield prediction based on the management parameters, climate parameters, soil parameters, and topographic parameters; Step 3: Constructing multiple machine learning prediction models based on the feature dataset for non-grain crop yield prediction, and training and evaluating the performance of the machine learning prediction models; Step 4: Adjusting parameters through machine learning to optimize the prediction accuracy of the multiple machine learning prediction models; Step 5: Selecting the optimal machine learning prediction model based on the evaluation results of the multiple machine learning prediction models; Step 6: Predicting the yield of non-grain crops within the marginal land area based on the optimal machine learning prediction model, generating spatial distribution results of non-grain crop yield.

[0007] Specifically, in step 1, the geographic coordinate information can be latitude and longitude information.

[0008] Specifically, in step 2, the management parameters, climate parameters, soil parameters, and terrain parameters are integrated and standardized before being used as the input feature set for the machine learning prediction model.

[0009] Specifically, in step 3, the model can be trained using leave-one-out cross-validation and hierarchical cross-validation, and the model performance can be evaluated based on the coefficient of determination, root mean square error, and mean absolute error.

[0010] Specifically, in step 5, the optimal configuration of the model is selected based on comprehensive performance indicators as a tool for predicting non-grain crop yields, which can ensure high accuracy, stability and adaptability to different regions and environmental conditions.

[0011] Specifically, in step 6, within the defined marginal land area, at least one of the following parameters closely related to the growth of non-grain crops is extracted: management parameters, climate parameters, and soil parameters. These include, but are not limited to: climate parameters such as temperature, precipitation, solar radiation, growth period, wind speed, humidity, and evaporation; soil indicators such as organic carbon, nitrogen, phosphorus, and potassium content, density, porosity, texture, water content, and bulk density; and multi-dimensional gradient management parameters, including fertilizer type, single fertilizer application rate, planting years, pesticide application rate, irrigation amount, planting density, and crop rotation / continuous cropping years. By constructing multiple management scenarios, flexible simulation of non-grain crop yield prediction can be achieved.

[0012] Ultimately, a complete feature dataset for marginal land was constructed, which integrates crop characteristics, climate, soil and management information. It has high resolution and is used for non-grain crop yield simulation under multiple crops, multiple management schemes and multiple climate scenarios, providing refined and quantifiable data support for non-grain crop development, raw material supply layout and resource management.

[0013] Using a constructed feature dataset for marginal land as input, and employing an optimized and selected optimal machine learning prediction model, yield prediction can be performed, generating a high-resolution spatial distribution map of non-grain crop yields. Analysis of the prediction results, including crop responses to management parameters, selection of suitable crop varieties, spatial distribution characteristics, and potential yield assessment, provides data support for non-grain crop raw material supply planning and spatial layout optimization.

[0014] In summary, the aforementioned technical solutions aim to achieve high-precision prediction of the potential non-grain crop yields on a large amount of marginal land in my country that is difficult to cultivate traditional food crops. This will provide reliable technical support for the scientific planning and optimized spatial layout of non-grain crop resources. The method systematically integrates global measured data on non-grain crops, multi-level climate parameters, refined soil parameters, and agricultural management practice data to construct and continuously optimize a machine learning-based prediction model. Ultimately, this achieves accurate and efficient prediction of the yield per unit area of ​​various non-grain crops on marginal land, thus assisting agricultural decision-making and regional sustainable development.

[0015] In one implementation, in step 1, the management parameters include at least one of the following: crop yield, planting years, planting density, crop growth cycle, crop rotation / continuous cropping years, fertilizer type, single fertilizer application rate, pesticide application rate, growing season, and irrigation amount. These data can be obtained through farmland sensors, agricultural records, or farmer surveys. In this way, the method of the present invention sets multidimensional gradient parameters to address the impact of management measures, enabling flexible simulation of non-grain crop yield prediction.

[0016] In one implementation, in step 2, the climate parameters include at least two of the following: temperature, precipitation, solar radiation intensity, growing season, wind speed, humidity, and evaporation rate. These data can be obtained, for example, by extracting key meteorological parameters from climate parameter databases such as ERA5, ESMAP, and EPA5.

[0017] In one implementation, in step 2, the soil parameters include at least two of the following: pH value, organic carbon content, nitrogen, phosphorus, and potassium content, density, porosity, texture, water content, and bulk density. These data can be obtained, for example, by extracting soil characteristic indicators from a soil parameter library (such as HWSD2).

[0018] In one implementation, in step 3, the multiple machine learning prediction models include at least two of the following: Random Forest, Gradient Boosting Tree, Extreme Gradient Boosting, Support Vector Machine, K-Nearest Neighbors, Artificial Neural Network, Convolutional Neural Network, Elastic Network, Lasso Regression, Ridge Regression, CatBoost, and Gaussian Process Regression. Random Forest, based on the ensemble of multiple decision trees, effectively avoids overfitting risks, is robust to outliers and noise, and can output feature importance to aid feature selection. Gradient Boosting Tree series (including XGBoost and CatBoost) iteratively optimizes residuals, exhibiting outstanding fitting capabilities; CatBoost can automatically process categorical features, significantly reducing manual preprocessing costs. Support Vector Machine performs well in high-dimensional data scenarios, especially suitable for tasks with small sample sizes. Artificial Neural Networks and Convolutional Neural Networks excel at uncovering deep nonlinear relationships in data and are well-suited for complex structured data such as images and time series. Elastic Network, Lasso Regression, and Ridge Regression can achieve feature selection and regularization in regression tasks, precisely controlling model complexity. Gaussian Process Regression has probabilistic output capabilities, quantifying prediction uncertainty.

[0019] In one implementation, before screening the machine learning prediction model, step 4 involves outlier removal or feature filtering on the feature dataset related to non-grain crop yield prediction. Specifically, this can be achieved by using variance filtering, outlier removal, and feature importance analysis to optimize input variables, eliminate extreme outliers, balance the contributions of climate, soil, and management factors, reduce the risk of overfitting, and improve the model's generalization ability.

[0020] In one implementation scheme, the non-grain crops include at least one of the following: Miscanthus sinensis, switchgrass, Napier grass, reeds, eucalyptus, poplar, willow, sweet sorghum, and sugarcane. If necessary, the non-grain crops may also include other non-grain crops that can be grown on marginal land. These crops play multiple core roles in ecological protection, resource utilization, and economic development: Among herbaceous non-grain crops, Miscanthus sinensis, switchgrass, and Napier grass have well-developed root systems and are tolerant of poor soil and saline-alkali conditions, enabling them to quickly cover exposed surfaces on marginal land, effectively conserving soil and water, improving soil structure, and curbing soil erosion and desertification; simultaneously, they have high biomass yields and stable calorific value, making them high-quality raw materials for biomass power generation and biofuel refining, which can replace fossil fuels and reduce carbon emissions. Reeds combine ecological restoration and economic value; in addition to their soil and water conservation function, their stems can be used for papermaking and weaving, broadening income channels for growers. Woody trees such as eucalyptus, poplar, and willow have short growth cycles, are fast-growing and high-yielding, and have outstanding carbon sequestration capabilities, significantly increasing the carbon storage of marginal land and contributing to the achievement of "dual carbon" goals. Their timber can be used in the paper industry and prefabricated building materials production, replacing natural forest resources, alleviating the supply and demand imbalance of forestry resources, and reducing damage to native ecosystems. Sweet sorghum and sugarcane are high-value-added non-grain crops. They can not only be planted on marginal lands such as saline-alkali land and tidal flats, but can also be processed into sugar and fuel ethanol. Byproducts such as bagasse and sorghum straw can be used for biomass power generation or feed processing, achieving multi-level utilization of resources, bringing stable income to growers, and at the same time avoiding competition with grain crops for land from the root, thus strengthening the bottom line of food security. Including other suitable non-grain crops when necessary can further enrich the planting structure and meet the personalized needs of ecological restoration and industrial development in different regions.

[0021] Based on land productivity, ecological protection requirements, and suitability standards for large-scale non-grain crop cultivation, potential marginal land within my country can be spatially identified and classified. In one implementation plan, marginal land includes at least one of the following: low-productivity farmland, saline-alkali land, dry land, degraded land, wasteland, sandy land, low-coverage grassland, shrubland, sparse woodland, and tidal flats. These marginal lands are areas unsuitable for large-scale grain crop cultivation, while excluding restricted areas such as ecologically sensitive areas, water source protection areas, and forest protection areas. Regarding low-productivity farmland, for example, farmland with a yield per unit area lower than the local average yield of similar arable land by 20% or more can be classified as low-productivity farmland. Cultivating non-grain crops on such marginal land has multiple significant values. First, it can efficiently revitalize idle and inefficient land, avoiding competition with grain crops for land: converting low-yield farmland to high-quality forage and silage corn, and promoting salt-tolerant wolfberry and salt-tolerant Suaeda salsa in saline-alkali land can transform land with meager yields into effective production resources, significantly improving land use efficiency without encroaching on grain cultivation space, and indirectly ensuring food security. Secondly, it also has ecological restoration benefits: planting sea buckthorn, caragana, and other sand-fixing plants in sandy areas, and planting perennial herbs and shrubs on degraded land, can effectively fix carbon and retain water, prevent wind erosion and sand, curb desertification and soil erosion, and gradually restore the regional ecological environment, achieving a win-win situation for both ecological and economic benefits. Thirdly, it can drive the development of characteristic industries and contribute to rural revitalization: planting salt-tolerant algae and medicinal herbs on tidal flats, and cultivating energy plants on wastelands, can extend the entire industrial chain of planting, processing, and sales, creating numerous jobs and broadening farmers' income channels. Furthermore, this type of planting can precisely fill market gaps. For example, forage planting can alleviate the pressure on feed supply in the livestock industry, and energy plants can provide raw material support for the development of biomass energy, enriching my country's agricultural product supply structure and promoting sustainable agricultural development.

[0022] This invention also proposes a non-grain crop yield prediction system, comprising: a feature dataset construction module for acquiring the management parameters, climate parameters, soil parameters, and topographic parameters of the non-grain crop to construct the feature dataset for non-grain crop yield prediction; a model construction and optimization module for constructing and optimizing the machine learning prediction model; a marginal land feature data construction module for constructing a feature dataset for marginal land as input for non-grain crop yield prediction, wherein the feature dataset for marginal land includes at least one of the management parameters, climate parameters, and soil parameters of the marginal land; and a yield prediction module for inputting the feature dataset for marginal land into the optimal machine learning prediction model to generate a non-grain crop yield prediction result.

[0023] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to implement the method proposed in the present invention.

[0024] The advantages of this invention are:

[0025] This invention presents a groundbreaking method, system, and computer-readable storage medium for predicting non-grain crop yields, completely departing from traditional technological approaches. It enables high-precision, universally robust spatial prediction of non-grain crop yields under arbitrary climatic conditions, soil types, and management models. This invention overcomes three major technical challenges that traditional process models have long struggled to address: excessive parameter dependence, extremely poor regional adaptability, and virtually zero ability to characterize nonlinear relationships, filling a long-standing technological gap in this field. The invention has a wide range of applications, applicable to various non-grain crop types and cross-regional, cross-scale land resources, providing optimal yield prediction solutions and serving as a reference for crop and region selection. It not only provides unprecedented scientific decision-making support for large-scale non-grain crop development but also achieves highly optimized resource spatial allocation, significantly improving marginal land development efficiency and system sustainability, while maximizing ecological and economic benefits. Particularly noteworthy is that this invention integrates multiple cutting-edge machine learning models and dynamic optimization strategies, endowing the technical solution with superior adaptability and scalability, completely eliminating the fatal flaw of traditional methods that require complex system architecture processes for specific regions. This invention establishes a highly robust and universally applicable technical system for the whole-space prediction and strategic planning of non-grain crops, representing a significant and groundbreaking advancement in this field.

[0026] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0027] The above and other objects, features, and advantages of the present invention will become more apparent from the more detailed description of the embodiments of the invention in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0028] Figure 1 A flowchart of the method according to the present invention is shown.

[0029] Figure 2 The figure shows the prediction effect of random forest on Miscanthus yield before and after optimization according to Embodiment 1 of the method according to the present invention, wherein the left figure shows the prediction effect before optimization and the right figure shows the prediction effect after optimization.

[0030] Figure 3The yield distribution of Miscanthus sinensis in marginal land under no-fertilization conditions according to Embodiment 1 of the method according to the present invention is shown.

[0031] Figure 4 The maximum yield distribution of Miscanthus sinensis on marginal land is shown in Embodiment 1 of the method according to the present invention.

[0032] Figure 5 The figure shows the prediction effect of random forest on poplar yield before and after optimization according to Embodiment 2 of the method according to the present invention, wherein the left figure shows the prediction effect before optimization and the right figure shows the prediction effect after optimization.

[0033] Figure 6 The maximum yield distribution of poplar trees on marginal land is shown in Embodiment 2 of the method according to the present invention.

[0034] Figure 7 The figure shows the prediction effect of XGBoost on Miscanthus yield before and after optimization in Embodiment 3 of the method according to the present invention, wherein the left figure shows the prediction effect before optimization and the right figure shows the prediction effect after optimization.

[0035] Figure 8 The maximum yield distribution of Miscanthus sinensis on marginal land is shown in Embodiment 3 of the method according to the present invention. Detailed Implementation

[0036] Figure 1 A flowchart of the method according to the present invention is shown.

[0037] This invention proposes a machine learning-based method for predicting the potential yield of non-grain crops on marginal land, comprising the following steps: Step 1: Obtaining management parameters and geographic coordinates of the planting location of the non-grain crops, wherein the management parameters are measured planting data; Step 2: Based on the geographic coordinates, extracting corresponding climate parameters, soil parameters, and topographic parameters from climate parameter sources, soil parameter sources, and topographic data sources respectively, and constructing a feature dataset for non-grain crop yield prediction based on the management parameters, climate parameters, soil parameters, and topographic parameters; Step 3: Constructing multiple machine learning prediction models based on the feature dataset for non-grain crop yield prediction, and training and evaluating the performance of the machine learning prediction models; Step 4: Adjusting parameters through machine learning to optimize the prediction accuracy of the multiple machine learning prediction models; Step 5: Selecting the optimal machine learning prediction model based on the evaluation results of the multiple machine learning prediction models; Step 6: Predicting the yield of non-grain crops within the marginal land area based on the optimal machine learning prediction model, generating spatial distribution results of non-grain crop yield.

[0038] Specifically, in step 1, the geographic coordinate information can be latitude and longitude information.

[0039] Specifically, in step 2, the management parameters, climate parameters, soil parameters, and terrain parameters are integrated and standardized before being used as the input feature set for the machine learning prediction model.

[0040] Specifically, in step 3, the model can be trained using leave-one-out cross-validation and hierarchical cross-validation, and the model performance can be evaluated based on the coefficient of determination, root mean square error, and mean absolute error.

[0041] Specifically, in step 5, the optimal configuration of the model is selected based on comprehensive performance indicators as a tool for predicting non-grain crop yields, ensuring high accuracy, stability, and adaptability to different regions and environmental conditions. In an alternative solution, the present invention integrates a model selection and evaluation module. Through cross-validation and multi-indicator comprehensive evaluation, the optimal machine learning prediction model is selected, and outlier removal and feature selection are performed on the feature dataset, achieving the expected yield prediction accuracy.

[0042] Specifically, in step 6, within the defined marginal land area, management parameters, climate parameters, and soil parameters closely related to the growth of non-grain crops are extracted. These include, but are not limited to: climate parameters such as temperature, precipitation, solar radiation, growth period, wind speed, humidity, and evaporation; soil indicators such as organic carbon, nitrogen, phosphorus, and potassium content, density, porosity, texture, water content, and bulk density; and multi-dimensional gradient management parameters, including fertilizer type, single fertilizer application rate, planting years, pesticide application rate, irrigation amount, planting density, and crop rotation / continuous cropping years. By constructing multiple management scenarios, flexible simulation of non-grain crop yield prediction can be achieved.

[0043] Ultimately, a complete feature dataset for marginal land was constructed, which integrates crop characteristics, climate, soil and management information. It has high resolution and is used for non-grain crop yield simulation under multiple crops, multiple management schemes and multiple climate scenarios, providing refined and quantifiable data support for non-grain crop development, raw material supply layout and resource management.

[0044] Using a constructed feature dataset for marginal land as input, and employing an optimized and selected optimal machine learning prediction model, yield prediction can be performed, generating a high-resolution spatial distribution map of non-grain crop yields. Analysis of the prediction results, including crop responses to management parameters, selection of suitable crop varieties, spatial distribution characteristics, and potential yield assessment, provides data support for non-grain crop raw material supply planning and spatial layout optimization.

[0045] In one implementation, in step 1, the management parameters include at least one of the following: crop yield, planting years, planting density, crop growth cycle, crop rotation / continuous cropping years, fertilizer type, single fertilizer application rate, pesticide application rate, growing season, and irrigation amount. These data can be obtained through farmland sensors, agricultural records, or farmer surveys. In this way, the method of the present invention sets multidimensional gradient parameters to address the impact of management measures, enabling flexible simulation of non-grain crop yield prediction.

[0046] In one implementation, in step 2, the climate parameters include at least two of the following: temperature, precipitation, solar radiation intensity, growing season, wind speed, humidity, and evaporation rate. These data can be obtained, for example, by extracting key meteorological parameters from climate parameter databases such as ERA5, ESMAP, and EPA5.

[0047] In one implementation, in step 2, the soil parameters include at least two of the following: pH value, organic carbon content, nitrogen, phosphorus, and potassium content, density, porosity, texture, water content, and bulk density. These data can be obtained, for example, by extracting soil characteristic indicators from a soil parameter library (such as HWSD2).

[0048] In one implementation, in step 3, the multiple machine learning prediction models include at least two of random forest, gradient boosting tree, extreme gradient boosting, support vector machine, K-nearest neighbors, artificial neural network, convolutional neural network, elastic network, Lasso regression, Ridge regression, CatBoost, and Gaussian process regression.

[0049] In one implementation, before screening the machine learning prediction model, in step 4, outlier removal or feature screening is performed on the feature dataset regarding non-grain crop yield prediction. Specifically, input variables can be optimized by using variance screening, outlier removal, and feature importance analysis to remove extreme outliers, balance the contributions of climate, soil, and management factors, reduce the risk of overfitting, and improve the model's generalization ability. In an alternative, the present invention develops a feature optimization algorithm based on variance screening, which uses variance threshold analysis, feature variance ranking, and other methods to screen high-contribution features and remove low-variance redundant features, thereby achieving automated simplification and accuracy tuning of the input features for each prediction model.

[0050] In one implementation, the non-grain crops include at least one of miscanthus, switchgrass, foxtail grass, reeds, eucalyptus, poplar, willow, sweet sorghum, and sugarcane. If necessary, the non-grain crops may also include other non-grain crops that can be grown on marginal land.

[0051] Based on land productivity, ecological protection requirements, and suitability standards for large-scale non-grain crop cultivation, potential marginal land within my country can be spatially identified and classified. In one implementation plan, the marginal land includes at least one of the following: low-productivity farmland, saline-alkali land, dry land, degraded land, wasteland, sandy land, low-coverage grassland, shrubland, sparse woodland, and tidal flats. These marginal lands are areas unsuitable for large-scale grain crop cultivation, while excluding restricted areas such as ecologically sensitive areas, water source protection areas, and forest protection areas. For example, low-productivity farmland can be classified as farmland whose yield per unit area is 20% or more lower than the average yield of similar arable land in the local area. In one alternative approach, to address the issues of high noise and numerous outliers in marginal land data, an adaptive outlier detection and removal algorithm is developed for different types of marginal land, such as saline-alkali land, dry land, and degraded land. Combined with feature variance screening, a cleaning process suitable for low-quality agricultural data is constructed. The yield response characteristics of non-grain crops (Miscanthus sinensis, switchgrass, poplar, etc.) under different marginal land types are studied, and differentiated feature screening rules based on land type are established to remove redundant features unrelated to the yield of specific marginal land. A dedicated model calling interface for marginal land yield prediction is developed, which automatically identifies marginal land parameters based on input geographic coordinates, dynamically matches the optimal feature subset and pre-trained model, and achieves accurate prediction. Prediction is conducted on typical marginal lands across the country (coastal saline-alkali land, northwest dry land, and southern wasteland). The system's yield predictions for different land types show minimal discrepancies between the predicted results and actual planting.

[0052] This invention also proposes a non-grain crop yield prediction system, comprising: a feature dataset construction module for acquiring the management parameters, climate parameters, soil parameters, and topographic parameters of the non-grain crop to construct the feature dataset for non-grain crop yield prediction; a model construction and optimization module for constructing and optimizing the machine learning prediction model; a marginal land feature data construction module for constructing a feature dataset for marginal land as input for non-grain crop yield prediction, wherein the feature dataset for marginal land includes at least one of the management parameters, climate parameters, and soil parameters of the marginal land; and a yield prediction module for inputting the feature dataset for marginal land into the optimal machine learning prediction model to generate a non-grain crop yield prediction result.

[0053] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to implement the method proposed in the present invention.

[0054] In summary, on the one hand, the method of this invention pioneers a mechanism of "spatial alignment and standardized fusion of multi-source heterogeneous feature data" to achieve seamless integration and scale matching of management parameters, climate parameters and soil parameters; on the other hand, it realizes a prediction strategy of "variance-based screening and parallel selection of multiple models" to break through the bottleneck of the generalization ability of a single model among different marginal land types (saline-alkali land, dry land, degraded land, wasteland, etc.).

[0055] On the other hand, the method of this invention proposes for the first time a "marginal land type-driven hierarchical feature screening" technique, which establishes a feature importance ranking for saline-alkali land, dry land, degraded land, etc., eliminates invalid features, and significantly improves prediction stability. At the same time, it realizes the system-level capability of "batch prediction and dynamic adaptation of marginal land": the system can import the geographical coordinates and management parameters of hundreds or thousands of marginal land plots at one time, automatically identify the land type of each plot, automatically match features and models, and automatically output the spatial distribution of output, solving the inefficiency problem of manual adaptation of each plot in traditional methods.

[0056] Through internal cross-validation, the average prediction accuracy of this invention for non-grain crop yields under different marginal land types reaches 84%, which is better than traditional crop models and can provide reliable decision support for the large-scale planting layout of non-grain crops.

[0057] In addition, actual yield data of non-grain crops on marginal land recorded in published academic papers were selected as a benchmark, and the prediction results of the model of this invention were compared with the above-mentioned measured data. The results show that the error between the predicted value of the model of this invention and the actual yield data is small, verifying the effectiveness and reliability of the model in real-world scenarios.

[0058] Example

[0059] To make the present invention easier to understand, specific embodiments are provided below to illustrate the practical effects of the method of the present invention. These embodiments are for illustrative purposes only and are not limited to the application scope of the present invention.

[0060] The following examples use barren land as a model for prediction, with extremely barren land in China as the verification objects: such as the Loess Plateau with severe soil erosion, ecologically fragile low-coverage grasslands, saline-alkali lands in Hunan and Hubei with extremely low crop yields, and severely acidified and compacted red soils in Guangdong and Guangxi. Under the harsh site conditions where any of the above traditional models are almost ineffective and the prediction error is unacceptably high, the method, system, and computer-readable storage medium proposed in this invention can still achieve ultra-high accuracy and stable cross-regional prediction of non-grain crop yields.

[0061] In the following examples, the fertilizers are urea, superphosphate, and potassium chloride, with a marginal land ratio of 1:1:1.

[0062] Example 1: A method for predicting Miscanthus yield based on random forest

[0063] 1) Sample data acquisition and processing

[0064] We obtained global measured data on Miscanthus sinensis (including yield, fertilizer application, planting years, and latitude and longitude), and standardized the fertilizer application and planting years to eliminate data discrepancies.

[0065] 2) Construction of feature variables

[0066] Based on latitude and longitude, climate and topographic parameters are extracted. The climate parameters include horizontal radiation (GHI), annual mean temperature (TEM), annual precipitation (PRE), and growing season length (GSL), sourced from the European Centre for Medium-Range Weather Forecasts' Generation 5 Reanalysis (ERA5) data and the World Bank's Energy Sector Management Assistance Program. Soil parameters are extracted, including soil organic matter content (SOM), soil volumetric water content (SWC), and soil available water holding capacity (AWC), sourced from the Harmonized World Soil Parameters Database 2.0 (HWSD2) published by the Food and Agriculture Organization of the United Nations (FAO). Management parameters are extracted, including planting age (AGE) and annual fertilizer application rate (FER). These four parameters are then integrated to form the model training dataset.

[0067] 3) Model building and optimization

[0068] Construct a Radio Frequency (RF) machine learning model, train it using leave-one-out cross-validation, and implement it through R. 2 Performance was evaluated using root mean square error, and the model was optimized by screening model configurations and eliminating extreme production samples. Figure 2 The diagram illustrates the prediction effect of random forest on Miscanthus yield before and after optimization according to Embodiment 1 of the method of the present invention. The left graph shows the prediction effect before optimization, and the right graph shows the prediction effect after optimization. The horizontal axis represents the observed Miscanthus yield (t / ha), and the vertical axis represents the predicted Miscanthus yield (t / ha). R0 2 N represents the coefficient of determination. masked Indicates the total number of masked lexical units. From Figure 2 As can be seen, the R² of Miscanthus sinensis was improved to 0.82, reaching the high reliability prediction standard, and the fitting ability was excellent.

[0069] 4) Delineation of marginal land boundaries

[0070] Marginal land refers to areas that are unsuitable for large-scale cultivation of food crops due to limitations in soil, water, heat, and topography, resulting in low agricultural productivity, limited economic returns, and fragile ecosystems.

[0071] Example 1 combines the requirements of farmland protection, ecological security and large-scale planting of non-grain crops to determine the constraints for extracting marginal land space, including low-productivity farmland, saline-alkali land, dry land, degraded land, and slopes, and excluding control areas such as permanent basic farmland, ecologically sensitive areas, water source protection areas, and forest protection areas, to generate a unified resolution map of China's marginal land space distribution.

[0072] 5) Construction of Marginal Land Machine Learning Prediction Dataset

[0073] Based on the spatial distribution map of marginal land, corresponding climate and soil characteristic variables are extracted to construct a climate-soil basic dataset; for the two management parameters of fertilizer application and planting years, gradient intervals are set to form multi-scenario management variable data.

[0074] 6) Marginal land Miscanthus yield forecast

[0075] The constructed marginal land prediction dataset was input into the optimized random forest (RF) model of the example to predict the yield of Miscanthus sinensis. The resulting spatial distribution map of Miscanthus sinensis yield on marginal land in China and the yield comparison dataset are shown in Table 1. The total yield of Miscanthus sinensis on marginal land under different planting years and fertilization amounts are shown in Table 1.

[0076] Table 1. Total marginal land yield (Gt) of Miscanthus sinensis under different planting years and fertilization levels.

[0077]

[0078] on the other hand, Figure 3 The diagram shows the (aggregate) yield distribution of Miscanthus sinensis on provincial border lands under no-fertilization conditions according to Embodiment 1 of the method according to the present invention. Figure 4 The maximum yield distribution of Miscanthus sinensis in marginal lands of various provinces (200 kg / ha – 8 years) according to Example 1 of the method according to the invention is shown.

[0079] Overall, Miscanthus sinensis, as a low-input, high-biomass energy crop, is particularly suitable for the ecological and climatic conditions of China's vast marginal lands, making it a preferred species for the development of localized energy crops.

[0080] This embodiment uses Miscanthus giganteus, considered a representative crop of next-generation energy crops and also the most suitable energy crop for Chinese land, as the research object. It selects Random Forest (RF), an advanced machine learning algorithm, as the carrier to fully demonstrate the complete implementation process of the non-grain crop yield prediction method based on machine learning of this invention. This embodiment fully demonstrates that even on complex marginal land where traditional models have completely failed, this invention can still achieve extremely high prediction accuracy and spatial generalization ability.

[0081] This invention utilizes machine learning methods to accurately predict yields on barren land, fundamentally changing the traditional planting decision-making model that relies on repeated trial and error in the field. By constructing a multi-scenario prediction framework encompassing climate, soil, and dynamic management parameters, it can quantify the production capacity response under different fertilization levels and planting years, providing farmers and land managers with pre-planting yield estimates. This method significantly reduces the time cost and financial risk of large-scale development of non-grain crops on marginal land, improves land use efficiency, and provides an actionable decision-making basis for the scientific layout of China's bioenergy industry.

[0082] Example 2: Poplar Yield Prediction Method Based on Random Forest

[0083] 1) Sample data acquisition and processing

[0084] We obtained global measured data on poplar trees (including yield, fertilization amount, planting years, and latitude and longitude), and standardized the fertilization amount and planting years to eliminate data discrepancies.

[0085] 2) Construction of feature variables

[0086] Consistent with Example 1.

[0087] 3) Model building and optimization

[0088] Construct a Radio Frequency (RF) machine learning model, train it using leave-one-out cross-validation, and implement it through R. 2 Performance was evaluated using root mean square error, and the model was optimized by screening model configurations and eliminating extreme production samples. Figure 5 The diagram illustrates the prediction effect of random forest on poplar yield before and after optimization in Embodiment 2 of the method according to the present invention. The left graph shows the prediction effect before optimization, and the right graph shows the prediction effect after optimization. The horizontal axis represents the observed poplar yield (t / ha), and the vertical axis represents the predicted poplar yield (t / ha). Where R... 2 N represents the coefficient of determination. masked Indicates the total number of masked lexical units. From Figure 5 It can be seen that the final R² of poplar wood improved to 0.85, reaching the high reliability prediction standard, and the fitting ability was excellent.

[0089] 4) Delineation of marginal land boundaries

[0090] Consistent with Example 1.

[0091] 5) Construction of Marginal Land Machine Learning Prediction Dataset

[0092] Consistent with Example 1.

[0093] 6) Marginal land poplar yield forecast

[0094] The constructed marginal land prediction dataset was input into the optimized RF model of the example to carry out poplar yield prediction, and a spatial distribution map of poplar yield on marginal land in China and a yield comparison dataset were generated. The total marginal land yield of poplar under different planting years and fertilization amounts is shown in Table 2, where the corresponding planting years are 10 years and the fertilization amount is 500 kg / ha.

[0095] Table 2. Marginal land total yield (Gt) of poplar under different planting years and fertilization levels.

[0096]

[0097] on the other hand, Figure 6 The maximum yield distribution of poplar trees in border (provincial) lands is shown in Embodiment 2 of the method according to the present invention.

[0098] Overall, poplar, with its rapid growth and wide adaptability to different locations, has become an irreplaceable woody resource in China's marginal land biomass economy.

[0099] This embodiment uses poplar as the research object. Poplar is a fast-growing tree species with rapid growth, short rotation period, and wide adaptability. It is widely considered a preferred non-grain woody energy and timber crop suitable for temperate and warm temperate marginal lands, reforestation areas, and riparian buffer zones in China. Random Forest (RF), an advanced machine learning algorithm, is selected as the carrier to fully demonstrate the complete implementation process of the non-grain crop yield prediction method based on machine learning of this invention. This embodiment fully demonstrates that even in complex marginal forest conditions where traditional growth model parameters are difficult to determine, regional adaptability is poor, and long-term observation costs are high, this invention can still achieve extremely high prediction accuracy and spatial generalization ability.

[0100] Example 3: A method for predicting Miscanthus yield based on XGBoost

[0101] This embodiment is similar to the method in Embodiment 1, the main difference being the use of some XGBoost models to predict the yield of Miscanthus sinensis on marginal land.

[0102] 1) Sample data acquisition and processing

[0103] Consistent with Example 1.

[0104] 2) Construction of feature variables

[0105] Consistent with Example 1.

[0106] 3) Model building and optimization

[0107] Construct an XGBoost machine learning model, train it using leave-one-out cross-validation, and implement it through R. 2Performance was evaluated using root mean square error, and the model was optimized by screening model configurations and eliminating extreme production samples. Figure 7 The figure illustrates the prediction effect of XGBoost on Miscanthus yield before and after optimization in Embodiment 3 of the method according to the present invention. The left figure shows the prediction effect before optimization, and the right figure shows the prediction effect after optimization. The horizontal axis represents the observed poplar yield (t / ha), and the vertical axis represents the predicted poplar yield (t / ha). Where R... 2 N represents the coefficient of determination. masked Indicates the total number of masked lexical units. From Figure 7 It can be seen that the R² of Miscanthus sinensis was improved to 0.940, reaching the extremely high reliability prediction standard, and the fitting ability was excellent.

[0108] 4) Delineation of marginal land boundaries

[0109] Consistent with Example 1.

[0110] 5) Construction of Marginal Land Machine Learning Prediction Dataset

[0111] Consistent with Example 1.

[0112] 6) Marginal land Miscanthus yield forecast

[0113] The constructed marginal land prediction dataset was input into the optimized XGBoost model of the implementation example to predict the yield of Miscanthus sinensis on marginal land in China. Subsequently, data from various provinces were collected for provincial yield comparison. The total yield of Miscanthus sinensis on marginal land under different planting years and fertilization rates are shown in Table 3, where the corresponding planting years are 10 years and the fertilization rate is 500 kg / ha. The fertilization rate listed here refers to the total of all kinds of chemical fertilizers; the maximum yield of Miscanthus sinensis on marginal land in each province is shown in Table 3.

[0114] Table 3. Total marginal land yield (Gt) of poplar under different planting years and fertilization levels.

[0115]

[0116] on the other hand, Figure 8 The maximum yield distribution of Miscanthus sinensis in marginal lands (in provinces) is shown in Embodiment 3 of the method according to the present invention.

[0117] Overall, the performance is similar to that of RF prediction, with XGBoost predicting a maximum Miscanthus yield of 12.31 Gt. However, the XGBoost model has higher prediction accuracy, with Rt... 2 The accuracy reached 0.940. This further demonstrates that even in complex marginal forest conditions where traditional growth model parameters are difficult to determine, regional adaptability is poor, and long-term observation costs are high, this invention can still achieve extremely high prediction accuracy and spatial generalization ability, and enable the learning and prediction of multiple models.

[0118] This invention utilizes machine learning methods to accurately predict the yield performance of non-grain crops on barren land, breaking away from the traditional forestry model that relies on long-term fixed-location observations. Through multi-scenario dynamic modeling, it can quantify the differences in yield under different planting times and fertilization scenarios, providing operators with yield estimates before planting. This method significantly reduces the time cost and financial risk in large-scale poplar farming, revitalizes some low-growth land resources, and provides a data-driven decision-making tool for the construction of biomass energy raw material bases and timber security strategies.

[0119] It should be noted that the embodiments described above are merely preferred embodiments of the present invention, used to explain the present invention, and do not constitute any limitation on the present invention. The present invention has been described with reference to typical embodiments, but it should be understood that the terms used therein are descriptive and explanatory terms, not limiting terms. Modifications can be made to the present invention within the scope of the claims, and revisions can be made to the present invention without departing from the scope and spirit of the present invention. Although the present invention described herein relates to specific methods, materials, and embodiments, it does not mean that the present invention is limited to the specific examples disclosed herein; on the contrary, the present invention can be extended to all other methods and applications having the same function.

Claims

1. A method for predicting the potential yield of non-grain crops on marginal land based on machine learning, characterized in that, Includes the following steps: Step 1: Obtain the management parameters and geographic coordinates of the planting location for non-grain crops, wherein the management parameters are measured planting data; Step 2: Based on the geographic coordinate information, extract the corresponding climate parameters, soil parameters, and terrain parameters from the climate parameter source, soil parameter source, and terrain data source, respectively, and construct a feature dataset for non-grain crop yield prediction based on the management parameters, the climate parameters, the soil parameters, and the terrain parameters; Step 3: Construct multiple machine learning prediction models based on the feature dataset for non-grain crop yield prediction, and train and evaluate the performance of the machine learning prediction models; Step 4: Adjust parameters through machine learning to optimize the prediction accuracy of the various machine learning prediction models; Step 5: Select the optimal machine learning prediction model based on the evaluation results of the various machine learning prediction models; Step 6: Based on the optimal machine learning prediction model, predict the yield of non-grain crops within the marginal land area and generate the spatial distribution results of non-grain crop yields.

2. The potential production forecasting method according to claim 1, characterized in that, In step 1, the management parameters include at least one of the following: crop yield, planting years, planting density, crop growth cycle, crop rotation / continuous cropping years, fertilizer type, amount of single fertilizer applied, amount of pesticide applied, growth time, and irrigation amount.

3. The potential production forecasting method according to claim 1, characterized in that, In step 2, the climate parameters include at least two of the following: temperature, precipitation, solar radiation intensity, growing season, wind speed, humidity, and evaporation rate.

4. The potential production prediction method according to claim 1, characterized in that, In step 2, the soil parameters include at least two of the following: pH value, organic carbon content, nitrogen, phosphorus and potassium content, density, porosity, texture, water content, and bulk density.

5. The potential production prediction method according to claim 1, characterized in that, In step 3, the various machine learning prediction models include at least two of the following: random forest, gradient boosting tree, extreme gradient boosting, support vector machine, K-nearest neighbors, artificial neural network, convolutional neural network, elastic network, Lasso regression, Ridge regression, CatBoost, and Gaussian process regression.

6. The potential production forecasting method according to claim 1, characterized in that, Before filtering the machine learning prediction model, in step 4, outlier removal or feature filtering is performed on the feature dataset concerning the non-grain crop yield prediction.

7. The potential production forecasting method according to claim 1, characterized in that, The non-grain crops include at least one of the following: Miscanthus sinensis, switchgrass, foxtail grass, reed, eucalyptus, poplar, willow, sweet sorghum, and sugarcane.

8. The potential production forecasting method according to claim 1, characterized in that, The marginal land includes at least one of the following: low-productivity farmland, saline-alkali land, dry land, degraded land, wasteland, sandy land, low-coverage grassland, shrubland, sparse woodland, and tidal flats.

9. A non-grain crop yield prediction system for implementing the potential yield prediction method according to any one of claims 1 to 8, characterized in that, include: A feature dataset construction module is used to obtain the management parameters, climate parameters, soil parameters, and terrain parameters of non-grain crops to construct the feature dataset for non-grain crop yield prediction. The model building and optimization module is used to build and optimize the machine learning prediction model. The feature data construction module for marginal land is used to construct a feature dataset for marginal land as input for non-grain crop yield prediction. The feature dataset for marginal land includes at least one of the management parameters, climate parameters, and soil parameters of the marginal land. The yield prediction module is used to input the feature dataset for marginal land into the optimal machine learning prediction model to generate yield prediction results for non-grain crops.

10. A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the potential yield prediction method of any one of claims 1 to 8.

Images