Abstract
Using data mining, the purpose of this study is to forecast and analyze the growth of the cotton cultivation industry and the policy financial support demands in the Aksu region. Data mining is a method for maximizing the value of data via the application of numerous algorithms. In contrast to conventional data mining, which adheres to specific algorithms, data mining employs a variety of analysis algorithms to analyze raw data, such as image and panel data, and produce accurate results. In this paper, we propose a data mining method that combines the semantic segmentation algorithm of remote sensing images with various nonlinear regression algorithms to predict the demand for policy-based financial support in a specific region based on a combination of multiple factors, including agricultural crop cultivation area, catastrophe analyses, agricultural price and inflation rates, etc. This paper intends to estimate and analyze actual data pertaining to the cotton cultivation industry in Aksu, and this methodology can further improve the policy-based financial inverse model. The methods presented in this paper can further improve countercyclical regulation of policy finance.
Introduction
As the country’s primary industry, it is self-evident that agricultural production work plays a crucial role in its economy. It should be noted that, as part of the continued strengthening of the national strategy for rural revitalization, the state has steadily increased financial support for rural areas. Due to the peculiar characteristics of rural finance, such as information asymmetry and high contract costs, general financial institutions rarely provide financial assistance to farmers, and government-sponsored policy financial assistance has become an important tool for the regulation and control of agricultural production. Effectively forecasting policy-based financial support can enhance policy-based financial support on the ground, thereby contributing to the expansion of agricultural work in China.
The need for a high level of model interpretability is the impetus for the problem investigated in this paper. This paper proposes a new data mining method that combines deep learning and machine learning algorithms, and applies it to the most important cotton-producing region in our country, the Aksu region. This method will be useful for future agricultural finance research.
Review of the literature
Modern researchers have been confronted with a number of difficult questions. Academics continue to investigate and analyze the many facets of rural finance now that its development is well under way. Clearly, they note that financial support can alleviate financial inhibitions in certain counties to a certain extent, and that this alleviation has a greater impact in densely populated and economically developed regions [1]. Guijun Zhang and Minhua Ouyang used the propensity matching method to analyze the income enhancement status of rural households and found that policy-based financial programs increase the disposable income of low-income households [2]. By applying the Oaxaca-Blinder decomposition method to insurance, banking, and Internet finance, Fang Su and Lei Fang investigated the underlying causes of supply exclusion, self exclusion, no demand, and physical exclusion in financial exclusion [3] and obtained promising analytic results.
The preceding studies provide a statistical analysis of agricultural finance issues, from which clear conclusions can be drawn. Nonetheless, if the data for model analysis come from statistical yearbooks, regional panel data, or data from banks on their own, the model loses its predictive power and can only perform statistical analysis on data that becomes available after the publication of yearbooks and panel data, which is typically at the beginning of the following year; thus, it has lost its early warning capability.
In the field of policy finance, increasing the accuracy of forecasting capabilities has been a valuable tool for addressing crises and rationally organizing capital operations. Similarly, it is possible to provide immediate financial aid in the event of a natural disaster or a bountiful harvest. For economic forecasting models, the academic community typically employs a linear or nonlinear forecasting algorithm. Song Lian, for example, suggested using the nonlinear gray model (GM) (1, 1). Even after accounting for independent variables such as fertilizer, labor, water, and electricity, the Aksu cotton production forecast for 2020 still contains an error rate of more than 100,000 tons [4]. On the basis of single linear or nonlinear prediction algorithm, although some scholars have studied the method of adding remote sensing images and then using basic NDVI parameters and SAR parameters for image analysis, trying to accurately predict the yield of planting industry. For example, VN Sridhar et al. predicted the yield of wheat in Madhya Pradesh from 1991 to 1992, and used the RVI parameters of remote sensing images to identify on the basis of ARIMA model, that the results are still unsatisfactory. The yield error of wheat in 1992 was as high as 60.02% [5]. J. Betbeder et al. proposed SAR parameter estimation for the yield estimation of rapeseed and wheat through remote sensing, but its effect did not show a consistent improvement in the performance of the two crops (compared with NDVI parameter method) [6]. The reason is that the prediction error of planting area produced by the above research based on the regularization method of processing remote sensing images is large. Therefore, when combining the output of unit output to predict the output, there is a large prediction error compared with the actual output. It also prevents the current scholars from further exploring the deeper research such as the prediction of the output value of planting industry. At the same time, previous studies also face the interference brought by time series to varying degrees, and there is a problem of model failure in long-span prediction.
Using data from previous years for forecasting current production output, regardless of the data’s quality or number of variables, makes forecasting more difficult as the time span increases. Due to disaster cycle fluctuations, changes in planting scale, crop market fluctuations, and other factors, determining how to more accurately predict the output of the cotton plantation industry in advance has become a pressing financial policy issue. The paper proposes a new high-precision plantation forecasting method after exhaustively examining the limitations of existing plantation forecasting methods. Initially, during the cotton planting process, the cotton plantation area is analyzed by the semantic segmentation algorithm of remote sensing images, and the yield is predicted based on variables including time series fluctuations in plantation area and disaster cycle fluctuations. This method is more accurate than those that rely solely on data from previous years, and it also uses remote sensing images at various stages of the planting process to predict crop yield. This paper compares the remote sensing images of cotton fields at different stages between 2013 and 2019 to analyze the yield of disaster reduction using remote sensing images from different periods of the planting process. This paper attempted to predict the output value of the cotton planting industry from 2014 to 2019 by calculating the output value using the price-value system, inflation coefficient, and yield per unit area in order to conduct a comprehensive analysis of the output value of the cotton planting industry. These results demonstrate that the proposed method can accurately predict the output value of the cotton plantation industry in the region before the yearbook is published, and that the predicted financial support demand is conducive to the sound and orderly development of the cotton harvesting and deep processing industries, as well as to the policy financial institutions’ efforts to further control risk and improve their capacity to serve agriculture.
Aksu cotton yield prediction based on time series analysis of remote sensing images
Utilizing remote sensing images for yield estimation has become an increasingly popular method of remote sensing image application in agriculture. Various index systems for multispectral image analysis are used in traditional satellite image processing, with the Normalized Difference Vegetation Index (NDVI) or Ratio Vegetation Index (RVI) being the most frequently used one in remote sensing image yield estimation, as shown in Eq. (1).
Both NDVI and RVI methods in Eq. analyze surface properties by comparing the near-infrared band to the red band, where NIR represents the value of the near-infrared band reflection and R represents the value of the infrared band reflection. In general, if the IR band reflection value is high, the density of vegetation is high, and vice versa, the vegetation cover may be sparse or have died for a variety of reasons if the value is low. The NDVI and RVI indices have a strong correlation for predicting crop yields because crops are typically planted at high densities. This made vegetation indices the primary method for assessing crop planting until the development of deep learning algorithms. However, images of sparse vegetation captured by remote sensing could be the result of natural disasters, pests, or diseases. In addition, other non-target crops, green landscapes, geothermal sources, and other non-crop interference situations may also contribute to dense vegetation. Consequently, using simple exponential calculation methods directly in practical applications can be error-prone, as the identified area is typically much larger than the annual statistics. In a study conducted by Sijia Li, the NDVI index was utilized to estimate the prediction of maize yield in Lishu County, with the predicted results in many years exceeding the actual results. In 2010, for example, the predicted area was 114769 ha, while the actual area was 97003 ha [7]. He Fuwei calculated the winter wheat planting area in North China by normalizing the LAI from remote sensing images and analyzing the yield and total yield by combining data from agricultural monitoring stations; the results were comparable to reality [8]. In recent years, multispectral remote sensing image recognition has gradually surpassed traditional image processing algorithms in terms of accuracy and recall, thanks to the progressive development of deep learning technology. Deep learning algorithms are able to recognize complex texture features in images and have strong generalization capabilities, allowing them to gradually replace complex image recognition analysis [9]. Hao-Lu Li, for instance, used the enhanced DenseNet to segment and identify the planted area of cotton fields in Kuche County, Xinjiang, and then used NDVI, wind speed, and precipitation to analyze the yield after calculating the yield, employing the ConvGRU algorithm to predict the yield error in recent years about 200 kg/ha and the total yield error about 4
In light of the aforementioned examples, it is not difficult to identify a dearth of current research.
First, the area identified by remote sensing images is directly compared with the planting area and other government-published information, ignoring the possibility that the identification results in remote sensing images have a linear relationship with the actual planting area (e.g., some misidentified image elements may continue to be misidentified in future images, causing a linear error in the predicted planting area).
Second, the cyclical variation data of natural disasters was not incorporated into the pattern of temporal variation of planted area.
Thirdly, neither the CPI, production, nor crop price forecasting analyses nor the work on production value forecasting were analyzed.
Fourth, the analysis was not aligned with policy-based financial requirements, and the projection results were not translated into substantial agricultural development instruments.
Despite the use and improvement of numerous advanced algorithms in deep learning, the current research has low prediction accuracy, is not integrated with rural agricultural development requirements, and is primarily focused on yield prediction.
This paper proposes a more accurate deep learning algorithm, Cascade-PSPNET image segmentation, for remote sensing images in order to accurately identify crop planting at the pixel level. We have computed yield estimates using ARIMAX and multiple precipitation, wind, NDVI, and seed quality factors. The loss area is determined by the linear relationship between the predicted area of a time series and the disaster-affected area. The yield prediction accuracy is further enhanced by the linear regression correction between the predicted yield and the actual yield, and the next step is to develop a production forecast while providing financial support. A two-stage object detection model, was proposed to solve the problem of detection area discovery through different threshold output, improve the accuracy of object detection, and was improved several times in subsequent studies [11]. At the same time, as the current classic remote sensing image semantic segmentation network, PSPNET has a good effect on the segmentation of regions of different sizes because it uses a pyramid pooling layer to carry out different scales on the middle features of the network [12]. This paper attempts to combine the advantages of the two networks, and proposes Cascade-PSPNET through the combination of threshold division technology and PSPNET network structure, hoping to further improve the recognition ability of semantic segmentation network for remote sensing images.
It is essential to determine the time-phase of images gathered by remote sensing to accurately estimate crop yields. According to their production areas and climates, the cotton-producing regions of Xinjiang can be divided into eastern Xinjiang, southern Xinjiang, and northern Xinjiang. The southern region of Xinjiang is distinguished by ample sunlight and a longer frost-free period. The cotton-growing region of Aksu is located in southern Xinjiang. According to local meteorological observatories, media reports on sowing date forecasts, and agricultural reports, the optimal sowing date is typically around mid-April, prior to and following the Grain Rain [13, 14]. The seeds germinate approximately 15 days after sowing. The blooming period occurs between the beginning and middle of July. Generally speaking, early September is harvest time in southern Xinjiang, while early October is harvest time in northern Xinjiang. In the region of Aksu, harvest typically occurs between late September and early October. Cotton varieties are grown for a variety of purposes, including lint, upland, and long-staple cotton. There is no significant difference between these varieties’ planting cycles. Additionally, wheat and maize are the principal crops of the Aksu region. The planting area for spring wheat in 2021 [15] is 2.75 million mu, spring wheat is 30,000 mu, winter wheat is harvested in April and May, and spring wheat is harvested in August, according to statistical data. Similarly, the corn harvest occurs at the end of August. The optimal months for acquiring remote sensing images are therefore early June and early September. Winter wheat is typically harvested in early June, while corn is typically planted in the middle of June. Thus, other crop characteristics can be excluded more efficiently. Spring wheat and corn are harvested in early September, and cotton is nearing harvest. Therefore, identifying cotton planting characteristics in remote sensing images is advantageous. The schematic diagram of remote sensing phases and time is shown in Fig. 1.
Diagram of remote sensing image time phase selection.
It is difficult to apply the Cascade-PSPNET image segmentation algorithm proposed in this paper to achieve pixel-level semantic segmentation because a selection of June images cannot eliminate the overlap of winter wheat harvesting and corn planting stages. Four channels of green and near-infrared bands are trained through the visible light band and near-infrared band’s characteristics and have the ability to identify different types of plants. Figure 2 illustrates the application of the Cascade-PSPNET structure to multispectral remote sensing images as described in this paper.
Res-Conv structure.
Based on Eq. (2), each residual head is assigned the binary loss function Focal-Loss, and the total loss function Loss-total of the first-stage network is equal to the sum of the three residual head loss functions Final-Loss. The Focal-Loss formula in Eq. (2)
Cascade-PSPNET network parameters table
To train the Cascade-PSPNET two-stage network, the Res-Conv4 residual convolution module is used. The second-stage loss function is a single binary Focal-Loss function with a threshold of 0.5 to output the final fused segmented images. This design concept has been developed based on the advantages of the two-stage network. In the target detection network, the Cascade-RCNN network has advantages over other two-stage networks because the first stage employs multiple thresholds to effectively deal with difficult-to-score samples, thereby reducing the imbalance in the output of the feature map in the first stage. The second-stage network analyzes the classification problem, thereby removing the issue of an excessive number of negative samples during the target detection stage. When it comes to frame and pixel segmentation metrics, the semantic segmentation and target detection algorithms have many similarities; both can use metrics such as IOU to assess location accuracy and category loss to assess the classification accuracy of a single frame or pixel and perform a comprehensive analysis. Therefore, attempting to incorporate Cascade’s two-stage network structure into the conventional segmentation model is worthwhile. There are a number of parameters in each stage of the model depicted in Table 1, and the total number of parameters in this paper is 21.9
Taking into account model training data sources, this study should analyze multispectral remote sensing images for crop identification in order to better distinguish crop status, and remote sensing images should account for multiple factors affecting identification accuracy, such as spatial resolution, cloud cover, band, satellite strip width, and return period. Lantsat8 and Gaofen-1 provide the best resolution accuracy, return period, band and strip width among current civil high-precision satellites, and both can be acquired through commercial and educational channels. A detailed comparison is shown in Table 2.
Comparison of high-resolution civil satellite images
Gaofen-1 has a greater advantage, primarily due to its short return period, which makes it possible to intercept remote sensing images for stitching within a short period of time, preventing cloud cover from causing errors in the determination of the planting surface. Lansat8, on the other hand, has a 16-day return period, making it easy to miss the most advantageous time period. In addition, the strip width of the Weigan River Oasis is 322 kilometers north to south and 194 kilometers east to west, which is greater than the strip width of Lansat8. Therefore, the Gaofen-1 satellite image for 2015-2021 is chosen as the input image, and the Weigan River Oasis in the Aksu region with the latitude and longitude ranges (41
Set image training set
While Model is Training:#While Model is Training
For
In the field of image enhancement, this paper introduces the concept of combining probability and image enhancement methods. For instance, after vertical flipping with a random probability of 0.5, the image has a 50% chance of becoming either the top or bottom mirror image or the original mirror image, and after horizontal inversion, the image may or may not be vertically flipped, which greatly improves the data set for spatially dense distribution. Since the semantic segmentation of images in this study is a position-sensitive task, flipping the images may improve the robustness of the model for pixel position judgment, and a 10% random brightness noise is added to the image enhancement process to improve the recognition effect under different lighting conditions and seasonal changes. Thus, the spatial distribution of the data set has been significantly enhanced.
In terms of parameters, training parameters were set using gradient descent optimization for Adamw, and the training parameters as well as evaluation indexes (IOU, F1Score, Focal Loss value, etc.) are shown in Table 3.
Comparison of training parameters and model metrics
The comparison of this model’s segmentation effect to that of similar models trained with the same parameters in Table 3 demonstrates the sophistication of this model. By incorporating Focal-Loss, it is possible to assess the classification status of the model’s challenging samples. While the two-stage segmentation model presented in this paper slightly increases the model’s computational volume and number of parameters, it outperforms similar segmentation models in a variety of indexes and has significant application potential.
In this paper, the trained Cascade-PSPNET network and other comparison networks are used to compare the cotton field area recognition effect in random image images, as shown in Fig. 3 of the model comparison image recognition effect graph. On Fig. 3, the original image is depicted in pseudo-color, and the recognition effect of the text-trained Cascade-PSPNET, UNET
Results of cotton fields identified by different models with original images: (a) pseudo-color multispectral remote sensing images, (b) cotton field mask labels, (c) Cascade-PSPNET, (d) UNET
Figure 3 shows the comparison between the planting area and the actual planting area of Aksu Region in all years between 2014 and 2019 (from the manual statistics of the Xinjiang Statistical Yearbook). The results of the analysis of Table 3 combined with Fig. 3 can be found in the remote sensing image recognition results. Some fixed recognition error image elements are easily misidentified as cotton fields, resulting in an overestimation of the cotton field recognition area. Therefore, it can be inferred that there should be a first-order linear relationship between the identified cotton field area and the actual cotton field area, and that neither the bias term nor the weight term should be zero.
Furthermore, since the cotton yield in Aksu region is adversely affected by major disasters such as hailstorms, mudslides, windy weather and heavy rainfall throughout the year, the timing of these disasters is unpredictable and difficult to predict. As a result, a prediction method is applied to complete the first-order residuals. Since current methods of forecasting acreage include LSTM, GRU, and other neural network methods, but the amount of data in this paper is too small, making it very easy for neural networks to overfit and thereby making future predictions difficult, this study attempts to predict the primary term residual of the current acreage by ARIMAX by using the fluctuation of acreage and disaster area from June to September in order to more precisely estimate the acreage. As shown in Eq. (3),
Where
Comparison of deep learning recognition prediction results and correction results of Eq. (3).
Since the relationship between area unit yield
This is then combined with
Financial support forecasting system solution.
Figure 5 depicts the flowchart for the financial support system based on the data mining methodology proposed in this paper. In the first section of this paper, node 1 completes the preliminary prediction for cotton planting area, node 2 completes the accurate prediction for cotton planting area, node 3 completes the accurate prediction for cotton yield, and node 4 completes the accurate prediction for cotton production value. In order to achieve the objective of accurate financial support estimation, the scheme incorporated the concept of residual neural networks by utilizing cross-layer variables and removing residuals.
The residual prediction complementary method presented in this paper is employed in node 5 for the analysis, and the amount of financial support required in this year’s Aksu region is predicted by the same method using unit prices, the amount of financial support in previous years, and taking inflation into account, as shown in Eq. (6).
In Eq. (6),
It is worth noting that there is a simple linear relationship between acreage, yield, and output in Eqs (3)–(6). The linear relationship between policy financial support and output value has not been demonstrated in previous studies, but Fig. 6 has verified the stable sequence relationship of residuals, so the rationality of using ARIMAX is fully guaranteed. Finally, this paper analyzes the linear relationship between the loans of financial institutions and the output value of the cotton industry through

Table 4 provides a summary of the methods that will be used in this paper to predict financial support based on the algorithm flow depicted in Fig. 5. The parameters of
Algorithm flow parameters table
Loan forecast error curve for financial institutions by algorithm over the years.
According to the same method, the amount of financial assistance needed in the Aksu region this year was predicted by using cotton yields, unit prices and production output values, as well as the amount of financial assistance provided in the cotton farming industry in the previous year, and taking inflation into account. This forecast is based on Xinjiang Statistical Yearbook data from 2013–2018 (Open source public release) [20] and satellite images of the corresponding time period. Forecasts are made for data from 2014–2019. The analysis results are shown in Fig. 7, which illustrates the prediction results of this paper regarding financial support and crop output. As evidenced by the results of previous years, the financial support prediction method proposed in this paper has significantly improved accuracy in almost all years compared to linear regression and ARIMA prediction, and the final predicted MAE has reached the level of applicability. It has a high potential for application.
This paper preliminarily summarizes the methods of predicting the output value of planting industry in the previous studies, and analyzes why the current research has not completed the full link prediction from predicting the planting area – output value – to further financial support quota of planting industry. The root cause lies in the prediction deviation of planting area, which makes it difficult to obtain high accuracy of link prediction. In order to better complete the fine planting area prediction, inspired by the network structure of CascadeRCNN and PSPNET, this paper proposes Cascade-PSPNET to complete a more accurate planting area prediction, which improves at least 5% IOU compared with other semantic segmentation algorithms, so that the prediction yield and prediction output value goals proposed in this paper have been achieved. Finally, by using the concept of residuals, this paper constructs an automated sequence stabilization method, which extends and reorganizes the ARIMAX prediction link to form a financial prediction system scheme, and achieves the goal of accurately predicting the total amount of policy financial support for the cotton planting industry in the Aksu region. This paper hopes to provide a new perspective and method for agricultural financial analysis.
