Deep learning-based cross-regional crop optimal harvest period monitoring method

Abstract

Provided in the present invention is a deep learning-based cross-regional crop optimal harvest period monitoring method, comprising: acquiring high-temporal-spatial-resolution remote sensing data of a target region in a corresponding crop growth season, and generating multiple groups of vegetation index time series; on the basis of a classification model, performing pixel extraction on each group of vegetation index time series, so as to obtain temporal features of pixels and, on the basis of the temporal features, obtaining a classification result; on the basis of a deep learning algorithm, performing deep feature extraction on space and spectral space of the high-temporal-spatial-resolution remote sensing data, so as to obtain feature-level data; constructing an initial neural network model and, on the basis of the feature-level data and the classification result, training the initial neural network model; and inputting data under test into a cross-regional crop optimal harvest period monitoring model, so as to obtain a monitoring result. The present invention can use large-scale remote sensing images to monitor crop optimal harvest periods with high precision, so as to improve the degree of automation and intelligence of cross-regional crop optimal harvest period monitoring.

Description

A Cross-Regional Crop Harvest Readiness Monitoring Method Based on Deep Learning Technical Field The present invention relates to the technical field of remote sensing monitoring, and particularly to a cross-regional crop harvest readiness monitoring method based on deep learning. Background Art Agricultural production has strong seasonality, and timely operations have a great impact on yield, especially the timeliness of harvesting operations. For example, if rice is harvested too early, the crop is not fully developed, the grain moisture content is high but the dry matter content such as starch is low, the 1000-grain weight is low and the broken rice rate during processing is high; if harvested too late, the 1000-grain weight of the grain decreases, and the probability of reduction in yield or even crop failure due to adverse weather increases. It is more vulnerable to attacks from insects, birds, animals and microorganisms, resulting in yield reduction. Timely harvesting is crucial for reducing losses in the crop harvesting process. (1) Research Status of Timeliness Loss Theory Since ancient times, agricultural production in China has emphasized observing the farming seasons. "Among all the principles of farming, waiting for the right season is the most precious" is mentioned in *Lv's Spring and Autumn Annals*. Timely operations are the premise for stable yields. The losses caused by operating against the schedule are called timeliness losses, which consist of three parts: agronomic, natural and operation process losses. Among them, the timeliness losses in the planting and harvesting links are the key points. In terms of planting, Wang Jinwu et al. (2004) set up 22 plots in the same paddy field and transplanted rice seedlings in one plot every day for 22 consecutive days. After the rice matured, the yield of each plot was measured, and the quantitative relationship between rice yield and transplanting time was calculated. However, he did not reveal the mechanism of yield change. The transplanting time of rice has a significant impact on yield, mainly because there are significant differences in the effective tiller number, panicle number, population growth rate, temperature and light utilization efficiency, and photosynthetic matter production characteristics of rice at different transplanting times. Huang Zheng et al. (2021) used five hybrid rice varieties such as Chuanyou 6203 as experimental materials and adopted a two-factor split-plot design to study the effects of staged direct seeding on the growth process and yield of hybrid rice in order to explore the suitable sowing date for regional rice. It was found that the rice yield showed a trend of first increasing and then decreasing as the transplanting time was postponed. In terms of harvest, Wang Jinwu (2004) used the randomized block experiment method to measure the yield of rice transplanted on the same field and on the same day every other day, and obtained the quantitative relationship between rice yield loss and harvest period. Huyen et al. (2010) measured the chemical components of miscanthus organs that can be converted into fermentable sugars at different harvest times and found that earlier harvest can obtain greater benefits than later harvest. Godin et al. (2013) evaluated the impact of harvest time on the yield of various energy crops converted into biofuels and found that the biofuel yield will decrease if harvested too early or too late. Heidari et al. (2012) measured the impact of harvest time on the yield of fennel through experiments, and the results showed that the fennel yield was the highest at maturity. Qiao Jinyou et al. (2017) studied the relationship between the actual harvested yield of soybeans and the harvest date and obtained the law of timeliness loss of soybean mechanical harvesting. Du Zhimin et al. (2018) found through experiments that as the harvest period is postponed, the brown rice rate and milled rice rate of japonica rice gradually increase first, reach the maximum value 65 days after full heading and then decrease significantly. The head rice rate fluctuates greatly before 65 days after full heading, without obvious regularity, and decreases significantly after 65 days after full heading. (2) Research progress on the method for determining the appropriate harvest period Li Zhenqing (1986) proposed that the harvest day with the lowest timeliness loss of crop harvest is the optimal harvest day, and the harvest period distribution with the smallest total timeliness loss of rice harvest is the appropriate harvest period. Since the rice harvest loss changes with the harvest period in a quadratic equation, the optimal harvest day should be in the middle of the appropriate harvest period. At present, the main methods for determining the appropriate harvest period of crops are: determination based on crop canopy characterization and determination based on the content of specific substances in agricultural products. Determination based on the surface characteristics of crop canopies. Hyperspectral Remote Sensing is widely used in agricultural crops and is mature in the inversion of crop bio-chemical parameters, agricultural production monitoring, and agricultural information monitoring. Wan et al. (2018) designed a mature detection device based on computer vision technology to determine the maturity by extracting the color feature values on the surface of tomatoes, with an average accuracy of 99.31%. Pereira et al. (2018) used digital imaging technology and the random forest algorithm to predict the maturity of papaya fruits and verified the prediction results by measuring the pulp hardness of papayas. Such determination methods are fast and convenient, do not damage crops, but the equipment cost is relatively high. Shen Yu et al. (2020) used hyperspectral imaging technology to identify the effective bands of an automatic grading instrument for slightly damaged apples, and the recognition rate of the support vector machine (SVM) damage recognition model for slightly damaged apples reached 90.63% at the 811 nm band. Ge Yonghui et al. (2022) used hyperspectral imaging technology (400 - 1000 nm) to study the non-destructive discrimination of kiwifruit chilling injury. The model with the characteristic wavelengths selected by the successive projections algorithm (SPA) is superior to the model with all wavelengths, and the correct rates of the calibration set and prediction set are 100% and 94.2% respectively. Guo Jingjing et al. (2022) proposed that the model established based on the SG+FD+CARS+LSSVM combined method is the optimal lettuce greenness determination model, which can achieve 64.59% of the sensitive wavelengths extracted accounting for all wavelengths. Compared with the original hyperspectrum (1.25%), the number of extracted sensitive wavelengths increased by 63.34%. Determination based on the content of specific substances in agricultural products. Wei Changling et al. (2017) extracted the volatile oil from perilla leaves by steam distillation method and determined the relative contents of its characteristic components perillaldehyde, perillaketone or perillene, and thus determined the maturity of perilla. Such discrimination methods have low cost, low efficiency, and poor timeliness

[0053] . Pan Yihong et al. (2017) took the middle leaves of different varieties as the research object and found that from the appropriate maturity, maturity to over-maturity, the contents of chlorophyll a, chlorophyll b, total chlorophyll and SPAD value in tobacco leaves decreased significantly by measuring the chlorophyll content and SPAD value of tobacco leaves. Zhang Yuqin et al. (2018) used an automatic colorimeter to measure the change characteristics of color parameters (L*, a*, C*) and color difference parameters (△L*, △a*, △C*) of tobacco leaves with different maturities, and established a Fisher discriminant function with color parameters as indicators to discriminate tobacco leaves with different maturities. Ren Erfang et al. (2019) compared and analyzed the results of measuring the soluble solids content, Vc content and pH value of fresh strawberry juice by electronic nose and electronic tongue, and thus determined the maturity of strawberries. Feng Jianying et al. (2020) conducted a follow-up study on the working principles and results of three typical intelligent sensory technologies, namely electronic nose technology, electronic tongue technology and texture analysis technology, and analyzed the accuracy of determining the maturity of strawberries from the soluble solids content, Vc content and pH value. At present, the methods for monitoring the appropriate harvest period of crops in China are inefficient, costly, and the data is scattered and the formats are not unified. Therefore, large-scale centralized management and update as well as subsequent decision-making use cannot be achieved. Therefore, how to use large-scale remote sensing images to monitor the appropriate harvest period of crops and obtain a series of corresponding effective crop growth data sets has become the key research goal and direction, and it is also a key technology for popularizing precision agriculture. Therefore, in view of the above problems and their subsequent impacts such as the low informatization, intelligence, and decision-making dependence of farmland agricultural machinery in China, how to achieve the automation and intelligence of cross-regional crop appropriate harvest period monitoring in China is the primary technical problem to be solved. Summary of the Invention In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a method for monitoring the cross-regional crop appropriate harvest period based on deep learning. To achieve the above purpose, the present invention provides the following solutions: A method for monitoring the cross-regional crop appropriate harvest period based on deep learning, comprising: Obtaining high spatio-temporal resolution remote sensing data within the target area during the corresponding crop growth season, and generating multiple groups of vegetation index time series; Performing pixel extraction on each group of the vegetation index time series based on a pixel-level crop biological growth situation classification model to obtain the temporal features of the pixels, and obtaining a classification result according to the temporal features; Performing deep feature extraction on the spatial and spectral spaces of the high spatio-temporal resolution remote sensing data based on a deep learning algorithm to obtain feature-level data; Constructing an initial neural network model, and training the initial neural network model based on the feature-level data and the classification result to obtain a trained cross-regional crop appropriate harvest period monitoring model; Inputting the data to be measured into the cross-regional crop appropriate harvest period monitoring model to obtain a monitoring result; Among them, performing deep feature extraction on the spatial and spectral spaces of the high spatio-temporal resolution remote sensing data based on a deep learning algorithm to obtain the feature-level data, including: Obtaining a training remote sensing data set; Performing feature quantization on the corresponding information of the high spatio-temporal resolution remote sensing data based on the vegetation information, spatial information, and spectral space information in the training remote sensing data set to determine the key factors for the interaction between the high spatio-temporal resolution remote sensing data and each piece of information; Constructing a feature extraction model for the high spatio-temporal resolution remote sensing data, training the feature extraction model based on the key factors and the training remote sensing data set, using the key factors as the training labels, and iteratively training to obtain the network parameters of the feature extraction model; The trained model extracts features from the high spatio-temporal resolution remote sensing data to obtain the feature-level data. Preferably, the high spatio-temporal resolution remote sensing data includes the microwave data Sentinel-1 dataset and the multispectral Sentinel-2 dataset. Preferably, based on the pixel-level crop biogrowth classification model, pixel extraction is performed on each group of the vegetation index time series to obtain the temporal features of the pixels, and the temporal features are classified to obtain a classification result, including: Construct a pixel-level crop biogrowth classification model; Use the pixel-level crop biogrowth classification model to extract the temporal features of each pixel from the vegetation index time series; According to the temporal features, perform pixel-level classification on the corresponding crops in the target area, and perform fusion extraction according to the single-channel, double-channel or multi-channel fusion method to extract crop pixels and form the corresponding classification result. Preferably, after performing pixel extraction on each group of the vegetation index time series based on the pixel-level crop biogrowth classification model to obtain the temporal features of the pixels and classifying the temporal features to obtain a classification result, it further includes: Perform image denoising, image enhancement and image registration on the high spatio-temporal resolution remote sensing data respectively to obtain the preprocessed high spatio-temporal resolution remote sensing data. Preferably, the initial neural network model is a UNet convolutional neural network structure. Preferably, the pixel-level crop biogrowth classification model is any one of a long short-term memory model or a Transformer model. Preferably, training a feature extraction model for the high spatio-temporal resolution remote sensing data based on the key factors and the training remote sensing dataset includes: Construct a feature estimation multi-dimensional dataset according to the key factors; Each element in the feature estimation multi-dimensional dataset reflects the association between vegetation information, spatial information, spectral spatial information and the high spatio-temporal resolution remote sensing data, and the feature extraction model is used to extract the vegetation information, spatial information and spectral spatial information in the high spatio-temporal resolution remote sensing data. According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention provides a method for monitoring the suitable harvest period of crops across regions based on deep learning, including: obtaining high spatio-temporal resolution remote sensing data within a target region during the growth season of corresponding crops, and generating multiple sets of vegetation index time series; performing pixel extraction on each set of the vegetation index time series based on a pixel-level crop biological growth condition classification model to obtain the temporal features of pixels, and obtaining a classification result according to the temporal features; performing deep feature extraction on the spatial and spectral spaces of the high spatio-temporal resolution remote sensing data based on a deep learning algorithm and the classification result to obtain the feature-level data; constructing an initial neural network model, and training the initial neural network model based on the feature-level data and the classification result to obtain a trained cross-region crop suitable harvest period monitoring model; inputting the data to be measured into the cross-region crop suitable harvest period monitoring model to obtain a monitoring result. The present invention can use large-scale remote sensing images to monitor the suitable harvest period of crops with high precision, and improve the automation and intelligence levels of cross-region crop suitable harvest period monitoring. BRIEF DESCRIPTION OF THE DRAWINGS In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings without creative efforts based on these drawings. FIG. 1 is a flowchart of the method provided by an embodiment of the present invention; FIG. 2 is a schematic diagram of the technical route provided by an embodiment of the present invention; FIG. 3 is a schematic diagram of the HSI model provided by an embodiment of the present invention; FIG. 4 is a structural diagram of the UNet network provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention. The purpose of the present invention is to provide a method for monitoring the suitable harvest period of crops across regions based on deep learning, which can use large-scale remote sensing images to monitor the suitable harvest period of crops with high precision, and improve the automation and intelligence levels of cross-region crop suitable harvest period monitoring. In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments. FIG. 1 is a flowchart of the method provided by the embodiment of the present invention. As shown in FIG. 1, the present invention provides a method for monitoring the suitable harvest period of cross-regional crops based on deep learning, including: Step 100: Obtain high spatio-temporal resolution remote sensing data within the target area during the growth season of the corresponding crops, and generate multiple groups of vegetation index time series; Step 200: Perform pixel extraction on each group of the vegetation index time series based on the pixel-level crop biological growth situation classification model to obtain the temporal features of the pixels, and obtain the classification result according to the temporal features; Step 300: Based on the deep learning algorithm and the classification result, perform deep feature extraction on the spatial and spectral spaces of the high spatio-temporal resolution remote sensing data to obtain the feature-level data; Step 400: Construct an initial neural network model, and train the initial neural network model based on the feature-level data and the classification result to obtain a trained cross-regional crop suitable harvest period monitoring model; Step 500: Input the data to be measured into the cross-regional crop suitable harvest period monitoring model to obtain the monitoring result. The data to be measured is real-time collected remote sensing data. After feature extraction to obtain the feature-level data, it is input into the cross-regional crop suitable harvest period monitoring model, and the suitable harvest period prediction result of the crops in different regions of the remote sensing image is obtained from the current remote sensing data. Preferably, the high spatio-temporal resolution remote sensing data includes the microwave data Sentinel-1 dataset and the multispectral Sentinel-2 dataset. Preferably, performing pixel extraction on each group of the vegetation index time series based on the pixel-level crop biological growth situation classification model to obtain the temporal features of the pixels, and classifying the temporal features to obtain the classification result includes: Construct a pixel-level crop biological growth situation classification model; Use the pixel-level crop biological growth situation classification model to extract the temporal features of each pixel from the vegetation index time series; Perform pixel-level classification on the corresponding crops in the target area according to the temporal features, and perform fusion extraction according to the single-channel, two-channel or multi-channel fusion method to extract crop pixels to form the corresponding classification result. As shown in FIG. 2, the present invention mainly focuses on the monitoring research based on remote sensing time series. Fully considering the temporal cumulative process and dynamic change characteristics of crop growth, it is intended to construct a cross-regional crop monitoring model based on remote sensing time series by combining deep learning time series models such as the Long Short-Term Memory (LSTM) model and the Transformer model with the remote sensing time series data during the crop growth season. The main contents include the following: ① Acquisition of high spatio-temporal resolution remote sensing time series data: For the geographical information of the fields and the distribution of main crops within the study area, high spatio-temporal resolution remote sensing data during the growth season of the corresponding crops are acquired, and multiple groups of vegetation index time series are generated to analyze the growth characteristics of each crop from sowing to maturity, preliminarily evaluate the growth status and development speed of the crops, form a complete set of crop growth data, provide the corresponding data basis for constructing the crop biological growth classification model described below, and also provide data support for model training. ② Pixel-level classification of crop biological growth: Based on deep learning methods such as the Long Short-Term Memory (LSTM) model and the Transformer model, a pixel-level crop biological growth classification model is constructed to extract the temporal features of each pixel from the vegetation index time series, classify the crops in the study area at the pixel level, conduct fusion extraction experiments according to the R, G, B monochromatic channels, nR-G, nR-B, nG-B and other two-color or multi-color channel fusion formulas, extract crop pixels, form the corresponding pixel data, and at the same time, based on the RGB channel pixels that have been obtained, obtain the HSI pixel model data, analyze and compare the two channels, and select the suitable channel pixels. The data of the above two channel pixels both provide data support for subsequent model training. The specific description of the HSI channel is shown in Figure 3. In Figure 3, the top and bottom ends of the cone in the model represent white and black respectively, the vertical axis represents the intensity of light, and the circumference represents the hue. In a color image, there is a strict mathematical relationship between the HSI model and the RGB model. In this embodiment, RGB can be converted to HSI. The following is the conversion method: (A) First, normalize the values of R, G, B in the range of [0, 255], and thus obtain three r, g, b values in the range of [0, 1]: (B) The calculation formulas for the corresponding three components h, s, i obtained in this embodiment are as follows: s = 1 - 3·min(r, g, b), S ∈ [0, 1]; (C) From the above formulas, the value range of h is [0, 2π], the value range of s is [0, 1], and the value range of i is [0, 1]. However, for the sake of understanding and calculation, this is converted to [0°, 360°], [0, 100], [0, 255]: Through the conversion of the above calculation formulas, this embodiment can obtain the calculation method of the HSI color model and apply it to the operation of actual images. Preferably, after performing pixel extraction on each group of the vegetation index time series based on the pixel-level crop biogrowth classification model to obtain the temporal features of the pixels and classifying the temporal features to obtain a classification result, the following steps are further included: Perform image denoising, image enhancement, and image registration on the high spatio-temporal resolution remote sensing data respectively to obtain the preprocessed high spatio-temporal resolution remote sensing data. Preferably, based on the deep learning algorithm and the classification result, perform deep feature extraction on the space and spectral space of the high spatio-temporal resolution remote sensing data to obtain the feature-level data, including: Obtain a training remote sensing data set; Quantify the corresponding information of the high spatio-temporal resolution remote sensing data based on the vegetation information, space information, and spectral space information in the training remote sensing data set to determine the key factors for the interaction between the high spatio-temporal resolution remote sensing data and each piece of information; Construct a feature extraction model for the high spatio-temporal resolution remote sensing data, train the feature extraction model based on the key factors and the training remote sensing data set, use the key factors as the training labels, and iteratively train to obtain the network parameters of the feature extraction model; the feature extraction model can adopt the commonly used deep learning convolutional neural network, and the original feature extraction model is optimized according to the training data during training. Use the trained model to perform feature extraction on the high spatio-temporal resolution remote sensing data to obtain the feature-level data. Furthermore, for the feature quantification of space information and spectral information, there are relatively mature technical solutions in the prior art, which will not be elaborated in this embodiment. In this embodiment, the key factors are used as the training labeled samples to enable the feature extraction model to learn the feature recognition of the vegetation information, space information, and spectral space information in the remote sensing data set. Preferably, the initial neural network model is a UNet convolutional neural network structure. Preferably, the pixel-level crop biogrowth classification model is any one of a long short-term memory model or a Transformer model. Preferably, training the feature extraction model for the high spatio-temporal resolution remote sensing data based on the key factors and the training remote sensing data set includes: Construct a feature estimation multi-dimensional data set according to the key factors; Each element in the feature estimation multi-dimensional data set reflects the association between the vegetation information, space information, and spectral space information and the high spatio-temporal resolution remote sensing data, and the feature extraction model is used to extract the vegetation information, space information, and spectral space information in the high spatio-temporal resolution remote sensing data. The last step in the technical route of this embodiment is specifically: ③ Monitoring of the suitable harvest period for cross-regional crops: Based on the classification results of pixel-level crop biological growth conditions, combined with remote sensing image data, a cross-regional crop suitable harvest period monitoring model is constructed through the training of crop growth data and pixel-level crop biological growth condition classification models based on convolutional neural network structures such as UNet, forming the monitoring form of the cross-regional crop suitable harvest period. The specific UNet convolutional neural network structure is shown in Figure 4. UNet uses a fully convolutional neural network. The left network is a feature extraction network: using conv and pooling; the right network is a feature fusion network: using the feature map generated by upsampling to perform a concatenate operation with the left-side feature map. (The pooling layer will lose image information and reduce image resolution and is permanent, which has some impact on image segmentation tasks and little impact on image classification tasks. Upsampling can make a low-resolution picture containing high-level abstract features become high-resolution while retaining high-level abstract features, and then perform a concatenate operation with the high-resolution picture of the low-level surface features on the left); finally, after two convolutional operations, a feature map is generated, and then two convolutions with a kernel size of 1*1 are used for classification to obtain the final two heatmaps. For example, the first one represents the score of the first category, and the second one represents the score heatmap of the second category, and then it is used as the input of the softmax function to calculate the softmax with a relatively large probability, and then perform loss and backpropagation calculations. The beneficial effects of the present invention are as follows: The present invention can use a large amount of remote sensing images to monitor the suitable harvest period of crops with high precision, improving the automation and intelligence of cross-regional crop suitable harvest period monitoring. The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A cross-regional crop harvest period monitoring method based on deep learning, characterized in that: include: Obtain high temporal and spatial resolution remote sensing data in the target area during the corresponding crop growing season, and generate multiple sets of vegetation index time series; Extracting pixels from each group of vegetation index time series based on a pixel-level crop biological growth classification model to obtain time series features of the pixels, and obtaining classification results based on the time series features; Based on a deep learning algorithm, deep feature extraction is performed on the spatial and spectral spaces of the high temporal and spatial resolution remote sensing data to obtain feature-level data; Constructing an initial neural network model, and training the initial neural network model based on the feature-level data and the classification results to obtain a trained cross-regional crop harvest period monitoring model; Inputting the data to be tested into the cross-regional crop harvest period monitoring model to obtain monitoring results; The deep feature extraction of the space and spectral space of the high spatiotemporal resolution remote sensing data based on the deep learning algorithm to obtain the feature-level data includes: Obtain training remote sensing dataset; Based on the vegetation information, spatial information and spectral spatial information in the training remote sensing data set, feature quantification is performed on the corresponding information of the high temporal and spatial resolution remote sensing data to determine the key factors for the interaction between the high temporal and spatial resolution remote sensing data and each information; Constructing a feature extraction model for the high spatiotemporal resolution remote sensing data, training the feature extraction model based on the key factors and the training remote sensing data set, using the key factors as training labels, and iteratively training to obtain network parameters of the feature extraction model; The trained model is used to extract features from high temporal and spatial resolution remote sensing data to obtain the feature-level data.

2. The cross-regional crop harvest period monitoring method based on deep learning according to claim 1 is characterized in that: The high temporal and spatial resolution remote sensing data include microwave data Sentinel-1 data set and multispectral Sentinel-2 data set.

3. The cross-regional crop harvest period monitoring method based on deep learning according to claim 1 is characterized in that: Based on the pixel-level crop biological growth classification model, pixels are extracted from each group of vegetation index time series to obtain the time series characteristics of the pixels, and the time series characteristics are classified to obtain classification results, including: Construct pixel-level crop biological growth classification model; Extracting the time series features of each pixel from the vegetation index time series using the pixel-level crop biological growth classification model; According to the time series features, the corresponding crops in the target area are classified at the pixel level, and fusion extraction is performed according to a single-color channel, a two-color channel or a multi-color channel fusion method to extract crop pixels and form the corresponding classification result.

4. The cross-regional crop harvest period monitoring method based on deep learning according to claim 1 is characterized in that: After extracting pixels from each group of vegetation index time series based on the pixel-level crop biological growth classification model to obtain the time series features of the pixels, and classifying the time series features to obtain the classification results, the method further includes: Image denoising, image enhancement and image registration are performed on the high temporal and spatial resolution remote sensing data to obtain preprocessed high temporal and spatial resolution remote sensing data.

5. The cross-regional crop harvest period monitoring method based on deep learning according to claim 1 is characterized in that: The initial neural network model is a UNet convolutional neural network structure.

6. The cross-regional crop harvest period monitoring method based on deep learning according to claim 1 is characterized in that: The pixel-level crop biological growth classification model is any one of a long short-term memory model or a Transformer model.

7. The cross-regional crop harvest period monitoring method based on deep learning according to claim 1 is characterized in that: Training a feature extraction model for the high spatiotemporal resolution remote sensing data based on the key factors and the training remote sensing data set includes: constructing a feature estimation multidimensional data set based on the key factors; Each element in the feature estimation multidimensional data set reflects the association between vegetation information, spatial information and spectral spatial information and high temporal and spatial resolution remote sensing data, and the feature extraction model is used to extract vegetation information, spatial information and spectral spatial information from high temporal and spatial resolution remote sensing data.

Images