Climate-adaptive emergency crop monitoring method and device based on deep learning

Abstract

The present invention relates to the field of deep learning technology, and in particular to a climate-adaptive emergency crop monitoring method and device based on deep learning. The method obtains input data of the area to be monitored for emergency crops within a preset time before the current moment, and the input data includes satellite image data and time-series climate data; then the input data is input into a pre-trained monitoring model, and a crop map containing various crop types in the monitoring area is output. This method does not require on-site label collection or model retraining, and reduces the impact of climate factors on phenological characteristics from an interpretable perspective, thereby achieving high-precision crop mapping.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a climate adaptability emergency crop monitoring method and device based on deep learning. Background Art

[0002] In recent years, the intensification of extreme climate events and the frequent occurrence of catastrophic events such as hurricanes, floods and regional conflicts have significantly increased the uncertainty of the global food supply chain. In order to stabilize the food supply chain and maximize the allocation of limited rescue resources, emergency crop mapping is essential, that is, quickly confirming the types of crops grown in a certain area.

[0003] With the increase in available satellite remote sensing data, the combination of remote sensing data and machine learning algorithms has become the mainstream strategy for large-scale crop mapping and yield estimation. Remote sensing technology has the advantages of wide coverage, strong timeliness, and periodic monitoring, and is an important technical means for crop identification and yield estimation. Despite significant progress in this field, emergency mapping of these inaccessible areas affected by war or disasters remains challenging, especially due to difficulties in sample collection. Moreover, when the monitoring model is trained in one region and applied to inaccessible areas with different agro-ecological conditions, the performance of crop mapping will decrease as the phenological differences increase.

[0004] Based on this, the present invention proposes a climate-adaptive emergency crop monitoring method and device based on deep learning to solve the above technical problems. Summary of the invention

[0005] The present invention describes a climate-adaptive emergency crop monitoring method and device based on deep learning, which can effectively improve the accuracy of emergency crop mapping.

[0006] According to a first aspect, the present invention provides a climate-adaptive emergency crop monitoring method based on deep learning, comprising:

[0007] Acquire input data of the area to be subjected to emergency crop monitoring within a preset time before the current moment; wherein the input data includes satellite image data and time series climate data;

[0008] The input data is input into a pre-trained monitoring model, and a crop map of the monitoring area containing various crop types is output; wherein the monitoring model is obtained by training a target neural network using known sample pairs, and the known sample pairs include known input data and known crop maps.

[0009] According to a second aspect, the present invention provides a climate-adaptive emergency crop monitoring device based on deep learning, comprising:

[0010] An acquisition unit is configured to acquire input data of the area to be subjected to emergency crop monitoring within a preset time before the current moment; wherein the input data includes satellite image data and time series climate data;

[0011] The output unit is configured to input the input data into a pre-trained monitoring model, and output a crop map of the monitoring area containing various crop types; wherein the monitoring model is obtained by training a target neural network using known sample pairs, and the known sample pairs include known input data and known crop maps.

[0012] According to a third aspect, the present invention provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method of the first aspect is implemented.

[0013] According to a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method of the first aspect.

[0014] According to the deep learning-based climate-adaptive emergency crop monitoring method and device provided by the present invention, input data of the area to be emergency crop monitored within a preset time before the current moment is obtained, and the input data includes satellite image data and time-series climate data; the input data is then input into a pre-trained monitoring model, and a crop map containing various crop types in the monitored area is output. This method does not require on-site label collection or model retraining, and achieves high-precision crop mapping by reducing the impact of climate factors on phenological characteristics from an explainable perspective. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 A schematic flow chart of a climate-adaptive emergency crop monitoring method based on deep learning according to one embodiment is shown;

[0017] Figure 2 A schematic block diagram of a climate-adaptive emergency crop monitoring device based on deep learning according to one embodiment is shown;

[0018] Figure 3 A flowchart of a deep learning-based climate-adaptive emergency crop monitoring method according to one embodiment is shown. DETAILED DESCRIPTION

[0019] The solution provided by the present invention is described below in conjunction with the accompanying drawings.

[0020] Figure 1 FIG. 1 is a flow chart of a method for climate-adaptive emergency crop monitoring based on deep learning according to an embodiment. It is understood that the method can be executed by any device, equipment, platform, or device cluster with computing and processing capabilities. Figure 1 As shown, the method includes:

[0021] Step 101, obtaining input data of the area to be monitored for emergency crops within a preset time before the current moment; wherein the input data includes satellite image data and time series climate data;

[0022] Step 102: input the input data into a pre-trained monitoring model, and output a crop map of the monitoring area containing various crop types; wherein the monitoring model is obtained by training a target neural network using known sample pairs, and the known sample pairs include known input data and known crop maps.

[0023] In this embodiment, input data of the area to be emergency crop monitored within a preset time before the current moment is obtained, and the input data includes satellite image data and time-series climate data; the input data is then input into a pre-trained monitoring model, and a crop map of each crop type in the monitoring area is output. This method does not require on-site label collection or model retraining, and achieves high-precision crop mapping by reducing the impact of climate factors on phenological characteristics from an explainable perspective.

[0024] It should be noted that a significant challenge in emergency crop mapping at this stage is the difference in crop phenological timing of the same crop under different climatic conditions. For example, satellite image sequences can be used to distinguish different crops, reflecting their unique characteristics from planting to harvesting. However, the phenological characteristics of the same crop vary significantly under different climatic conditions. Therefore, when the model is trained in one region and applied to inaccessible areas with different agricultural ecological conditions, the performance of crop mapping will decrease as the phenological difference increases, with a decrease of up to 97%.

[0025] To address the above challenges, it is critical to extract features that remain constant under a variety of agro-ecological conditions. To this end, several approaches have been proposed, including: feature selection strategies (features in one region must be manually selected to match features in another region), self-supervised learning strategies (the model is pre-trained on a large amount of unlabeled data and fine-tuned with a small amount of labeled data in each target new region. Fine-tuning is required for each region), and domain adaptation strategies (the model must be retrained for each new region to reduce domain shift between cross-regional features). Although the first two strategies reduce the model's reliance on labels, they still require collecting some labels in inaccessible areas to extract similar features between different regions, which is unrealistic in many disaster areas. The third strategy aims to reduce domain shift between regions by retraining the model in each new region, but this approach is time-consuming and cumbersome when conducting large-scale mapping. In addition, performance drops sharply if the proportion of crops in the inaccessible area changes.

[0026] This paper is the first attempt to address the long-standing and unsolved challenge of large-scale emergency crop mapping without retraining models or additional label collection. The proposed method is called Climate Adaptive Crop Mapping (CACM) strategy (automatically extracting invariant features across regions based on regional climate conditions to ensure consistency of extracted features), which is based on the phenological response of crops to agricultural ecological conditions, resulting in significant differences between different regions. CACM uses contrastive learning and cross-attention mechanisms to reduce the impact of climate factors on crop phenological characteristics by designing an interpretable climate response mitigation module to ensure that the features extracted in different regions are consistent. Unlike existing machine learning-based strategies, CACM does not require retraining models or label collection. CACM significantly outperformed the current state-of-the-art methods when conducting large-scale emergency mapping in six disaster-prone countries covering approximately 948,200 square kilometers with diverse agricultural ecological conditions. In the assessment of additional crop losses in the 2021 Henan flood, CACM was able to conduct emergency crop mapping in undeveloped or disaster-affected inaccessible areas, which is critical for optimizing agricultural practices, crop management, and food production monitoring, thereby enhancing food security.

[0027] In one embodiment of the present invention, the time series climate data includes maximum surface temperature, minimum surface temperature, downward surface shortwave radiation, precipitation and soil properties. The soil properties may include soil nutrients, soil organic carbon and soil cation exchange capacity, etc.

[0028] In one embodiment of the present invention, the target neural network includes an embedding module, a climate response mitigation module and a feature fusion module connected in sequence, the embedding module is used to extract feature information of input data, the climate response mitigation module is used to mitigate changes in crops caused by climate, and the feature fusion module includes fusing features output by the climate response mitigation module to identify subtle changes in crop growth.

[0029] In one embodiment of the present invention, the embedding module includes a time embedding component and a climate embedding component. The time embedding component is used to retain the time position information of the satellite image data, and the climate embedding component is used to map multi-time climate data to a higher-dimensional representation space.

[0030] In one embodiment of the present invention, the climate response mitigation module includes:

[0031] Multiple response mitigation components, each of which is used to perform: extracting key vectors and value vectors from satellite image data and extracting query vectors from time series climate data; calculating attention features based on the key vectors and query vectors; performing softmax processing on the attention features and weighting them with the value vector;

[0032] A concatenation component is used to concatenate the features output by each response mitigation component.

[0033] In one embodiment of the present invention, the attention feature is obtained by the following formula:

[0034]

[0035] In the formula, A represents the attention feature, represents the query vector, represents the key vector, T represents the transpose of the matrix, d q express The channel dimension.

[0036] In one embodiment of the present invention, the target neural network is optimized by the following objective function:

[0037]

[0038] In the formula, L represents the objective function, C represents the total number of categories, and y i represents the true label, p i represents the probability that a sample belongs to category i, i∈I≡{1…N} is the index of the sample, N is the number of samples in a batch, λ is the balance parameter, |P(i)| is the cardinality of the set P(i), and z i 、z p 、z arepresents the output of the feature fusion module in different dimensions, τ represents the temperature parameter used to adjust the sharpness of the similarity distribution, · represents the dot product operation, and A(i)≡I\i represents the set of remaining indices after excluding index i.

[0039] like Figure 3 As shown in the figure, CACM uses pixel-by-pixel multi-temporal Sentinel-2 satellite remote sensing observation data (from April to August) and time-series climate data (maximum / minimum air temperature and downward shortwave radiation) from TerraClimate25 as input and generates crop type maps. The strategy includes two climate response mitigation modules (CRMs), a feature extractor, and a classifier. Since crop phenological changes are mainly related to maximum / minimum surface temperature and downward surface shortwave radiation, two CRMs (CRM_T and CRM_R, i.e., two response mitigation components) are designed to mitigate the impact of these factors on phenological characteristics. It should be noted that other factors such as precipitation and nutrients may also affect changes in crop phenology.

[0040] Next, the features extracted from the CRM are fused using a feature fusion module while maintaining a lightweight structure. The classifier is connected to the fusion block for crop mapping with a cross entropy loss. The training process is optimized by a contrastive loss, which is defined as minimizing the distance between crop features of the same category while maximizing the distance between crop features of different categories. In this way, the proposed method is able to separate the effects of temperature and shortwave radiation through a climate response mitigation module, thereby extracting crop features that remain unchanged under different agro-climatic conditions.

[0041] To evaluate the emergency mapping capabilities of CACM, the model was trained using 6,345,093 samples in six states of the United States and applied to hypothetical unlabeled inaccessible areas to map three major crops including corn, soybeans, and rice. To ensure diversity and comprehensiveness of the test, we selected twelve regions across six countries, covering approximately 948,200 square kilometers, to account for different agro-ecological conditions in inaccessible areas. These regions have four main characteristics: (1) high disaster risk, with more than 15% of national natural disasters occurring in these regions; (2) climate diversity covering six –Geiger climate type; (3) differences in crop distribution; (4) significant phenological differences, especially significant changes in crop sowing dates.

[0042] For example, during the week of July 19, 2021, Henan Province, China experienced widespread flooding (referred to as the “2021 Henan Floods”) triggered by record rainfall. As one of the cities hardest hit by the floods, Jiaozuo suffered severe crop damage during this event. Corn is a major crop in the region that had not yet been harvested before the floods, and crop maps are missing, making timely assessment of corn losses both critical and challenging. Using CACM, we generated a 10-meter resolution crop map of Jiaozuo using satellite imagery from 2020 and 2021, and estimated the severity and area of ​​corn damage by comparing the vegetation condition index before and after the floods (July 18-23).

[0043] The CACM-generated damaged corn area in Jiaozuo is 33,922.33 hectares, of which 1,357.42 hectares are severely damaged, highlighting the widespread impact of the floods on corn cultivation in Jiaozuo. The southern and southeastern regions of Jiaozuo were the most severely affected. Specifically, Wuzhi had the most severe corn damage, with a total damaged area of ​​11,415.79 hectares. Other regions such as Xiuwu, Qinyang, Mengzhou, and Boai also recorded large amounts of damage, with each region having an affected area of ​​more than 3,000 hectares.

[0044] To verify the reliability of the estimation results, we estimated the maize acreage based on the map generated by CACM and compared it with the official statistical records. The maize acreage generated by CACM is highly consistent with the statistical records (R 2 =0.82, root mean square error RMSE = 6 hectares). Official statistical records also contain information on the area of ​​cultivated land lost due to floods. We use this information to further verify the cultivated land area loss estimated by CACM. The total cultivated land loss caused by the 2021 Henan floods in Jiaozuo was 119,251 hectares, which is 0.8% lower than the official record of 118,340 hectares.

[0045] To further estimate the corn losses caused by the floods, we used the average corn yield of 6.135 tons per hectare in Henan Province in 2020 as the expected yield before the floods. Assuming that the extreme damage caused by the floods would lead to a complete loss of the crop (if no recovery measures were taken), the potential corn production reduction in Jiaozuo due to extreme damage was 8,327.78 tons, which was calculated by multiplying the extremely damaged area by the yield.

[0046] That is, to address the urgent crop mapping needs under the intensification of extreme climate events and catastrophic events, we propose a first-of-its-kind climate-adaptive deep learning method (CACM) for rapid crop mapping. CACM mitigates the impact of climate factors on phenological features without retraining models or collecting labels. Through contrastive learning and cross-attention mechanisms, the CACM method integrates temperature and radiation knowledge into the deep learning network to make the extracted features consistent across regions with different climate conditions. CACM was trained in the United States and tested in 12 regions in six countries covering approximately 948,000 square kilometers and six different climate types. CACM achieved an average F1 score of 86.44%, which is much higher than three state-of-the-art methods, including RF (71.89%), DCM (76.13%), and PAN (76.65%). CACM performed particularly well in the recognition of maize and rice (improved by at least 17.90% and 20.95%, respectively), crops whose phenology is significantly affected by climate. The training time of CACM was reduced by 30 times compared to the second best performing method PAN. The application of CACM to mapping crops affected by the 2021 Henan floods further demonstrated its accuracy and timeliness in large-scale emergency crop mapping.

[0047] Crop type maps generated by CACM can serve as basic data for enhancing food security. For example, it is able to quickly assess damage after post-disaster disasters such as floods and wars, helping government decision makers understand the area and spatial distribution of damaged crops. This helps to efficiently allocate relief resources such as manpower, materials and funds, thereby ensuring effective support for agricultural recovery and disaster relief strategies. However, most methods require satellite imagery throughout the growing season. If the disaster occurs before harvest, related technologies usually use satellite images from the previous year for mapping, which may introduce bias. In contrast, CACM only uses satellite images taken from April to August, which is able to timely and accurately assess crop damage before the disaster occurs.

[0048] However, the method provided by the embodiment of the present invention can be further improved by adding more factors that affect crop phenology changes. However, data from many regions, such as crop varieties, may not be available. For example, if only temperature or solar radiation data were used in this study, the performance would decrease by 7.03% or 4.12% respectively compared to using both factors at the same time. Therefore, these two factors that affect crop phenology changes are the most important.

[0049] The monitoring model provided by the present invention is implemented on the NVIDIA GeForce RTX 3090 GPU using PyTorch. The stochastic gradient descent (SGD) optimizer is used with an initial learning rate of 5×10^-3, a momentum of 0.9, and a weight decay of 1×10 -4We use a multi-step learning rate schedule with γ being 0.1 and the learning rate adjustment points being the 30th, 60th, and 90th epochs. The batch size and learning epochs for all experiments are set to 4096 and 100, respectively, and λ is set to 0.1.

[0050] This study evaluates the model from two perspectives: pixel scale and regional scale. For pixel scale evaluation, F1 score (F1) and mean F1 score (mF1) are used as metrics. F1 represents the harmonic mean of precision and recall (Formula (gs1)), which can be used to comprehensively evaluate the performance of the model. mF1 is the average F1 of all crop categories (Formula (gs 2)). Specifically, for each category i, the category is considered as the positive category, while other categories are considered as negative categories. TP i Indicates the number of true positive samples, FN i Indicates the number of false negative samples, FP i Represents the number of false positive samples. C represents the total number of categories.

[0051]

[0052] At the regional scale, the total area of ​​target crops in each sub-region is calculated based on the mapping results and compared with official statistics. In Jilin Province, the United States, and Hungary, the sub-regions refer to cities, agricultural statistical areas, and states, respectively. The indicators R2 (coefficient of determination) and RMSE (root mean square error) are introduced to evaluate spatial accuracy. R2 indicates the goodness of fit between the predicted crop area and the actual crop area, while RMSE is used to quantify the difference between the predicted value and the actual value.

[0053]

[0054] Among them, a i represents the actual crop area of ​​each sub-region, represents the predicted crop area, a represents the average crop area of ​​the sub-region, and m represents the number of sub-regions.

[0055] In summary, the embodiments of the present invention propose a climate-adaptive crop mapping (CACM) strategy to quickly use satellite imagery for crop monitoring. CACM does not require on-site label collection or model retraining, and achieves high-precision crop mapping by mitigating the impact of climate factors on phenological characteristics from an interpretable perspective. CACM was applied in six disaster-prone countries, spanning 948,000 square kilometers of various climatic conditions. The results show that CACM has an average F1 score of 86.44%, which is at least 17.90% and 20.95% higher than the state-of-the-art technology in distinguishing between corn and rice, which are highly affected by climate. CACM is also able to quickly and accurately estimate corn losses in hard-to-reach areas caused by the 2021 Henan floods. Our research promotes the development of emergency crop mapping and helps enhance food security hindered by disasters and regional conflicts.

[0056] The above describes specific embodiments of the present invention. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0057] According to an embodiment of another aspect, the present invention provides a climate-adaptive emergency crop monitoring device based on deep learning. Figure 2 A schematic block diagram of a climate-adaptive emergency crop monitoring device based on deep learning according to an embodiment is shown. It can be understood that the device can be implemented by any device, equipment, platform and device cluster with computing and processing capabilities. Figure 2 As shown, the device includes: an acquisition unit 201 and an output unit 202. The main functions of each component unit are as follows:

[0058] The acquisition unit 201 is configured to acquire input data of the area to be subjected to emergency crop monitoring within a preset time before the current moment; wherein the input data includes satellite image data and time series climate data;

[0059] The output unit 202 is configured to input the input data into a pre-trained monitoring model, and output a crop map of the monitoring area containing various crop types; wherein the monitoring model is obtained by training a target neural network using known sample pairs, and the known sample pairs include known input data and known crop maps.

[0060] As a preferred implementation, the time series climate data includes maximum surface temperature, minimum surface temperature, downward surface shortwave radiation, precipitation and soil properties.

[0061] As a preferred embodiment, the target neural network includes an embedding module, a climate response mitigation module and a feature fusion module which are connected in sequence, the embedding module is used to extract feature information of the input data, the climate response mitigation module is used to mitigate the changes in crops caused by climate, and the feature fusion module includes fusing the features output by the climate response mitigation module to identify subtle changes in crop growth.

[0062] As a preferred implementation, the embedding module includes a time embedding component and a climate embedding component, wherein the time embedding component is used to retain the temporal position information of the satellite image data, and the climate embedding component is used to map multi-time climate data to a higher-dimensional representation space.

[0063] As a preferred implementation, the climate response mitigation module includes:

[0064] A plurality of response mitigation components, each of which is used to perform: extracting a key vector and a value vector from the satellite image data and extracting a query vector from the time series climate data; calculating an attention feature based on the key vector and the query vector; performing softmax processing on the attention feature and performing weighted processing with the value vector;

[0065] A concatenation component is used to concatenate the features output by each of the response mitigation components.

[0066] As a preferred implementation, the attention feature is obtained by the following formula:

[0067]

[0068] In the formula, A represents the attention feature, represents the query vector, represents the key vector, T represents the transpose of the matrix, d q express The channel dimension.

[0069] As a preferred implementation, the target neural network is optimized by the following objective function:

[0070]

[0071] In the formula, L represents the objective function, C represents the total number of categories, and y i represents the true label, p irepresents the probability that a sample belongs to category i, i∈I≡{1…N} is the index of the sample, N is the number of samples in a batch, λ is the balance parameter, |P(i)| is the cardinality of the set P(i), and z i 、z p 、z a represents the output of the feature fusion module in different dimensions, τ represents the temperature parameter used to adjust the sharpness of the similarity distribution, · represents the dot product operation, and A(i)≡I\i represents the set of remaining indexes after excluding index i.

[0072] According to another embodiment, there is also provided a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute a combination of Figure 1 The method described.

[0073] According to another embodiment of the present invention, an electronic device is provided, including a memory and a processor, wherein an executable code is stored in the memory, and when the processor executes the executable code, the Figure 1 method.

[0074] The various embodiments of the present invention are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0075] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the present invention can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0076] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A climate-adaptive emergency crop monitoring method based on deep learning, characterized in that: include: Acquire input data of the area to be subjected to emergency crop monitoring within a preset time before the current moment; wherein the input data includes satellite image data and time series climate data; Inputting the input data into a pre-trained monitoring model, and outputting a crop map of the monitoring area containing various crop types; wherein the monitoring model is obtained by training a target neural network using known sample pairs, and the known sample pairs include known input data and known crop maps; The target neural network includes an embedding module, a climate response mitigation module and a feature fusion module connected in sequence, wherein the embedding module is used to extract feature information of the input data, the climate response mitigation module is used to mitigate changes in crops caused by climate, and the feature fusion module includes fusing features output by the climate response mitigation module to identify subtle changes in crop growth; The embedding module includes a time embedding component and a climate embedding component, wherein the time embedding component is used to retain the time position information of the satellite image data, and the climate embedding component is used to map the multi-time climate data to a higher-dimensional representation space; The climate response mitigation module includes: A plurality of response mitigation components, each of which is used to perform: extracting a key vector and a value vector from the satellite image data and extracting a query vector from the time series climate data; calculating an attention feature based on the key vector and the query vector; performing softmax processing on the attention feature and performing weighted processing with the value vector; A concatenation component is used to concatenate the features output by each of the response mitigation components.

2. The method according to claim 1, characterized in that The time series climate data include maximum surface temperature, minimum surface temperature, downward surface shortwave radiation, precipitation and soil properties.

3. The method according to claim 2, characterized in that The attention feature is obtained by the following formula: In the formula, A represents the attention feature, represents the query vector, represents the key vector, T represents the transpose of the matrix, d q express The channel dimension.

4. The method according to claim 3, characterized in that The target neural network is optimized by the following objective function: In the formula, L represents the objective function, C represents the total number of categories, and y i represents the true label, p i represents the probability that a sample belongs to category i, i∈I≡{1…N} is the index of the sample, N is the number of samples in a batch, λ is the balance parameter, |P(i)| is the cardinality of the set P(i), and z i 、z p 、z a represents the output of the feature fusion module in different dimensions, τ represents the temperature parameter used to adjust the sharpness of the similarity distribution, · represents the dot product operation, and A(i)≡I\i represents the set of remaining indexes after excluding index i.

5. A climate-adaptive emergency crop monitoring device based on deep learning, characterized in that: include: An acquisition unit is configured to acquire input data of the area to be subjected to emergency crop monitoring within a preset time before the current moment; wherein the input data includes satellite image data and time series climate data; an output unit configured to input the input data into a pre-trained monitoring model, and output a crop map of the monitoring area containing various crop types; wherein the monitoring model is obtained by training a target neural network using known sample pairs, and the known sample pairs include known input data and known crop maps; The target neural network includes an embedding module, a climate response mitigation module and a feature fusion module connected in sequence, wherein the embedding module is used to extract feature information of the input data, the climate response mitigation module is used to mitigate changes in crops caused by climate, and the feature fusion module includes fusing features output by the climate response mitigation module to identify subtle changes in crop growth; The embedding module includes a time embedding component and a climate embedding component, wherein the time embedding component is used to retain the time position information of the satellite image data, and the climate embedding component is used to map the multi-time climate data to a higher-dimensional representation space; The climate response mitigation module includes: A plurality of response mitigation components, each of which is used to perform: extracting a key vector and a value vector from the satellite image data and extracting a query vector from the time series climate data; calculating an attention feature based on the key vector and the query vector; performing softmax processing on the attention feature and performing weighted processing with the value vector; A concatenation component is used to concatenate the features output by each of the response mitigation components.

6. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method according to any one of claims 1 to 4 is implemented.

7. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 4.

Images