CN-121999369-A - Paddy field remote sensing image interpretation method and system based on deep learning and physical weather assistance

CN121999369ACN 121999369 ACN121999369 ACN 121999369ACN-121999369-A

Abstract

The invention provides a paddy field remote sensing image interpretation method and system based on deep learning and physical condition assistance, comprising the following steps of S1, constructing a double-branch encoder, respectively extracting space-spectrum characteristics and physical condition time sequence characteristics, establishing a multi-scale characteristic mapping relation, S2, establishing a physical condition phase alignment module, registering key physical condition characteristics with image time sequences to generate a physical condition phase weight matrix, S3, establishing a cross-modal attention mechanism, calculating the mutual attention score of the space-spectrum characteristics and the physical condition characteristics, realizing self-adaptive characteristic fusion, S4, constructing a physical condition constraint loss function, combining physical condition period priori knowledge constraint model prediction results, inhibiting non-paddy field misjudgment, S5, establishing a self-adaptive threshold optimization module, dynamically adjusting a segmentation threshold according to the physical condition confidence, improving boundary region extraction precision, S6, establishing a domain self-adaptive migration module, and improving the generalization performance of the model on cross-year and cross-region data through characteristic distribution alignment.

Inventors

LIU XIAOYU
ZHENG YUNYUN
HU YONG
Zhou Luhe
LI XIAOJUN
AO YING
YANG KAI
ZHAN SHIJUN
LIAO XU

Assignees

重庆市规划和自然资源调查监测院

Dates

Publication Date: 20260508
Application Date: 20260130

Claims (8)

1. The paddy field remote sensing image interpretation method based on deep learning and physical assistance is characterized by comprising the following steps of: s1, constructing a double-branch encoder, respectively extracting space-spectrum characteristics and climatic time sequence characteristics, and establishing a multi-scale characteristic mapping relation; S2, establishing a physical phase alignment module, registering key physical phase characteristics with an image time sequence, and generating a physical phase weight matrix; s3, a cross-modal attention mechanism is established, the mutual attention score of the space-spectrum characteristic and the physical characteristic is calculated, and self-adaptive characteristic fusion is realized; S4, constructing a conjugate weather constraint loss function, and restraining a model prediction result by combining conjugate weather period priori knowledge to inhibit misjudgment of non-rice features; s5, a self-adaptive threshold optimization module is established, a segmentation threshold is dynamically adjusted according to the confidence level of the object, and the extraction precision of the boundary region is improved; and S6, establishing a domain self-adaptive migration component, and improving the generalization performance of the model on the annual and regional data by means of characteristic distribution alignment.
2. The method for interpreting a remote sensing image of a paddy field based on deep learning and climate assistance as claimed in claim 1, wherein S1 comprises: because the single-branch network is difficult to capture the space details and long-time sequence climatic rules of the high-resolution image at the same time, the step adopts a decoupling double-branch architecture to realize modal separation coding, and the multi-scale space feature map is extracted through a residual error encoder Extracting the candidate feature vector by adopting time sequence convolution network Wherein 128 Is the time sequence coding dimension, the branch captures key climatic nodes such as water index sudden increase in the rice transplanting period, vegetation index peak in the heading period and the like, four-level multi-scale space features are generated through pyramid pooling And realizing multi-level characterization from ridge textures to planting patterns.
3. The method for interpreting a remote sensing image of a paddy field based on deep learning and climate assistance as claimed in claim 1, wherein S2 comprises: Construction of a phase weight matrix Where σ is the Sigmoid activation function, For the matrix of phase-mapped weights, For key period embedded vectors based on xx city rice crop weather table codes, Is an offset item, and aims at the image acquisition time Phase registration weights are calculated by gaussian kernel function Where k=1 corresponds to the transplanting period , Is the basic weight coefficient of the kth weathered period, meets the following conditions And the transplanting period For the term weight coefficient of the model, the term weight coefficient, The day is a time tolerance parameter, and continuous mapping of discrete object weather labels to actual image acquisition time is achieved.
4. The method for interpreting a remote sensing image of a paddy field based on deep learning and climate assistance as claimed in claim 1, wherein said S3 comprises: after the aligned climatic features are obtained, a dynamic association between the spatial position and the climatic state is required to be established, and the spatial features are used Flattened along the spatial dimension into Wherein hw= Representing the total number of pixels, obtaining a space query matrix through projective transformation Wherein In order to project the matrix of the light, Scaling factor for key value dimension, projection of the weathered features to obtain key value matrix Sum value matrix Wherein Computing a cross-modal attention map Multiplying the obtained product with a numerical matrix to obtain Generating a cross-modal attention feature after remodelling into a space dimension, and inquiring a vector when the space feature detects the water body reflection characteristic, namely near infrared band low reflection The high attention weight is obtained by calculating the physical key value of the transplanting period, while the lotus pool area has similar water body characteristics, but the attention weight is restrained and the characteristics are fused due to the lack of the special follow-up vegetation growth physical mode of the transplanting period Wherein beta is a fusion scaling factor of the climate, For the channel projection matrix, And recalibrating the operation for each channel.
5. The method for interpreting a remote sensing image of a paddy field based on deep learning and climate assistance as claimed in claim 1, wherein S4 comprises: establishing a priori constraint of the weathers, defining constraint loss of the weathers Where N is the total number of pixels, CE is the cross entropy loss, The ith pixel is manually labeled with a label, For model predictive probability, lambda is the weathered constraint weight, in a specific implementation, Representing the weatherable sensitivity weight of the jth phase, e.g. transplanting period Heading stage , For the KL divergence, the average value of the power supply is calculated, For model predicted probability distribution of the j-th rice, Is a priori distribution constructed based on rice weather in Chongqing.
6. The method for interpreting a remote sensing image of a paddy field based on deep learning and climate assistance as claimed in claim 1, wherein said S5 comprises: Computing adaptive segmentation threshold Wherein As a matter-level confidence level adjustment factor, As the current phase object candidate feature variance, As a result of the normalization factor, For the spatial average attention confidence level, For the boundary region of the seven-pond Zhenkan, introducing space continuity constraint Wherein ∇ is a Sobel gradient operator, Is a boundary sensitivity parameter.
7. The method for interpreting a remote sensing image of a paddy field based on deep learning and climate assistance as claimed in claim 1, wherein S6 comprises: Construction domain adaptive loss Wherein phi is a domain arbiter, a three-layer fully connected network, As a spatial feature of the source domain, As a feature of the target domain, =0.5 Is the time-series gradient consistency weight, And calculating the NDVI change rate between adjacent months as a climatic time sequence gradient operator.
8. A computer system, comprising: A processor; A memory for storing processor-executable instructions; The processor is configured to implement the paddy field remote sensing image interpretation method based on deep learning and climate assistance according to one of claims 1 to 9 when executing the executable instructions.

Description

Paddy field remote sensing image interpretation method and system based on deep learning and physical weather assistance Technical Field The invention relates to the field of circuit control, in particular to a paddy field remote sensing image interpretation method and system based on deep learning and physical assistance. Background With the rapid development of remote sensing technology, intelligent interpretation methods based on deep learning have become an important means for land utilization classification and crop monitoring. At present, various deep learning models are widely applied to remote sensing image interpretation tasks, such as Convolutional Neural Networks (CNNs) are excellent in spatial feature extraction, a transducer model captures long-range dependency relations through a self-attention mechanism, and a U-Net, deep Lab and the like partition network shows remarkable advantages in high-resolution image classification. However, paddy fields often have difficulty achieving high accuracy interpretation due to their unique spectral features (e.g., quaternary water changes and vegetation cover dynamics) and high confusion with other features (e.g., wetlands, aquatic vegetation), relying solely on deep learning methods of spatio-spectral features. The climatic information is used as a time sequence representation of crop growth, so that the defect of space characteristics can be effectively overcome. Research shows that the critical climatic period (such as water body characteristics in transplanting period and vegetation index change in tillering period) of rice can obviously improve the distinguishing capability of the classification model. For example, zeng et al (2020) realized accurate identification of paddy fields by fusing an NDVI time sequence curve with a random forest model, and Wang et al (2022) further combined with Sentinel-1/2 multisource data and climatic features to increase classification accuracy to more than 90%. However, the existing research focuses on direct embedding of the climatic features, and fails to fully mine a multi-level cooperative mechanism of the climatic features and a deep learning model, and has limited generalization capability especially in a complex planting environment. In order to improve the remote sensing interpretation precision of the paddy field, a multi-temporal remote sensing image paddy field intelligent interpretation framework based on the combination of deep learning and physical weather assistance is provided, and the spatial-spectral characteristics and physical weather time sequence rules of images are synchronously extracted by designing a double-branch network structure, and a attention mechanism is introduced to dynamically fuse multi-mode characteristics so as to solve the problem of confusion between the paddy field and similar objects. The research not only provides a new thought for remote sensing monitoring of crops, but also provides technical support for accurate agricultural management driven by multi-time phase data. There is a need for a person skilled in the art to solve the corresponding technical problems. Disclosure of Invention The invention aims at least solving the technical problems in the prior art, and particularly creatively provides an electromagnetic pulse welding effect regulating and controlling method based on current measurement and waveform characteristics. In order to achieve the above object of the present invention, the present invention provides a paddy field remote sensing image interpretation method based on deep learning and climate assistance, comprising the following steps: s1, constructing a double-branch encoder, respectively extracting space-spectrum characteristics and climatic time sequence characteristics, and establishing a multi-scale characteristic mapping relation; S2, establishing a physical phase alignment module, registering key physical phase characteristics with an image time sequence, and generating a physical phase weight matrix; s3, a cross-modal attention mechanism is established, the mutual attention score of the space-spectrum characteristic and the physical characteristic is calculated, and self-adaptive characteristic fusion is realized; S4, constructing a conjugate weather constraint loss function, and restraining a model prediction result by combining conjugate weather period priori knowledge to inhibit misjudgment of non-rice features; s5, a self-adaptive threshold optimization module is established, a segmentation threshold is dynamically adjusted according to the confidence level of the object, and the extraction precision of the boundary region is improved; and S6, establishing a domain self-adaptive migration component, and improving the generalization performance of the model on the annual and regional data by means of characteristic distribution alignment. Preferably, in the above technical solution, the S1 includes: because the single-branch network is difficult t