Method for reconstructing flood disaster image of cloud and fog blocked area based on precipitation prior and sar constraint

Abstract

The application discloses a cloud and fog sheltered area flood disaster image reconstruction method based on precipitation prior and SAR constraint, and relates to the technical field of remote sensing image processing. Firstly, permanent water body information is extracted by using pre-disaster cloud-free optical images, and daily flood state prior is constructed by combining multi-temporal SAR images, precipitation data and terrain hydrological information. Then, according to the cloud sheltered area, flood state category and background stability, a partitioned differentiated optical reconstruction prior is constructed. Finally, a conditional diffusion model is adopted to reconstruct the disaster image of the cloud and fog sheltered area under the guidance of SAR flood space constraint, precipitation driving weight and flood state prior, while the real observation information of the non-cloud visible area is reserved. The application can reduce the flood information missing problem caused by cloud sheltering of traditional optical images, improve the spectral texture recovery capability of the newly added expansion area, shallow stagnant water area and water recession transition area under the cloud, and improve the continuity and reliability of flood disaster remote sensing monitoring.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints. Background Technology

[0002] Floods are typically accompanied by continuous rainfall and widespread cloud cover. Post-disaster optical remote sensing images often suffer from severe cloud obscuration, making it difficult to obtain timely and complete information about the true surface conditions of the affected areas. For emergency response, the extent of flooding, land-water boundaries, blocked roads, and flooded farmland all need to be identified from remote sensing images as quickly as possible. Cloud obscuration in optical images directly reduces the timeliness and reliability of disaster assessment.

[0003] Synthetic Aperture Radar (SAR) possesses all-weather, all-day observation capabilities, enabling it to acquire surface backscattering information even under cloudy and rainy conditions. Open water bodies, floodplains, and some low-roughness submerged areas typically exhibit low backscattering in SAR imagery, making SAR imagery useful for spatial extent estimation of post-disaster floods. However, SAR imagery differs from optical imagery in its imaging mechanism, exhibiting issues such as speckle noise, geometric distortion, building overlap, mountain shadows, and difficulties in visual interpretation. Therefore, it is less intuitive than optical imagery when directly used for post-disaster analysis.

[0004] Existing methods for removing clouds from optical remote sensing images mainly include multi-temporal image replacement, traditional image inpainting, image generation based on generative adversarial networks, SAR and optical data fusion, and diffusion model conditional generation. Multi-temporal replacement methods typically rely on pre-disaster or historical cloud-free optical images and are suitable for restoring stable ground features. However, in flood disaster scenarios, the actual state of the area under clouds may have changed from farmland, bare land, roads, or wetlands to floodwaters. Directly using pre-disaster image replacement can easily lead to incorrect restoration of the land structure from the pre-disaster state.

[0005] Diffusion models possess strong capabilities in conditional generation and image detail restoration, enabling the generation of high-quality images through progressive denoising. However, when diffusion models are directly applied to the free generation of remotely sensed cloud-occluded areas, in the absence of spatial constraints on floods, constraints on flood processes, and prior knowledge of real post-disaster water bodies, the models tend to generate content that does not conform to the post-disaster flood state based on the distribution of common land features in the training set. For example, they might generate farmland textures in actual flood areas or expand water body textures in non-flood areas.

[0006] Furthermore, existing methods mostly aim at reconstructing single images, and rarely consider the driving role of precipitation events in flood expansion, retention, and recession processes. When the optical observation interval is long and clear optical images are lacking on intermediate dates, it is difficult to express the daily changes of flood processes simply based on image replacement or single-phase SAR constraints. Summary of the Invention

[0007] To address the above technical problems, this invention provides a method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints, comprising the following steps: Step M1: Acquire pre-disaster cloudless optical images, cloudy optical images during the disaster, post-disaster multi-temporal synthetic aperture radar (SAR) images, daily precipitation data, and topographic and hydrological auxiliary data of the target area, and perform registration, cropping, resampling, and sequence statistical preprocessing. Step M2: Extract the permanent water body reference area based on the cloudless optical image before the disaster, and identify flood water bodies in the multi-temporal SAR image after the disaster to obtain SAR flood water body masks corresponding to different observation dates; use the SAR flood water body mask and its land-water boundary as flood spatial anchor points, and identify perennial water accumulation areas, newly expanded flood areas, and receding water transition areas based on the temporal comparison of SAR flood water body masks on different observation dates, and extract the corresponding flood spatial boundary information as spatial constraints for the subsequent daily flood state prior construction; Step M3: Construct precipitation event driving features and daily flood state priors based on the daily precipitation process and the multi-temporal SAR flood change relationship. The daily flood state priors include precipitation priors and SAR flood spatial constraints. Divide the target area into different state zones, including stable land surface zone, perennial water accumulation zone, flood expansion zone, shallow water retention zone, receding water transition zone, and cloud cover uncertainty zone, and determine the spatial distribution of each state zone. Step M4: Extract cloud and cloud shadow occlusion areas from the cloud-covered optical image during the disaster to obtain the disaster-time cloud occlusion mask; spatially intersect the disaster-time cloud occlusion mask with the state partitions in the daily flood state prior to obtain the joint reconstruction mask; the joint reconstruction mask is used to define the flood-oriented reconstruction area of ​​the conditional diffusion model, so that the diffusion reconstruction preferentially acts on the pixel areas that are obscured by clouds and fog and are identified as perennial water accumulation areas, newly expanded flood areas, shallow water retention areas, or receding water transition areas under the precipitation prior and SAR flood spatial constraints; and determine the disaster-time spatial partitions based on the flood state area, stable background constraints, and cloud occlusion mask, including non-cloud visible areas, flood state areas under clouds, stable background areas under clouds, and uncertain state areas under clouds; Step M5: Based on the disaster spatial zoning results determined in Step M4, construct a zoning-differentiated optical prior. Among them, the original observation of the disaster optical image is retained in the non-cloud visible area, the flood state area under the cloud adopts the state flood optical prior, the stable background area under the cloud adopts the stable background prior constructed by the pre-disaster cloudless optical image, and the uncertain state area under the cloud adopts the weak constraint prior. The weak constraint prior refers to the constraint only using the pre-disaster optical image, the texture information of the adjacent area and the generation capability of the diffusion model itself, without applying the flood category constraint. Its reconstruction result has a lower credibility than the determined state zoning. Step M6: Input the cloud-covered optical image during the disaster, the cloud occlusion mask during the disaster, the joint reconstruction mask, the regional differentiated optical prior, the SAR flood spatial constraint characteristics, the precipitation event-driven characteristics, and the adaptive diffusion intensity map into the conditional diffusion model to generate a cloudless flood optical reconstruction image during the disaster. Step M7: Based on the cloudless flood optical reconstruction image and the daily flood status prior, extract the flood inundation range, newly expanded flood zone, receding drawdown transition zone, land-water boundary, reconstruction uncertainty map, and post-disaster analysis results.

[0008] The beneficial effects of this invention are: (1) In this invention, SAR images are upgraded from single-phase flood auxiliary information to dual-phase or multi-phase flood spatial anchors. By comparing the SAR water body range on different dates, stable water bodies, expansion zones and receding zones are constructed to provide spatial constraints for flood state inference during disasters. Based on the daily flood state prior, the cloud-covered area is partitioned. For the flood state area under the cloud, a state-specific flood optical prior is constructed. For the non-flood stable area under the cloud, a pre-disaster stable background optical prior is constructed to avoid incorrectly filling the texture of pre-disaster farmland, bare land, roads or residential areas into the real flood change area. At the same time, the pre-disaster cloudless optical images are fully utilized to restore the stable background under the cloud. (2) In this invention, a joint reconstruction mask is constructed by using a disaster-time cloud-shading mask and a flood state area. Flood-guided diffusion sampling is performed only on cloud-shading areas that are subject to flood state changes. This can accurately limit the flood reconstruction range and reduce the risk of the generated model erroneously modifying non-flood areas, stable ground features, and areas visible outside the clouds. (3) In this invention, non-flood stable areas are separated from the joint reconstruction mask by using a cloud-based stable background mask, and pre-disaster cloudless optical images are used as stable background priors after tone normalization, so that flood change areas and stable land areas adopt different reconstruction strategies, thereby improving the disaster semantic consistency and background spatial continuity of the post-disaster reconstruction images. (4) In this invention, a multi-scale adaptive diffusion intensity map is generated by using the flood boundary distance, cloud area confidence, SAR confidence, precipitation intensity, effective optical observation interval and scale ratio, so that the cloud core area is strongly repaired, the cloud edge is weakly repaired, and the joint reconstruction mask remains unchanged, thereby effectively avoiding boundary abrupt changes, repeated textures, ring artifacts and excessive smoothing problems. (5) In this invention, the uncertainty map of daily reconstruction and post-disaster analysis products are further output, so that the cloudless optical reconstruction results can be used not only for visual interpretation, but also for flood change response, disaster range assessment and emergency decision support. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0010] This embodiment provides a method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints, such as... Figure 1 As shown, it includes the following steps: Step M1: Acquire pre-disaster cloudless optical images, cloudy optical images during the disaster, post-disaster multi-temporal SAR images, daily precipitation data, and topographic and hydrological auxiliary data of the target area, and perform registration, cropping, resampling, and sequence statistical preprocessing.

[0011] Pre-disaster cloudless optical images are used to extract permanent water bodies and stable backgrounds, while cloudy optical images during the disaster are images to be reconstructed. Post-disaster multi-temporal SAR images include at least two post-disaster SAR observation phases to provide spatial constraints on floods under cloud and rain conditions. Daily precipitation data are used to characterize the temporal driving features of flood expansion, retention, and recession processes.

[0012] Optical images undergo band selection, radiometric normalization, out-of-cloud visibility enhancement, and study area cropping. SAR images undergo orbit correction, radiometric calibration, topographic correction, speckle filtering, incident angle normalization, and VV or VH polarization processing. Daily precipitation data undergo spatial cropping, outlier removal, time series completion, regional averaging, or watershed zoning statistics. Daily precipitation, previous day's precipitation, three-day cumulative precipitation, five-day cumulative precipitation, and precipitation peak information are also statistically analyzed to ensure that multi-source data are in the same spatial coordinate system, the same spatial resolution, and the same pixel grid.

[0013] Step M2: Extract the permanent water body reference area based on the pre-disaster cloudless optical image, and identify flood water bodies in the post-disaster multi-temporal SAR image to obtain SAR flood water body masks corresponding to different observation dates; use the SAR flood water body mask and its land-water boundary as flood spatial anchor points, and identify perennial water accumulation areas, newly expanded flood areas, and receding drawdown transition areas based on the temporal comparison of SAR flood water body masks on different observation dates, and extract the corresponding flood spatial boundary information as spatial constraints for subsequent daily flood state prior construction.

[0014] The improved normalized water index (MNDWI) was calculated using the green band and shortwave infrared band of pre-disaster cloudless optical imagery. A permanent water mask was then obtained by combining threshold segmentation and morphological processing. Its expression is as follows: ; Wherein, B03 represents the green band reflectivity, and B11 represents the shortwave infrared band reflectivity.

[0015] When extracting flood water bodies from multi-temporal SAR images, based on the preprocessing results in step M1, dB conversion is performed on the SAR images of each temporal phase, and a backscatter intensity feature map is constructed according to VV and VH polarization information to obtain SAR feature data for flood water body identification. Since open water bodies and flood-inundated areas usually exhibit low backscattering characteristics, this embodiment adopts a low backscattering water body extraction method based on Otsu adaptive threshold segmentation. Threshold segmentation is performed on the backscattering characteristics of VV polarization, VH polarization, or VV and VH dual polarization to obtain candidate areas for SAR water bodies in each temporal phase.

[0016] For the t-th SAR observation phase, its low backscattering characteristic map is denoted as follows: The Otsu adaptive thresholding method was used to determine the water body segmentation threshold for this time phase. For cell (x,y), when Furthermore, this pixel is located within a hydrologically constrained area. If the water body is within the specified range, it is classified as a SAR flood water body pixel; otherwise, it is classified as a non-SAR flood water body pixel. The expression is as follows: ; in, This represents the SAR flood water body determination result at pixel (x,y) in the t-th SAR observation phase. A value of 1 indicates a SAR flood water body pixel, and a value of 0 indicates a non-SAR flood water body pixel. The low backscattering characteristic value at pixel (x,y) represents the t-th SAR observation phase, which can be obtained from the backscattering characteristics of VV polarization, VH polarization, or VV and VH dual polarization. This represents the water body segmentation threshold for the t-th SAR observation phase, automatically determined by the Otsu method. The hydrological spatial constraint area is a constraint mask composed of spatial areas with the possibility of flooding, such as the buffer zone of permanent water bodies, the adjacent area of ​​river networks, low-lying terrain areas, and the range of historical water systems.

[0017] For the t-th SAR observation phase, let its backscattering characteristic map be as follows: The Otsu method was used to automatically determine the water segmentation threshold for the corresponding time phase. When pixel Less than At that time, the pixel was identified as a candidate pixel for a low backscattering water body. Subsequently, the hydrological spatial constraint area was formed by the low backscattering water body candidate area and the buffer zone of the permanent water body, the river network adjacent area, the low-lying terrain area, or the historical water system range. By performing spatial intersection, we obtain the spatially constrained SAR flood water body mask, and its determination relationship can be expressed as: ; in, This represents the SAR flood water body mask corresponding to the t-th SAR. This represents the backscattering characteristic map of the t-th SAR time phase. This represents the water body segmentation threshold determined by the Otsu method. It refers to the hydrologically constrained area formed by the vicinity of permanent water bodies, adjacent river networks, low-lying terrain areas, or the scope of historical water systems.

[0018] This spatial constraint filters out false positives caused by low backscattering due to mountain shadows, smooth bare ground, wet farmland, building overlays, and isolated noise patches. After obtaining the flood water body mask for two adjacent SAR observation phases, a spatial overlay method is used to identify the flood water body change type. Let... and These represent the previous SAR observation time phase and the next SAR observation time phase, respectively. ; and They represent Phase and The corresponding SAR flood water body mask. By comparing the water body determination results of the same pixel in two time phases, the study area can be divided into perennial water accumulation area, flood expansion area, receding water transition area, and stable land surface area. Their spatial overlay relationship is expressed as: ; in, , , and They represent to The perennial waterlogged area, the newly expanded flood zone, the transition zone of receding water level, and the stable land surface area within the time period; and These represent the flood water body mask in the previous and next SAR observation phases, respectively. A mask value of 1 indicates a water body pixel, and a value of 0 indicates a non-water body pixel. and They represent and Non-aquatic regions in the time phase; This represents finding the intersection of spaces.

[0019] Step M3: Construct a priori daily flood state based on the daily precipitation process and the relationship between multi-temporal SAR flood changes, divide the target area into stable land surface area, perennial water accumulation area, flood expansion area, shallow water retention area, water receding and drawdown transition area, and cloud cover uncertainty area, and determine the spatial distribution of each state partition.

[0020] Based on daily precipitation events, precipitation event-driven features are constructed. Combined with SAR time-series flood variation characteristics, permanent water body distribution, digital elevation model, and distance from the river network, a priori information on daily flood states is built. Precipitation event-driven features include daily precipitation, previous day's precipitation, three-day cumulative precipitation, five-day cumulative precipitation, maximum precipitation intensity within the target time window, and the number of days since the last heavy precipitation event. These features characterize the temporal driving effects of flood expansion, retention, and recession.

[0021] In this embodiment, precipitation priors are not merely used as ordinary auxiliary data inputs, but rather to control the temporal allocation of flood conditions between adjacent SAR flood spatial anchor points. Precipitation event-driven weights for day d are constructed based on precipitation event-driven features. This weight is used to characterize the relative intensity of flood expansion, maintenance, or recession on day d.

[0022] When the cumulative precipitation over three days or five days exceeds the 75th percentile of the corresponding historical statistical value, or when the daily precipitation reaches more than 80% of the peak precipitation of this precipitation event, the precipitation process is determined to be in an intensifying phase, increasing the probability of the release of new flood expansion areas and shallow water retention areas.

[0023] The precipitation process is considered to have entered a weakening phase when any of the following conditions are met: (1) no effective precipitation for 3 consecutive days, where no effective precipitation is defined as daily precipitation less than 1 mm; (2) daily precipitation less than 20% of the peak precipitation of this precipitation event; (3) the time since the most recent precipitation peak is 3 days or more. When the precipitation process enters a weakening phase, the release probability of the drawdown transition zone is increased. Thus, the discrete flood spatial anchor points provided by multi-temporal SAR images are converted into daily flood state priors;

[0024] Using the permanent water body reference area and the multi-temporal SAR flood water body mask as spatial constraints for water bodies, and the perennial water accumulation area, flood expansion area, and receding drawdown transition area between adjacent SAR temporal phases as constraints for flood changes, and combining precipitation event driving weights, topographic depression degree, and distance from the river network, state values ​​are assigned to pixels or ground objects in the study area, resulting in stable land surface area, perennial water accumulation area, flood expansion area, shallow perched water area, receding drawdown transition area, and cloud cover uncertainty area. The classification criteria and meanings of each state category are shown in Table 1.

[0025] Table 1. Prior Categories and Meanings of Flood States

[0026] In Table 1, the low-lying area or hydrological connectivity area refers to the area located in the lowest 20% of the DEM elevation in the study area, or within a 100m buffer zone of permanent water bodies or river networks; the weak water body response refers to the SAR backscatter intensity not reaching the clear water body determination threshold, but located within 2dB above the Otsu automatic segmentation threshold, and having continuous spatial distribution characteristics.

[0027] For intermediate dates lacking SAR image coverage between two adjacent SAR observation phases, the daily release process of flood change zones is controlled using precipitation event-driven weights. When a new water body area appears at the later SAR flood spatial anchor point relative to the earlier SAR flood spatial anchor point, this area is designated as a flood expansion candidate area; when some water bodies transition from water bodies in the earlier phase to non-water or weak water bodies in the later phase, this area is designated as a receding water candidate area. For flood expansion candidate areas, higher precipitation event-driven weights, closer proximity to the river network, lower elevation, and stronger connectivity with existing water bodies result in higher priority for being assigned as a new flood expansion area or shallow perched water area; for receding water candidate areas, lower precipitation event-driven weights, longer periods of continuous dry weather, greater distance from the river network, or higher elevation result in higher priority for being assigned as a receding water transition area.

[0028] Step M4: Extract cloud and cloud shadow occlusion areas from the cloud-covered optical image during the disaster to obtain the disaster-time cloud occlusion mask; perform spatial intersection between the disaster-time cloud occlusion mask and the six state partitions in the daily flood state prior determined by precipitation prior and SAR flood spatial constraints in Step M3 to obtain the joint reconstruction mask; use the joint reconstruction mask to limit the flood-oriented reconstruction area of ​​the diffusion model, so that the diffusion reconstruction preferentially acts on the pixel areas that are obscured by clouds and fog and are identified as perennial water accumulation areas, newly expanded flood areas, shallow water stagnant areas, or receding water transition areas under the precipitation prior and SAR flood spatial constraints; and determine the non-cloud visible area, the flood state area under the cloud, the stable background area under the cloud, and the uncertain state area under the cloud based on the flood state area, stable background constraints, and cloud occlusion mask.

[0029] Cloud regions are extracted from daily or target-day cloud-covered optical remote sensing images during disasters to obtain cloud occlusion masks and cloud region confidence maps. Cloud region extraction can employ spectral thresholding, quality control bands, the Fmask algorithm, cloud detection networks, or manual interactive correction methods. By spatially overlaying cloud-occluded areas with prior flood conditions and stable background constraints, disaster-time images are divided into non-cloud-visible areas, flood-state areas under clouds, stable background areas under clouds, and areas with uncertain cloud conditions. For thin clouds, haze, and cloud shadow areas, continuous confidence information can be retained as input for subsequent regional zoning and diffusion intensity control in reconstruction.

[0030] set up This indicates the cloud cover masking in the optical imagery during the disaster on day d. The water-affected area, which is determined a priori by the daily flood state, may include perennial water accumulation areas, newly expanded flood areas, shallow water retention areas, and receding drawdown transition areas. This represents the stable background constraint region, which is mainly determined by the stable land surface region.

[0031] Based on the above mask, the disaster imagery is divided into non-cloud visible area, sub-cloud flood state area, sub-cloud stable background area, and sub-cloud uncertain state area, and their spatial relationships are represented as follows: ; ; ; ; in, This represents the non-cloud-visible area in the optical imagery during the disaster on day d, used to preserve the original optical observations; This indicates that the study area on day d is not within the cloud cover / mask area. The pixel area that is not identified as thick cloud, thin cloud, cloud shadow or area where optical observation is unreliable; This represents the joint reconstructed mask obtained by intersecting the cloud-obstructed mask and the water-affected state area, corresponding to the flood state area under the cloud; This represents the stable background region under the clouds obtained by intersecting the cloud occlusion mask and the stable background constraint region. This indicates the area within the study area that is neither part of the joint reconstruction mask nor the stable background area under clouds; This indicates an area of ​​uncertainty in the sub-cloud state that cannot be clearly classified by prior flood conditions or stable background constraints within the cloud-covered area; the above inversion operations are all limited to the study area of ​​the optical image during the disaster on day d.

[0032] Step M5: Based on the four types of disaster spatial zoning results determined in Step M4, construct differentiated optical priors for each zoning. Among them, the original observations of the disaster-time optical images are retained for the non-cloud-visible area, the flood state area under clouds adopts the sub-state flood optical prior, the stable background area under clouds adopts the stable background prior constructed from the pre-disaster cloudless optical images, and the uncertain state area under clouds adopts the weakly constrained prior or uncertainty label. The weakly constrained prior refers to using only the pre-disaster optical images, the texture information of adjacent areas, and the generation capability of the diffusion model itself for constraints, without applying flood category constraints. The reliability of its reconstruction results is lower than that of the already determined state zoning.

[0033] For the sub-cloud flood state area within the joint reconstruction mask, sample areas that are not visible from clouds and belong to the same flood state category are extracted from post-disaster optical images as spectral statistical samples. When the number of samples in the same category is insufficient, samples are supplemented from adjacent state categories with continuous flood evolution relationships. When the number of effective sample pixels meeting the non-cloud visibility condition in a certain flood state category is less than 500, the spectral statistical samples for that category are deemed insufficient. According to the principles of flood evolution continuity and spatial proximity, 100–500 homogeneous sample pixels are supplemented from adjacent state categories to ensure that the number of effective samples for spectral, index, and texture statistics for that category reaches the stable statistical requirements. Adjacent state categories include perennial water accumulation areas and newly expanded flood areas, newly expanded flood areas and shallow perched water areas, and shallow perched water areas and receding drawdown transition areas.

[0034] The spectral, index, texture, and boundary transition features are calculated separately to construct a state-specific flood optical prior. Spectral features include reflectance or its normalized values ​​in the blue, green, red, near-infrared, and shortwave infrared bands; spectral indices include one or more of the Improved Normalized Water Index (MNDWI), Normalized Water Index (NDWI), Normalized Vegetation Index (NDVI), and Automatic Water Extraction Index (AWEI) to characterize the differences between water bodies, vegetation, and bare, wet surfaces; texture features include local mean, local variance, gradient magnitude, edge density, and contrast, homogeneity, and entropy features in the gray-level co-occurrence matrix to describe the smoothness of the flood surface, the texture of turbid water bodies, and the texture of bare surfaces exposed during receding floodwaters.

[0035] For stable background areas under clouds, pre-disaster or early-stage cloudless optical images are used as background reference images. Based on the non-cloud-visible areas of the disaster-time optical images, tone normalization, brightness matching, local contrast correction, and boundary feathering are performed to obtain stable background priors. Finally, disaster-time zone-differentiated optical priors are synthesized according to spatial partitioning. ; in, Let (x,y) represent the partition-differentiated optical prior value at pixel (x,y) on day d; (x,y) represent the spatial location of the pixel within the study area. This represents the original observation value of the cloud-covered optical image at pixel (x,y) during the d-th natural disaster. This represents the sub-state optical prior values ​​for flood conditions constructed for the sub-cloud flood state zone; This represents the stable background optical prior value constructed for a stable background region under clouds; This represents a weakly constrained prior value or uncertainty identifier value constructed for the uncertain region of the cloud state. This indicates the non-cloud-visible area during the disaster on day d, within which the original optical imagery observations during the disaster are directly preserved. This represents the joint reconstructed area obtained by intersecting the cloud-masking mask and the water-affected state area, corresponding to the cloud-under-flood state area; This represents the stable background region under the cloud that satisfies the stable background constraint within the cloud-occupied area. This indicates an area of ​​uncertain under-cloud status within the cloud-covered region that is neither part of the joint reconstruction area nor part of the stable under-cloud background area.

[0036] The aforementioned regions do not overlap spatially and together constitute a zonal representation of the disaster imagery. Using the above formula, original disaster observations can be preserved in non-cloudy areas, a sub-state flood optical prior can be introduced in the flood state region under clouds, a pre-disaster stable background prior can be introduced in the stable background region under clouds, and weak constraints can be applied to uncertain regions.

[0037] Step M6: Input the cloud-covered optical image during the disaster, the cloud occlusion mask during the disaster, the joint reconstruction mask, the regional differentiated optical prior, the SAR flood spatial constraint characteristics, the precipitation event-driven characteristics, and the adaptive diffusion intensity map into the conditional diffusion model to generate the cloudless flood optical reconstruction image during the disaster.

[0038] A conditional diffusion model for optical reconstruction of floods in cloud-covered areas is constructed. The conditional diffusion model includes a basic image coding module, a conditional constraint coding module, a diffusion denoising backbone network, a decoding reconstruction module, and a pixel backfilling fusion module. The basic image coding module encodes cloud-covered optical images during the disaster; the conditional constraint coding module encodes cloud cover masks, joint reconstruction masks, regionally differentiated optical priors, SAR flood spatial constraint features, precipitation event-driven features, and adaptive diffusion intensity maps; the diffusion denoising backbone network performs stepwise inverse denoising of latent noise variables under conditional constraints; the decoding reconstruction module outputs cloudless flood optical reconstruction images during the disaster; and the pixel backfilling fusion module preserves the original observations of non-cloud-visible areas during the disaster.

[0039] An adaptive diffusion intensity map is used to adjust the noise injection intensity, condition-guided intensity, and pixel fusion weights within different spatial partitions. Different diffusion intensities are assigned to different regions based on cloud obscuration confidence, SAR flood confidence, precipitation event driving weights, flood boundary distance, and flood state category. This allows the diffusion model to focus on areas of flood change while reducing disturbances in stable background areas. The adaptive diffusion intensity map can be jointly determined by cloud obscuration confidence, SAR flood confidence, precipitation event driving weights, flood boundary distance, and flood process state category. The adaptive diffusion intensity map can be represented as: ; in, This represents the adaptive diffusion intensity at pixel (x,y) on day d, with a value ranging from 0 to 1; The cloud cover confidence level at pixel (x,y) on day d is determined based on the SAR backscattering characteristics and flood threshold segmentation results, and is normalized by combining the flood spatial constraint region. The value range is 0 to 1, which is used to characterize the confidence level of the pixel belonging to the flood region, i.e. SAR flood confidence level. Indicates the SAR flood confidence level; This indicates the driving weight of the precipitation event on day d; The boundary influence factor is calculated from the Euclidean distance from the pixel to the nearest SAR flood boundary. The boundary influence factor is normalized to the range of 0 to 1 using a distance decay function based on the boundary distance. The closer to the flood boundary, the larger the boundary influence factor. This represents the state weight determined a priori by the daily flood state constructed in step M3. The level is assigned according to the flood state category to which the cell belongs. The newly expanded flood zone is given the highest level weight, the shallow water zone is given a relatively high level weight, the receding water transition zone is given a medium level weight, the perennial water accumulation zone is given a low level weight, and the stable land surface zone and the cloud-under-state uncertain zone are given the lowest level weight. to These are the weight coefficients of the corresponding factors, and , =1; min and max are used to limit the calculation results to the range of 0 to 1.

[0040] In the conditional diffusion model, precipitation driving weights On the one hand, it participates in conditional constraint encoding as a conditional feature; on the other hand, it participates in the calculation of adaptive diffusion intensity map. For the cloud flood state area within the joint reconstruction mask and with a high precipitation driving weight, the flood prior condition guiding weight and diffusion repair intensity are increased to make the model tend to generate optical textures consistent with new floods or shallow perched water. For areas with a low precipitation driving weight and where the SAR flood spatial anchor does not show expansion, the flood prior guiding weight is reduced to suppress unreasonable water body expansion in the cloud area.

[0041] In the reverse denoising process, the flood state region under clouds within the joint reconstruction mask adopts a diffusion intensity of 0.4–0.7 and a high flood prior condition guiding weight, enabling the model to generate optical textures matching the newly expanded flood zone, shallow perched water zone, or receding drawdown transition zone under the constraints of SAR flood boundaries, precipitation event-driven weights, and sub-state flood optical priors. The stable background region under clouds adopts a diffusion intensity of 0–0.4, and its ground feature structure is constrained by stable background optical priors. The non-cloud-visible region uses a pixel backfilling mechanism to retain the original optical observations during the disaster, that is, the corresponding pixels of the optical image during the disaster are directly copied to the final reconstruction result using the non-cloud-visible region mask, and diffusion reconstruction is only performed on the area covered by the joint reconstruction mask, while the remaining pixels retain the original observation values. The uncertain region under clouds is repaired with weak constraints or outputs uncertainty indicators. The diffusion intensity and flood prior condition guiding weight range is a preferred implementation of this embodiment and can be adjusted according to the flood process characteristics of the study area, SAR detection accuracy, and diffusion model parameters.

[0042] Conditional diffusion models can be implemented using latent space diffusion models, pixel space diffusion models, diffusion patching models, or diffusion models with control branches. The conditional diffusion model takes cloud-covered optical imagery during a disaster as its base input. Under the combined effects of joint reconstruction masking, regionally differentiated optical priors, SAR flood constraints, and precipitation-driven features, it generates cloudless flood-reconstructed optical imagery during a disaster through a conditionally controlled inverse denoising process. Its conditional inputs, structural modules, and output results are shown in Table 2.

[0043] Table 2. Structure, input conditions, and output results of the conditional diffusion model.

[0044] Step M7: Based on the cloudless flood optical reconstruction image and the daily flood status prior, extract the flood inundation range, newly expanded flood zone, receding drawdown transition zone, land-water boundary, reconstruction uncertainty map, and post-disaster analysis results.

[0045] The primary output of the conditional diffusion model is the optically reconstructed image of the flood during a cloudless event. After obtaining this image, daily flood state priors, water index calculations, change detection, or semantic segmentation methods can be used to extract the flood inundation extent, newly expanded flood zone, receding water level transition zone, land-water boundary, reconstruction uncertainty map, and post-disaster statistical results. Note that the flood inundation extent, newly expanded flood zone, receding water level transition zone, land-water boundary, and post-disaster statistical results are post-disaster analysis results further extracted from the cloudless flood optically reconstructed image, and are not the direct output of the conditional diffusion model. The direct output of the conditional diffusion model is the cloudless flood optically reconstructed image during the disaster; subsequent post-disaster analysis results are used to assist in flood process interpretation, disaster extent statistics, and post-disaster impact assessment.

[0046] In this embodiment, the diffusion model is constrained during the reverse denoising process by the post-disaster SAR flood time-series anchor points, daily precipitation event-driven features, flood state transition priors, joint masking of cloud and flood state areas, and state-specific optical priors. This ensures that the optical texture, spatial structure, and temporal evolution consistency of the flood region under clouds are restored without damaging the non-reconstructed areas. This addresses problems such as the discontinuity of post-disaster flood process observations caused by cloud and rain weather affecting optical remote sensing images during floods, the difficulty of ordinary multi-temporal replacement methods in reflecting real flood changes under clouds, the poor visual interpretability of simple SAR images, and the tendency of diffusion models to generate pseudo-textures that do not conform to the flood evolution law when lacking disaster process constraints.

[0047] In this embodiment, a flood event in a lake in my country in 2020 was used as the experimental subject. Pre-disaster Sentinel-2 cloudless or low-cloud optical images, during-disaster Sentinel-2 cloudy optical images, post-disaster multi-temporal Sentinel-1 SAR images, and daily precipitation data were selected as experimental inputs. Figure 1 The process shown is used to conduct experiments on the a priori construction of daily flood conditions and the reconstruction of optical images of cloudless floods during disasters.

[0048] Among them, the Sentinel-2 optical image on June 18, 2020, served as a cloudless optical image before or during the disaster, and was used to extract the baseline area of ​​permanent water bodies and the stable background; the Sentinel-2 optical image with clouds on July 15, 2020, served as an image to be reconstructed during the disaster; the Sentinel-1 SAR images on June 20, June 26, July 2, July 8, and July 14, 2020, served as multi-temporal SAR observation data after the disaster, and were used to provide spatial constraints on flood water bodies under cloud and rain conditions; and the daily precipitation data from June 18 to July 16, 2020, were used to construct precipitation event driving characteristics.

[0049] (1) Acquire multi-source remote sensing and precipitation data and perform preprocessing: Sentinel-2 optical images from June 18, 2020, were selected as pre-disaster cloudless or partly cloudy optical images; Sentinel-2 cloudy optical images from July 15, 2020, were selected as images to be reconstructed during the disaster; Sentinel-1 SAR images from June 20, June 26, July 2, July 8, and July 14, 2020, were selected as post-disaster multi-temporal SAR images; and daily precipitation data from June 18 to July 16, 2020, were selected as precipitation event driving data.

[0050] The above data underwent unified preprocessing. Pre-disaster optical imagery, cloud-covered optical imagery during the disaster, multi-temporal SAR imagery, and daily precipitation statistics were unified to the spatial reference, study area range, and pixel grid of the Sentinel-2 image on the target day. Specifically, the optical and SAR imagery were unified to the spatial reference, study area range, and pixel grid of the Sentinel-2 image on the target day. The Sentinel-2 optical imagery underwent study area cropping, band selection, radiometric normalization, and out-of-cloud visibility enhancement. The Sentinel-1 SAR imagery underwent orbit correction, radiometric calibration, topographic correction, speckle filtering, incident angle normalization, and dB conversion. The daily precipitation data underwent study area cropping, outlier checking, time series completion, and regional averaging statistics to obtain the daily precipitation time series.

[0051] (2) Extracting the permanent water body reference area and the SAR flood water body mask: The Improved Normalized Difference Water Index (MNDWI) was calculated using the green band B03 and shortwave infrared band B11 from pre-disaster Sentinel-2 optical imagery taken on June 18, 2020. Pixels with an MNDWI value greater than 0.05 were identified as candidate pixels for permanent water bodies; those with an MNDWI value less than or equal to 0.05 were identified as candidate pixels for non-permanent water bodies. Subsequently, small patch removal, connected component filtering, and morphological closure operations were performed on the permanent water body candidate results to obtain the permanent water body baseline area.

[0052] When extracting flood water bodies from multi-temporal Sentinel-1 SAR images, the SAR images for each date are first processed with orbit correction, radiometric calibration, topographic correction, speckle filtering, incident angle normalization, and dB conversion. Then, low backscattering water body candidate areas are extracted based on VV polarization, VH polarization, and VV / VH joint backscattering features. In this embodiment, the Otsu adaptive thresholding method is used to determine the SAR water body segmentation threshold for each observation date. Specifically, the water body segmentation threshold for the SAR image on July 8, 2020 is -27.192036 dB, and the water body segmentation threshold for the SAR image on July 14, 2020 is -26.528383 dB. For the SAR images on June 20, June 26, and July 2, 2020, the same Otsu adaptive thresholding method is used to determine the water body segmentation threshold for the corresponding dates.

[0053] After obtaining SAR water body candidate regions for each date, spatial constraints are applied to them with permanent water body buffer zones, river network adjacent areas, or low-lying terrain constraints to reduce misjudgments caused by mountain shadows, smooth bare land, wet farmland, and isolated noise patches. Connectivity filtering, morphological closing operations, and small patch removal are then performed on the spatially constrained SAR water body candidate regions to obtain the SAR flood water body mask corresponding to each observation date.

[0054] (3) Constructing a priori daily flood conditions: Statistical analysis was conducted on daily precipitation data from June 18, 2020 to July 16, 2020, calculating daily precipitation, previous day precipitation, three-day cumulative precipitation, five-day cumulative precipitation, maximum precipitation intensity, and the number of days between the most recent heavy precipitation event. Based on the precipitation intensification, peak retention, and precipitation weakening processes, combined with the spatial variation relationships between flood water body masks in multi-temporal SAR, the daily release ratio and state transition trend of the flood state from July 10, 2020 to July 15, 2020 were determined.

[0055] The permanent water body reference area is designated as the S1 perennial water accumulation area; newly appearing water bodies in multi-temporal SAR images relative to the permanent water body reference area are designated as flood expansion candidate areas; areas located adjacent to flood water bodies, low-lying areas, or hydrologically connected areas that are affected by precipitation processes but have weak water body responses are designated as shallow perched water areas; areas that were water bodies in the previous temporal phase but transitioned to non-water bodies or weak water body responses in the next temporal phase are designated as receding water transition candidate areas. Based on the matching relationship between the above candidate areas and precipitation event driving characteristics, a daily flood state prior map is generated from July 10, 2020 to July 15, 2020, and the target area is divided into stable land surface area, perennial water accumulation area, flood expansion area, shallow perched water area, receding water transition area, and cloud cover uncertainty area. In this embodiment, the state codes, category names, and meanings of the daily flood state prior are shown in Table 1.

[0056] (4) Extracting the cloud cover mask during a disaster and constructing a joint reconstruction mask: Cloud regions were extracted from Sentinel-2 cloud-covered optical images taken on July 15, 2020, during the disaster, to obtain a disaster-time cloud obscuration mask. The cloud obscuration mask included thick clouds, thin clouds, haze, cloud shadows, and areas where atmospheric effects made disaster-time optical observations unreliable. Cloud region extraction was determined through a combination of brightness thresholding, saturation assessment, spectral index, quality control bands, and manual interactive correction. Continuous cloud-covered areas were obtained through morphological dilation, erosion, and hole filling.

[0057] Subsequently, the spatial intersection of the disaster-time cloud obscuration mask with the flood-affected state areas in the daily flood state prior on July 15, 2020, yields a joint reconstruction mask. The joint reconstruction mask represents pixel regions in the disaster-time optical imagery that simultaneously satisfy the criteria of being "cloud-obscured" and "belonging to a flood-related state area," and is the core region for subsequent flood-oriented reconstruction using the conditional diffusion model. Regions within the cloud-obscured area that do not belong to the flood-affected state area but satisfy stable background constraints are classified as stable background areas under clouds; regions that are not clouded, thinly clouded, or cloud-shadowed and whose disaster-time optical observations are reliable are classified as non-clouded visible areas; and cloud-obscured areas lacking reliable SAR, precipitation, or stable background constraints are classified as cloud-obscured uncertain areas.

[0058] (5) Constructing regionally differentiated optical priors: Based on the disaster-time spatial partitioning results obtained in step M4, a partition-differentiated optical prior is constructed. For non-cloud-visible areas, the original observations of the Sentinel-2 optical imagery from July 15, 2020, are directly retained. For flood-state areas under clouds, visible samples corresponding to the same or similar flood state categories are extracted from post-disaster optical images, and flood optical priors are constructed by combining SAR flood boundaries and daily flood state priors. For stable background areas under clouds, pre-disaster Sentinel-2 cloudless or low-cloud optical images from June 18, 2020, are used as stable background priors, and brightness matching, local contrast correction, and boundary feathering are used to make them consistent with the non-cloud-visible areas of the disaster-time images. For uncertain areas under clouds, weak constraint repair or uncertainty labeling is used to avoid forcibly inputting unreliable priors into the model.

[0059] Through the above partitioning process, we obtain partitioned differential optical priors that both preserve the real observations during the disaster and distinguish between flood change zones, stable background zones, and uncertain state zones.

[0060] (6) Construct a conditional diffusion model and perform optical reconstruction of cloudless floods during disasters: The Sentinel-2 optical image with clouds on July 15, 2020, is used as the base image to be reconstructed. The cloud occlusion mask and joint reconstruction mask obtained in step M4 are used as constraints on the reconstruction area. The partitioned differential optical prior obtained in step M5 is used as the image condition input. The SAR flood water body mask obtained in step M2, the daily flood state prior obtained in step M3, and the precipitation event driving weight are used as auxiliary conditions input to the conditional diffusion model.

[0061] In the non-cloud-visible area, the original optical observations at the time of the disaster are preserved during the model's reverse denoising and result fusion process; the stable background area under clouds is repaired with low intensity under the prior constraints of the stable background before the disaster; the flood state area under clouds within the joint reconstruction mask is reconstructed with flood guidance under the joint constraints of the SAR flood boundary, daily flood state prior, and sub-state flood optical prior; the uncertain area under clouds is repaired with weak constraints or the uncertainty marker is retained. Through the above processing, the cloudless flood optical reconstruction image at the time of the disaster on July 15, 2020 is obtained.

[0062] The input conditions, structural modules, and output results of the conditional diffusion model are consistent with those in Table 2. The direct output of the conditional diffusion model is the optical reconstruction image of the flood without clouds during the disaster. The flood extent, land-water boundary, receding area, or post-disaster statistical results are subsequent analysis results obtained by further extraction based on this reconstruction image.

[0063] (7) Output cloudless flood optical reconstruction images and analysis results during the disaster: The direct output of the conditional diffusion model is the cloudless flood optical reconstruction image from July 15, 2020. Based on the reconstructed image and combined with daily flood state priors, the flood inundation extent, land-water boundary, newly expanded flood zone, receding drawdown transition zone, and reconstruction uncertainty area can be further extracted. The aforementioned flood inundation extent, expansion zone, receding zone, and uncertainty area are subsequent analysis results based on the reconstructed image and are not the direct output of the conditional diffusion model.

[0064] For multi-date results, the flood extent and land-water boundary changes of adjacent dates can be compared, and the daily newly flooded area, daily receding area, flood expansion process map, and reconstruction uncertainty map can be output. If auxiliary data such as farmland, roads, residential areas, or administrative divisions are further superimposed, statistics on flooded objects and post-disaster impact assessments can be carried out; if the corresponding auxiliary data is not superimposed, this embodiment does not directly output the area of ​​flooded farmland, the length of flooded roads, or the impact range of residential areas.

[0065] The above embodiments are merely preferred embodiments of the present invention and do not constitute a limitation on the scope of protection of the present invention. Any equivalent substitutions, improvements, or combinations made to the data source, mask generation method, precipitation event encoding method, state transition model, prior construction method, conditional injection method, diffusion model structure, loss function, or post-disaster analysis module under the technical concept of the present invention should be included within the scope of protection of the present invention.

[0066] In addition to the embodiments described above, the present invention may have other implementations. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints, characterized in that: Includes the following steps: Step M1: Acquire pre-disaster cloudless optical images, cloudy optical images during the disaster, post-disaster multi-temporal synthetic aperture radar (SAR) images, daily precipitation data, and topographic and hydrological auxiliary data of the target area, and perform registration, cropping, resampling, and sequence statistical preprocessing. Step M2: Extract the permanent water body reference area based on the cloudless optical image before the disaster, and identify flood water bodies in the multi-temporal SAR image after the disaster to obtain SAR flood water body masks corresponding to different observation dates; use the SAR flood water body mask and its land-water boundary as flood spatial anchor points, and identify perennial water accumulation areas, newly expanded flood areas, and receding water transition areas based on the temporal comparison of SAR flood water body masks on different observation dates, and extract the corresponding flood spatial boundary information as spatial constraints for the subsequent daily flood state prior construction; Step M3: Construct precipitation event driving features and daily flood state priors based on the daily precipitation process and the multi-temporal SAR flood change relationship. The daily flood state priors include precipitation priors and SAR flood spatial constraints. Divide the target area into different state zones, including stable land surface zone, perennial water accumulation zone, flood expansion zone, shallow water retention zone, receding water transition zone, and cloud cover uncertainty zone, and determine the spatial distribution of each state zone. Step M4: Extract cloud and cloud shadow occlusion areas from the cloud-covered optical image during the disaster to obtain the disaster-time cloud occlusion mask; spatially intersect the disaster-time cloud occlusion mask with the state partitions in the daily flood state prior to obtain the joint reconstruction mask; the joint reconstruction mask is used to define the flood-oriented reconstruction area of ​​the conditional diffusion model, so that the diffusion reconstruction preferentially acts on the pixel areas that are obscured by clouds and fog and are identified as perennial water accumulation areas, newly expanded flood areas, shallow water retention areas, or receding water transition areas under the precipitation prior and SAR flood spatial constraints; and determine the disaster-time spatial partitions based on the flood state area, stable background constraints, and cloud occlusion mask, including non-cloud visible areas, flood state areas under clouds, stable background areas under clouds, and uncertain state areas under clouds; Step M5: Based on the disaster spatial zoning results determined in Step M4, construct a zoning-differentiated optical prior. Among them, the original observation of the disaster optical image is retained in the non-cloud visible area, the flood state area under the cloud adopts the state flood optical prior, the stable background area under the cloud adopts the stable background prior constructed by the pre-disaster cloudless optical image, and the uncertain state area under the cloud adopts the weak constraint prior. The weak constraint prior refers to the constraint only using the pre-disaster optical image, the texture information of the adjacent area and the generation capability of the diffusion model itself, without applying the flood category constraint. Its reconstruction result has a lower credibility than the determined state zoning. Step M6: Input the disaster-time cloud-covered optical image, disaster-time cloud obscuring mask, joint reconstruction mask, regional differentiated optical prior, SAR flood spatial constraint features, precipitation event-driven features, and adaptive diffusion intensity map into the conditional diffusion model to generate a disaster-time cloudless flood optical reconstruction image; the SAR flood spatial constraint features are the spatial constraint information extracted from the flood water body of the multi-temporal SAR image, including the SAR flood water body range, flood expansion boundary, receding water change area, distance information from the pixel to the flood boundary, and SAR flood detection confidence, which are used to limit the range of flood state reconstruction in the under-cloud area; Step M7: Based on the cloudless flood optical reconstruction image and the daily flood status prior, extract the flood inundation range, newly expanded flood zone, receding drawdown transition zone, land-water boundary, reconstruction uncertainty map, and post-disaster analysis results.

2. The method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints according to claim 1, characterized in that: In step M1, the pre-disaster cloudless optical image is used to extract permanent water bodies and stable background; the cloudy optical image during the disaster is the image to be reconstructed; the post-disaster multi-temporal SAR image includes at least two post-disaster SAR observation phases to provide spatial constraints for floods under cloud and rain conditions; and the daily precipitation data is used to characterize the temporal driving features of flood expansion, retention and recession processes. Optical images undergo band selection, radiometric normalization, out-of-cloud visibility enhancement, and study area cropping. SAR images undergo orbit correction, radiometric calibration, topographic correction, speckle filtering, incident angle normalization, and VV / VH dual polarization data processing. Daily precipitation data undergo spatial cropping, outlier removal, time series completion, regional averaging, or watershed zoning statistics, and daily precipitation, previous day's precipitation, three-day cumulative precipitation, five-day cumulative precipitation, and precipitation peak information are calculated to ensure that multi-source data are in the same spatial coordinate system, the same spatial resolution, and the same pixel grid.

3. The method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints according to claim 2, characterized in that: In step M2, the improved normalized water index MNDWI is calculated using the green band and shortwave infrared band of pre-disaster cloudless optical imagery, and a permanent water mask is obtained by combining threshold segmentation and morphological processing. Its expression is: ； Wherein, B03 represents the green band reflectivity, and B11 represents the shortwave infrared band reflectivity; When extracting flood water bodies from multi-temporal SAR images, based on the preprocessing results in step M1, dB conversion is performed on the SAR images of each temporal phase, and backscatter intensity feature maps are constructed according to VV and VH polarization information to obtain SAR feature data for flood water body identification; a low backscatter water body extraction method based on Otsu adaptive threshold segmentation is adopted to perform threshold segmentation on the backscatter features of VV polarization, VH polarization, or VV and VH dual polarization to obtain candidate areas for SAR water bodies in each temporal phase; For the t-th SAR observation phase, its low backscattering characteristic map is denoted as follows: The Otsu adaptive thresholding method was used to determine the water body segmentation threshold for this time phase. For cell (x,y), when Furthermore, this pixel is located within a hydrologically constrained area. If the water body is within the specified range, it is classified as a SAR flood water body pixel; otherwise, it is classified as a non-SAR flood water body pixel. The expression is as follows: ； in, This represents the SAR flood water body determination result at pixel (x,y) in the t-th SAR observation phase. A value of 1 indicates a SAR flood water body pixel, and a value of 0 indicates a non-SAR flood water body pixel. The low backscattering characteristic value at pixel (x,y) represents the t-th SAR observation phase, which is obtained from the backscattering characteristics of VV polarization, VH polarization, or VV and VH dual polarization. The hydrological spatial constraint area is a constraint mask composed of spatial areas with the possibility of flooding. Spatial areas with the possibility of flooding include buffer zones adjacent to permanent water bodies, river network adjacent areas, low-lying terrain areas, and historical water system ranges.

4. The method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints according to claim 3, characterized in that: In step M2, when the pixel At that time, the pixel was identified as a candidate pixel for low backscattering water bodies; subsequently, the candidate region for low backscattering water bodies was compared with the hydrological spatial constraint region. By performing spatial intersection, we obtain the spatially constrained SAR flood water body mask, whose determination relationship is expressed as follows: ； in, This represents the SAR flood water body mask corresponding to the t-th SAR; After obtaining the flood water body mask for two adjacent SAR observation phases, the flood water body change type is identified by overlaying the preceding and following phases. and These represent the previous SAR observation time phase and the next SAR observation time phase, respectively. By comparing the water body determination results of the same pixel in two consecutive time phases, the study area is divided into a perennial water accumulation area, a flood-induced expansion area, a receding water transition area, and a stable land surface area. Their spatial overlay relationship is expressed as follows: ； in, , , and They represent to The perennial waterlogged area, the newly expanded flood zone, the transition zone of receding water level, and the stable land surface area within the time period; and They represent Phase and The corresponding flood water body mask, where a mask value of 1 represents a water body pixel and a value of 0 represents a non-water body pixel; and They represent and Non-aquatic regions in the time phase; This represents finding the intersection of spaces.

5. The method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints according to claim 4, characterized in that: In step M3, precipitation event-driven features are constructed based on the daily precipitation process, and combined with SAR time-series flood change features, permanent water body distribution, digital elevation model and distance from river network, a priori flood state is constructed. The precipitation event-driven features include the current day's precipitation, the previous day's precipitation, the three-day cumulative precipitation, the five-day cumulative precipitation, the maximum precipitation intensity within the target time window, and the number of days since the last heavy precipitation event, which are used to characterize the time-driven effect of flood expansion, retention and recession processes. Precipitation priors are used to control the temporal allocation of flood states between adjacent SAR flood spatial anchors, and precipitation event driving weights for day d are constructed based on precipitation event driving features. This weight is used to characterize the relative intensity of flood expansion, maintenance, or recession on day d; When the cumulative precipitation over three days or five days exceeds the 75th percentile of the corresponding historical statistical value, or when the precipitation on a given day reaches more than 80% of the peak precipitation of this precipitation event, the precipitation process is determined to be in an intensifying phase, increasing the probability of the release of new flood expansion areas and shallow water retention areas. A precipitation process is considered to have entered a weakening phase when any of the following conditions are met: Condition 1: No effective precipitation for 3 consecutive days, where no effective precipitation means that the daily precipitation is less than 1 mm; Condition 2: The daily precipitation is less than 20% of the peak precipitation of this precipitation event; Condition 3: The time since the most recent precipitation peak is 3 days or more; When a precipitation process enters a weakening phase, the release probability in the drawdown transition zone is increased. This transforms the discrete flood spatial anchor points provided by multi-temporal SAR images into daily flood state priors; Using the permanent water body reference area and the multi-temporal SAR flood water body mask as water body spatial constraints, and the perennial water accumulation area, flood expansion area and receding water drawdown transition area between adjacent SAR temporal phases as flood change constraints, and combining the precipitation event driving weight, the degree of terrain depression and the distance from the river network, the state values ​​of pixels or ground objects in the study area are assigned to obtain the stable land surface area, perennial water accumulation area, flood expansion area, shallow water stagnant area, receding water drawdown transition area and cloud cover uncertainty area; The criteria for classifying each state category are as follows: When pre-disaster and post-disaster observations do not show water bodies, and the area is not located within the SAR flood water body range, flood expansion candidate area, or receding candidate area, it is classified as a stable land surface area; low-lying areas or hydrological connectivity areas refer to areas located in the lowest 20% of the DEM elevation of the study area, or areas located within a 100m buffer zone of permanent water bodies or river networks. MNDWI was calculated from cloudless optical images before the flood. An initial water mask was obtained by Otsu threshold segmentation. Isolated patches were then removed by area threshold screening, hole filling and morphological opening and closing operations to obtain permanent water candidate areas. Then, stability was judged by combining existing water frequency products or historical water distribution information. Areas that maintained water status and spatial location stability during continuous observation were identified as permanent water areas and divided into perennial water accumulation areas. If the current SAR time phase is a non-water body and the next SAR time phase is a water body, or if it is located within the water body constraint range of the target date and does not belong to the permanent water body reference area, and is supported by enhanced precipitation, low-lying terrain or river network proximity, it is classified as a flood expansion zone. When a water body is located in the vicinity of a newly expanded flood zone, a low-lying area, or a hydrologically connected area, and is supported by precipitation, but has a weaker SAR or optical water body response than a clearly defined water body area, it is classified as a shallow perched water zone. When the current SAR time phase is a water body and the next SAR time phase is a non-water body or weak water body response, and the precipitation is weakened, it does not belong to a low-lying area or a hydrologically connected area, that is, it is not located in the area of ​​the lowest 20% of the DEM elevation in the study area, and it is more than 100m away from permanent water bodies or river network adjacent buffer zones, it is classified as a receding water transition zone; a weak water body response refers to a SAR backscatter intensity that does not reach the preset water body determination threshold, but is within 2dB above the Otsu automatic segmentation threshold, and has continuous spatial distribution characteristics. When the cloud cover is affected by clouds, thin clouds, cloud shadows, or missing data, and the state cannot be determined by SAR, precipitation, permanent water bodies, or topographic and hydrological constraints, it is classified as a cloud cover uncertainty zone.

6. The method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints according to claim 5, characterized in that: In step M3, for intermediate dates between two adjacent SAR observation phases where SAR image coverage is lacking, the daily release process of the flood change zone is controlled by precipitation event-driven weights. When a new water body area appears at the later SAR flood spatial anchor point relative to the previous SAR flood spatial anchor point, this area is designated as a flood expansion candidate area. When some water bodies change from water bodies in the previous phase to non-water bodies or weak water bodies in the later phase, this area is designated as a receding water candidate area. For flood expansion candidate areas, the higher the precipitation event-driven weight, the closer to the river network, the lower the terrain, and the stronger the connectivity with existing water bodies, the higher the priority of being assigned as a new flood expansion area or a shallow stagnant water area. For receding water candidate areas, the lower the precipitation event-driven weight, the longer the continuous rainless period, the farther away from the river network, or the higher the terrain, the higher the priority of being assigned as a receding water transition area.

7. The method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints according to claim 6, characterized in that: In step M4, cloud areas are extracted from the target day or daily cloud-covered optical remote sensing images during the disaster to obtain a cloud occlusion mask and a cloud area confidence map. The cloud area extraction adopts spectral thresholding, quality control bands, Fmask algorithm, cloud detection network or manual interactive correction method. For thin clouds, haze and cloud shadow areas, continuous confidence information is retained as input for subsequent reconstruction of regional zoning and diffusion intensity control. set up This indicates the cloud cover masking in the optical imagery during the disaster on day d. It represents the water-affected area determined a priori by the daily flood state, including the perennial water accumulation area, the newly expanded flood area, the shallow water retention area, and the receding water transition area; This represents the stable background constraint area determined by the stable land surface area; based on the above mask, the disaster imagery is divided into non-cloud visible area, cloud-below flood state area, cloud-below stable background area, and cloud-below uncertain state area, and their spatial relationships are expressed as follows: ； ； ； ； in, This represents the non-cloud-visible area in the optical imagery during the disaster on day d, used to preserve the original optical observations; This indicates that the study area on day d is not within the cloud cover / mask area. The pixel area that is not identified as thick cloud, thin cloud, cloud shadow or area where optical observation is unreliable; This represents the joint reconstructed mask obtained by intersecting the cloud-obstructed mask and the water-affected state area, corresponding to the flood state area under the cloud; This represents the stable background region under the clouds obtained by intersecting the cloud occlusion mask and the stable background constraint region. This indicates the area within the study area that is neither part of the joint reconstruction mask nor the stable background area under clouds; This indicates an uncertain region of the under-cloud state within the cloud-covered area that cannot be clearly classified by prior flood conditions or stable background constraints.

8. The method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints according to claim 7, characterized in that: In step M5, for the flood state area under the cloud within the joint reconstruction mask, sample areas that are not visible from the cloud and are located in the same or adjacent flood state categories are extracted from the post-disaster optical image. Their spectral, index, texture, and boundary transition features are calculated respectively to construct the sub-state flood optical prior. The spectral features include reflectance or normalized values ​​of the blue, green, red, near-infrared, and short-wave infrared bands. The spectral indices include one or more of the improved normalized water index MNDWI, normalized water index NDWI, normalized vegetation index NDVI, and automatic water extraction index AWEI, which are used to characterize the differences between water bodies, vegetation, and bare wet surfaces. The texture features include local mean, local variance, gradient magnitude, edge density, and contrast, homogeneity, and entropy features in the gray-level co-occurrence matrix, which are used to describe the smoothness of the flood water surface, the texture of turbid water bodies, and the texture of bare surfaces exposed by receding water. For stable background areas under clouds, pre-disaster or early-stage cloudless optical images are used as background reference images. Based on the non-cloud-visible areas of the disaster-time optical images, tone normalization, brightness matching, local contrast correction, and boundary feathering are performed to obtain stable background priors. Finally, disaster-time zone-differentiated optical priors are synthesized according to spatial partitioning. ； in, Let (x,y) represent the partition-differentiated optical prior value at pixel (x,y) on day d; (x,y) represent the spatial location of the pixel within the study area. This represents the original observation value of the cloud-covered optical image at pixel (x,y) on day d during the disaster. This represents the sub-state optical prior values ​​for flood conditions constructed for the sub-cloud flood state zone; This represents the stable background optical prior value constructed for a stable background region under clouds; This represents the weakly constrained prior values ​​constructed for the uncertain region of the cloud state; This indicates the non-cloud-visible area during the disaster on day d, within which the original optical imagery observations during the disaster are directly preserved. This represents the joint reconstructed area obtained by intersecting the cloud-masking mask and the water-affected state area, corresponding to the cloud-under-flood state area; This represents the stable background region under the cloud that satisfies the stable background constraint within the cloud-occupied area. This indicates an area within the cloud-covered zone that is neither part of the joint reconstruction area nor part of the stable background area under the clouds, representing an uncertain area under the clouds. These areas do not overlap spatially and together constitute the zoning representation of the disaster image.

9. The method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints according to claim 8, characterized in that: In step M6, a conditional diffusion model for optical reconstruction of floods in cloud-covered areas is constructed. This model includes a basic image coding module, a conditional constraint coding module, a diffusion denoising backbone network, a decoding reconstruction module, and a pixel backfilling fusion module. The basic image coding module is used to encode optical images with clouds during disasters. The conditional constraint coding module is used to encode cloud cover masks, joint reconstruction masks, regional differentiated optical priors, SAR flood spatial constraint features, precipitation event-driven features, and adaptive diffusion intensity maps. The diffusion denoising backbone network is used to perform stepwise reverse denoising of noise latent variables under conditional constraints. The decoding and reconstruction module is used to output optically reconstructed images of floods without clouds during disasters; the pixel backfilling and fusion module is used to preserve the original observations of non-cloud-visible areas during disasters. The adaptive diffusion intensity map is used to adjust the noise injection intensity, conditional guidance intensity, and pixel fusion weights in different spatial partitions. The adaptive diffusion intensity map is jointly determined by cloud obscuration confidence, SAR flood confidence, precipitation event driving weights, flood boundary distance, and flood process state category; the adaptive diffusion intensity map is represented as: ； in, This represents the adaptive diffusion intensity at pixel (x,y) on day d, with a value ranging from 0 to 1; This represents the confidence level of cloud occlusion at pixel (x,y) on day d; Indicates the SAR flood confidence level; This indicates the driving weight of the precipitation event on day d; This represents the boundary influence factor calculated from the Euclidean distance from the pixel to the nearest SAR flood boundary. This boundary influence factor is normalized to the range of 0 to 1 using a distance decay function based on the boundary distance. The closer to the flood boundary, the larger the boundary influence factor. The state weight is determined a priori by the daily flood state. The level is assigned according to the flood state category to which the cell belongs. The newly expanded flood area is assigned the highest level weight, the shallow water area is assigned a relatively high level weight, the receding water transition area is assigned a medium level weight, the perennial water accumulation area is assigned a lower level weight, and the stable land surface area and the cloud under uncertain state area are assigned the lowest level weight. to They are respectively , , , as well as The weighting coefficients, and , ; min and max are used to limit the calculation results to the range of 0 to 1; In the conditional diffusion model, precipitation events drive the weights. It not only participates in conditional constraint coding as a conditional feature, but also in the calculation of adaptive diffusion intensity map; In the reverse denoising process, the flood state area under clouds within the joint reconstruction mask adopts a diffusion intensity of 0.4–0.7 and a flood prior condition guiding weight of 0.5–1.0, enabling the model to generate optical textures that match the newly expanded flood area, shallow perched water area, or receding drawdown transition area under the constraints of SAR flood boundary, precipitation event driving weight, and sub-state flood optical prior. The non-water-contaminated stable background area under clouds adopts a diffusion intensity of less than 0.4 and a flood prior condition guiding weight of less than 0.5, and its ground structure is constrained by the stable background prior. The non-cloud-visible area retains the original optical observations during the disaster through a pixel backfilling mechanism, that is, the corresponding pixels of the optical image during the disaster are directly copied to the final reconstruction result using the non-cloud-visible area mask, and diffusion reconstruction is only performed on the area covered by the joint reconstruction mask, while the remaining pixels retain the original observation values. The uncertain state area under clouds is repaired with weak constraints.

10. The method for reconstructing flood-affected images in cloud-covered areas based on precipitation priors and SAR constraints according to claim 9, characterized in that: In step M7, the output of the conditional diffusion model is the optical reconstruction image of the flood during the disaster without clouds. After obtaining the optical reconstruction image of the flood during the disaster without clouds, the flood inundation range, newly expanded flood zone, receding drawdown transition zone, land-water boundary, reconstruction uncertainty map, and post-disaster statistical results are extracted by combining daily flood state priors, water body index calculations, change detection, or semantic segmentation methods. The flood inundation range, newly expanded flood zone, receding drawdown transition zone, land-water boundary, and post-disaster statistical results are post-disaster analysis results further extracted from the optical reconstruction image of the flood during the disaster without clouds, rather than the direct output of the conditional diffusion model. The direct output of the conditional diffusion model is the optical reconstruction image of the flood during the disaster without clouds, and the subsequent post-disaster analysis results are used to assist in the interpretation of the flood process, the statistics of the disaster range, and the assessment of post-disaster impact.

Images