Multi-stage lightweight natural fire remote sensing image inference method based on feature complexity self-adaptation

Abstract

The application provides a multi-stage lightweight natural fire remote sensing image reasoning method based on feature complexity self-adaption, comprising: extracting shallow surface features of an original fire remote sensing image, calculating a global complexity score of a fire scene based on the shallow surface features; constructing a feature gradient extraction rule according to the global complexity score, combining the feature gradient extraction rule and a gating network to perform classification reasoning calculation, and obtaining a self-adaptive reasoning path; designing coarse, medium and fine stages to perform staged feature extraction on the original fire remote sensing image, and obtaining multi-level features; compressing, caching and self-adaptively fusing the multi-level features, and outputting a natural fire remote sensing image reasoning result. The application can self-adaptively adjust a reasoning path and a calculation amount according to an input image feature complexity, significantly reduce a calculation cost and a reasoning time delay under the premise of ensuring fire identification and segmentation accuracy, and support various scenes such as natural fire monitoring, disaster risk assessment and emergency response.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a multi-stage lightweight inference method for natural fire remote sensing images based on adaptive feature complexity. Background Technology

[0002] Natural fires (such as forest fires and grassland fires) are characterized by their suddenness, rapid spread, and wide impact, posing a significant threat to ecosystems, human life and property, and socio-economic development. Currently, natural fire monitoring and identification technologies can be categorized into three main types: ground-based monitoring, aerial monitoring, and satellite remote sensing monitoring. These technologies complement each other and cover different monitoring scenarios. Among them, satellite remote sensing monitoring is the core method for large-scale, all-weather macroscopic monitoring. Natural fire monitoring requires the rapid and accurate identification of fire points, smoke, and burned areas within large-area, high-resolution remote sensing images. However, satellites, drones, and aerial platforms are typically limited by resources such as computing power, power consumption, and bandwidth, preventing them from running extremely large models for extended periods; furthermore, different areas within the same image can exhibit significant variations.

[0003] With the rapid development of deep learning technology, models based on Convolutional Neural Networks (CNNs) and Visual Transformers (ViTs) have become important techniques for remote sensing image recognition, scene classification, and change detection. However, these models share common characteristics such as large parameter counts and high computational complexity. For example, a typical ViT-B model has over 80 million parameters and extremely high FLOPs during inference, making it difficult to deploy efficiently on edge devices, drones, or satellites. Therefore, lightweight design is crucial. However, in remote sensing tasks, the models have complex numerical distributions and large dynamic ranges, and quantization may lead to a decrease in feature representation capabilities. This is particularly true in tasks requiring high-precision boundary recognition, such as fire monitoring and ground feature change detection, where quantization errors can significantly impact the results.

[0004] Currently, most lightweight methods focus on static structure compression or single-level dynamic adjustment, making it difficult to simultaneously balance inference efficiency and remote sensing mission accuracy. In real-world remote sensing scenarios, images from different regions vary greatly. For example, densely built-up urban areas have complex textures and clear boundaries; mountainous and forested areas have continuous textures but strong directionality; and desert and arid areas have simple features but large spectral variations. Using a uniform, fixed computation path would waste computational resources in simple areas, while inference might be insufficient in complex areas. While existing dynamic inference frameworks (such as SkipNet and DynamicConv) can partially alleviate computational redundancy, they still lack adaptive mechanisms for the complexity of remote sensing images. Therefore, how to dynamically adjust the inference structure based on the complexity of input features while maintaining model accuracy, thereby achieving optimal utilization of computational resources and a balance between task performance, has become a key technical challenge that urgently needs to be overcome in the field of lightweight remote sensing image deployment. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a multi-stage lightweight natural fire remote sensing image inference method based on adaptive feature complexity. This method can adaptively adjust the inference path and computational load according to the feature complexity of the input image, significantly reducing computational costs and inference latency while ensuring the accuracy of fire identification and segmentation. It supports various scenarios such as natural fire monitoring, disaster risk assessment, and emergency response.

[0006] To achieve the above objectives, this invention provides the following solution: a multi-stage lightweight natural fire remote sensing image inference method based on adaptive feature complexity, comprising: Acquire the original fire remote sensing image, extract the shallow surface features of the original fire remote sensing image, and calculate the global complexity score of the fire scene based on the shallow surface features. Based on the global complexity score, feature gradient extraction rules are constructed. Combining the feature gradient extraction rules and the gating network, classification and reasoning calculations are performed on the original fire remote sensing image to obtain an adaptive reasoning path. Based on the type of the adaptive inference path, a coarse-stage, medium-stage, and fine-stage feature extraction is performed on the original fire remote sensing image to obtain multi-level features. The multi-level features output by the adaptive inference path are compressed, cached, and adaptively fused to obtain reusable features, and the inference results of the remote sensing image of natural fire are output.

[0007] Optionally, the process involves acquiring original fire remote sensing images, extracting shallow surface features from the original fire remote sensing images, and calculating a global complexity score for the fire scene based on the shallow surface features, including: The original fire remote sensing image is acquired, and the primary texture, edge and color changes of the original fire remote sensing image are extracted using a lightweight convolutional network to obtain a shallow feature map. Based on the original fire remote sensing image, the complexity of fire points and ground feature outlines is defined, the feature gradient variance is obtained, the diversity of regional content is quantified, the feature entropy is obtained, and the burnt areas or continuous surface texture changes are identified to obtain the feature density. Based on the shallow feature mapping, the difficulty and complexity of the fire scene are calculated using the feature gradient variance, the information entropy, and the texture density to obtain a global complexity score; the calculation expression for the global complexity signal is: ; in, The feature gradient variance reflects the degree of drastic change in local edges. The feature entropy is the pixel distribution information density. To measure the feature density of local texture repeatability and spatial compactness, These are the learnable weight coefficients.

[0008] Optionally, feature gradient extraction rules are constructed based on the global complexity score. Combining these rules with a gating network, classification and inference calculations are performed on the original fire remote sensing image to obtain an adaptive inference path, including: Set a first adaptive threshold and a second adaptive threshold, determine whether the global complexity score is less than the first adaptive threshold, if so, select a lightweight convolution path for shallow feature extraction, determine whether the global complexity score is greater than or equal to the first adaptive threshold and less than the second adaptive threshold, if so, select a medium-depth feature extraction path for feature extraction, determine whether the global complexity is greater than or equal to the second adaptive threshold, if so, select a complete backbone network inference path for high-precision feature extraction of complex regions, and complete the construction of feature gradient extraction rules; By combining the feature gradient extraction rules and the gating network, a path dynamic selection mechanism is obtained. Based on the global complexity score, the path dynamic selection mechanism is used to perform classification and reasoning calculations on the original fire remote sensing image to obtain an adaptive reasoning path.

[0009] Optionally, based on the type of the adaptive inference path, a coarse-stage, medium-stage, and fine-stage feature extraction process is designed for the original fire remote sensing image to obtain multi-level features, including: When the adaptive inference path is a lightweight convolutional path, the lightweight network is used to quickly extract features from the low-resolution image in the original fire remote sensing image, obtain the location information of the fire scene and the background area information, and complete the coarse-stage feature extraction. When the adaptive inference path is a medium-depth feature extraction path, medium-resolution feature extraction is performed on the fire-related areas in the original fire remote sensing image, and the area boundaries and area categories are refined to complete the medium-stage feature extraction. When the adaptive inference path is a complete backbone network inference path, the Transformer layer or multi-scale fusion unit is used to extract high-resolution features from the multi-source fire points, smoke overlap areas, and detailed boundaries between burned and unburned areas in the original fire remote sensing image, thus completing fine-stage feature extraction and obtaining multi-level features.

[0010] Optionally, the multi-level features output by the adaptive inference path are compressed, cached, and adaptively fused to obtain reused features, and the inference results of the natural fire remote sensing image are output, including: Using 1×1 convolution or linear mapping, channel dimensionality reduction and feature summarization are performed on the feature tensors of coarse-stage features, medium-stage features and fine-stage features respectively to obtain coarse-compressed features, medium-compressed features and fine-compressed features. Then, lightweight memory buffer or shared memory cache is used to compress the features. The coarse-compressed features and the fine-compressed features are adaptively weighted and fused to obtain reused features, which are then used to output the inference results of the natural fire remote sensing image. The calculation expression for the reused features is as follows: ; in, These are the coarse-stage features output from the previous stage. This represents the detailed stage features output for the current stage. The learnable fusion coefficient.

[0011] This invention discloses the following technical effects by providing a multi-stage lightweight inference method for remote sensing images of natural fires based on adaptive feature complexity: 1. Balancing computational resources and task accuracy: Adaptive path and stage control using complexity, so that a large number of simple background areas no longer consume the computing power of the entire process; focus the computation on complex areas such as fire points, smoke, and burnt areas, so that the computation is performed only where it is complex.

[0012] 2. Specially designed for the characteristics of remote sensing scenarios: It takes into account the unique problems of remote sensing, such as high resolution, multi-scale, and multi-regional differences; unlike the dynamic computing mechanism in the general natural image field, it is specially adapted to natural fire remote sensing tasks.

[0013] 3. Multi-stage, multi-scale collaboration: The three stages of coarse, medium and fine are combined with caching and fusion to take into account both the global structure and detailed contours; it has advantages in precision-sensitive tasks such as fire detection, smoke boundary, and burn area segmentation.

[0014] 4. The unity of lightweight and dynamic reasoning: It does not rely solely on static lightweight techniques such as pruning, quantization, and knowledge distillation, but achieves lightweight and adaptiveness through the structural design itself; it is more suitable for deployment on various hardware platforms such as GPU, FPGA, and embedded systems.

[0015] 5. Engineering Applications and Expansion Potential: Applicable to various natural disaster scenarios such as fire monitoring, post-disaster assessment, and surface change detection; the framework can be extended to multimodal inputs (visible light, infrared, SAR) and cross-temporal analysis, supporting more complex remote sensing intelligent processing tasks.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this invention provides a multi-stage lightweight inference method for remote sensing images of natural fires based on adaptive feature complexity, including: Step 1: Acquire the original fire remote sensing image, extract the shallow surface features of the original fire remote sensing image, and calculate the global complexity score of the fire scene based on the shallow surface features; specifically including: The original fire remote sensing image is acquired, and the primary texture, edge and color changes of the original fire remote sensing image are extracted using a lightweight convolutional network to obtain a shallow feature map. Based on the original fire remote sensing image, the complexity of fire points and ground feature outlines is defined, the feature gradient variance is obtained, the diversity of regional content is quantified, the feature entropy is obtained, and the burnt areas or continuous surface texture changes are identified to obtain the feature density. Based on the shallow feature mapping, the difficulty and complexity of the fire scene are calculated using the feature gradient variance, the information entropy, and the texture density to obtain a global complexity score; the calculation expression for the global complexity signal is: ; in, The feature gradient variance reflects the degree of drastic change in local edges. The feature entropy is the pixel distribution information density. To measure the feature density of local texture repeatability and spatial compactness, These are the learnable weight coefficients.

[0022] Calculated The value, serving as a global complexity signal for the system, is passed to the dynamic path selection module to control the activation strategy of subsequent network layers. Through this module, the model gains the ability to quantitatively perceive the "difficulty" of the input image, thereby enabling it to automatically allocate computing resources to different areas in fire monitoring and laying the foundation for adaptive inference.

[0023] Step 2: Construct feature gradient extraction rules based on the global complexity score. Combine the feature gradient extraction rules and the gating network to perform classification inference calculations on the original fire remote sensing image to obtain an adaptive inference path; specifically including: Set a first adaptive threshold and a second adaptive threshold, determine whether the global complexity score is less than the first adaptive threshold, if so, select a lightweight convolution path for shallow feature extraction, determine whether the global complexity score is greater than or equal to the first adaptive threshold and less than the second adaptive threshold, if so, select a medium-depth feature extraction path for feature extraction, determine whether the global complexity is greater than or equal to the second adaptive threshold, if so, select a complete backbone network inference path for high-precision feature extraction of complex regions, and complete the construction of feature gradient extraction rules; By combining the feature gradient extraction rules and the gating network, a path dynamic selection mechanism is obtained. Based on the global complexity score, the path dynamic selection mechanism is used to perform classification and reasoning calculations on the original fire remote sensing image to obtain an adaptive reasoning path.

[0024] Specifically, scoring is based on complexity. The gating network automatically determines the inference path and uses an adaptive threshold. T 1. T 2. Compare. When At lower levels, the model only uses the basic feature extraction path; when When the threshold is high, mid-level or deep inference units are activated to enhance feature representation capabilities. The gating network employs learnable parameters to achieve end-to-end training. Its selection strategy is as follows: ; in: L 1: Lightweight convolutional paths perform only shallow feature extraction, suitable for regions with simple textures and clear boundaries; L 2: Medium-depth feature extraction path, maintaining a moderate amount of computation while ensuring a certain level of accuracy; L 3: Complete backbone network inference path, used for high-precision feature extraction in complex regions.

[0025] In the inference process of remote sensing images of natural fires, the texture complexity and spectral characteristics of different regions vary significantly. If a fixed inference path is used uniformly, it will lead to a waste of computational resources or insufficient recognition. Through the dynamic path control mechanism of this module, the system can adaptively activate the most suitable inference path according to the complexity of the input image, realizing the fine recognition of complex areas and the rapid processing of simple areas, thereby significantly improving the overall inference efficiency and task adaptability.

[0026] Step 3: Based on the type of the adaptive inference path, design coarse, medium, and fine stages to extract features from the original fire remote sensing image in stages, obtaining multi-level features; specifically including: When the adaptive inference path is a lightweight convolutional path, the lightweight network is used to quickly extract features from the low-resolution image in the original fire remote sensing image, obtain the location information of the fire scene and the background area information, and complete the coarse-stage feature extraction. Specifically, in the coarse-grained stage: a lightweight network is used to quickly extract low-resolution features to obtain global structural information of the scene. For regions with simple textures and gradual changes (such as unaffected surfaces or sky backgrounds), if the complexity score is low, the model can complete the prediction at this stage and terminate the inference early, thereby saving computational resources.

[0027] When the adaptive inference path is a medium-depth feature extraction path, medium-resolution feature extraction is performed on the fire-related areas in the original fire remote sensing image, and the area boundaries and area categories are refined to complete the medium-stage feature extraction. Specifically, in the intermediate stage: medium-resolution features are extracted from fire-related areas (such as the edges of smoke and the vicinity of hot spots) to further refine the boundary and region discrimination results. When the input image has local high-complexity variations but the overall scene is still relatively simple, this stage can maintain a moderate computational load while ensuring recognition accuracy.

[0028] When the adaptive inference path is a complete backbone network inference path, the Transformer layer or multi-scale fusion unit is used to extract high-resolution features from the multi-source fire points, smoke overlapping areas, and detailed boundaries between burned and unburned areas in the original fire remote sensing image, thus completing fine-stage feature extraction and obtaining multi-level features. Specifically, in the finer stages: for areas with high complexity (such as multi-source fire points, overlapping smoke areas, or the boundaries of burned areas), a complete high-resolution feature extraction module (such as a Transformer layer or a multi-scale fusion unit) is enabled to obtain more refined texture expression and boundary delineation capabilities.

[0029] This invention can also be based on CThe system can dynamically skip or prematurely terminate certain stages. In actual inference, if the system detects the characteristic complexity of the coarse-grained stage... C (1) It is below the threshold T If the input image is classified as having a "single fire source and low background complexity" scenario, the model skips the middle stage and retains only the fine stage for detail correction. Conversely, when the fire area is large and the background perturbation is strong, the system maintains the complete three-stage inference process to ensure semantic segmentation accuracy. Through this staged dynamic inference mechanism, the model can achieve significant optimization of inference latency and energy consumption while maintaining accuracy in natural fire scenarios.

[0030] Step 4: Compress, cache, and adaptively fuse the multi-level features output by the adaptive inference path to obtain reusable features, and output the inference result of the natural fire remote sensing image. Specifically, this includes: Using 1×1 convolution or linear mapping, channel dimensionality reduction and feature summarization are performed on the feature tensors of coarse-stage features, medium-stage features and fine-stage features respectively to obtain coarse-compressed features, medium-compressed features and fine-compressed features. Then, lightweight memory buffer or shared memory cache is used to compress the features.

[0031] Feature caching aims to avoid redundant computation and enable cross-stage feature reuse during multi-stage inference. Simultaneously, an attention fusion mechanism maintains consistency and contextual relevance of features across different stages. In remote sensing images of natural fires, fire points, smoke, and burnt areas often exhibit clear hierarchical and localized variations in spatial distribution. Processing features extracted at each stage independently can easily lead to information fragmentation between regions.

[0032] The coarse and fine compressed features are adaptively weighted and fused. Specifically, in the next stage, cached features are loaded and fused with newly extracted features according to adaptive weights to obtain reusable features, outputting the inference result of the natural fire remote sensing image. Cache reading and updating are transmitted in the form of feature packets, which contain feature vectors, spatial indices, and timestamp information to ensure cross-stage information matching.

[0033] The calculation expression for the reuse feature is: ; in, These are the coarse-stage features output from the previous stage. This represents the detailed stage features output for the current stage. The learnable fusion coefficient.

[0034] In fire monitoring tasks, this strategy can effectively avoid boundary offset problems in the multi-scale fusion process, enabling the model to maintain high consistency and spatiotemporal stability in detail-sensitive tasks such as fire detection and smoke region segmentation, thereby further improving the overall inference accuracy and robustness.

[0035] Through this progressive design—from feature complexity assessment to dynamic path selection, multi-stage inference, and feature caching fusion—the system can automatically adjust the computational load based on the complexity of the input fire remote sensing imagery, achieving the adaptive inference goal of being "lightweight and fast in simple scenarios, and precise and accurate in complex areas." This scheme significantly reduces the computational resource consumption of the model in natural fire monitoring and disaster identification while maintaining high-precision feature representation capabilities, demonstrating good versatility and engineering application value.

[0036] Therefore, this invention provides a multi-stage lightweight natural fire remote sensing image inference method based on adaptive feature complexity. This method can adaptively adjust the inference path and computational load according to the feature complexity of the input image. While ensuring the accuracy of fire identification and segmentation, it significantly reduces computational costs and inference latency, and supports various scenarios such as natural fire monitoring, disaster risk assessment, and emergency response.

[0037] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0038] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-stage lightweight inference method for natural fire remote sensing images based on adaptive feature complexity, characterized in that, include: Acquire the original fire remote sensing image, extract the shallow surface features of the original fire remote sensing image, and calculate the global complexity score of the fire scene based on the shallow surface features. Based on the global complexity score, feature gradient extraction rules are constructed. Combining the feature gradient extraction rules and the gating network, classification and reasoning calculations are performed on the original fire remote sensing image to obtain an adaptive reasoning path. Based on the type of the adaptive inference path, a coarse-stage, medium-stage, and fine-stage feature extraction is performed on the original fire remote sensing image to obtain multi-level features. The multi-level features output by the adaptive inference path are compressed, cached, and adaptively fused to obtain reused features, and the inference results of the remote sensing image of natural fire are output. Based on the global complexity score, feature gradient extraction rules are constructed. Combining these rules with a gating network, classification and inference calculations are performed on the original fire remote sensing image to obtain an adaptive inference path, including: Set a first adaptive threshold and a second adaptive threshold, determine whether the global complexity score is less than the first adaptive threshold, if so, select a lightweight convolution path for shallow feature extraction, determine whether the global complexity score is greater than or equal to the first adaptive threshold and less than the second adaptive threshold, if so, select a medium-depth feature extraction path for feature extraction, determine whether the global complexity is greater than or equal to the second adaptive threshold, if so, select a complete backbone network inference path for high-precision feature extraction of complex regions, and complete the construction of feature gradient extraction rules; Combining the feature gradient extraction rules and the gating network, a path dynamic selection mechanism is obtained. Based on the global complexity score, the path dynamic selection mechanism is used to perform classification reasoning calculations on the original fire remote sensing image to obtain an adaptive reasoning path. Based on the type of the adaptive inference path, a coarse-stage, medium-stage, and fine-stage approach is designed to extract features from the original fire remote sensing image in stages, resulting in multi-level features, including: When the adaptive inference path is a lightweight convolutional path, the lightweight network is used to quickly extract features from the low-resolution image in the original fire remote sensing image, obtain the location information of the fire scene and the background area information, and complete the coarse-stage feature extraction. When the adaptive inference path is a medium-depth feature extraction path, the coarse-stage features are used to extract medium-resolution features from the fire-related areas in the original fire remote sensing image, refine the area boundaries and area categories, and complete the medium-stage feature extraction. When the adaptive inference path is a complete backbone network inference path, the coarse-stage features and the medium-stage features are used to extract high-resolution features from the multi-source fire points, smoke overlap areas, and detailed boundaries between burned and unburned areas in the original fire remote sensing image through the Transformer layer or multi-scale fusion unit, thereby completing the fine-stage feature extraction and obtaining multi-level features.

2. The multi-stage lightweight natural fire remote sensing image inference method based on adaptive feature complexity according to claim 1, characterized in that, Acquire raw fire remote sensing images, extract shallow surface features from the raw fire remote sensing images, and calculate the global complexity score of the fire scene based on the shallow surface features, including: The original fire remote sensing image is acquired, and the primary texture, edge and color changes of the original fire remote sensing image are extracted using a lightweight convolutional network to obtain a shallow feature map. Based on the original fire remote sensing image, the complexity of fire points and ground feature outlines is defined, the feature gradient variance is obtained, the diversity of regional content is quantified, the feature entropy is obtained, and the burnt areas or continuous surface texture changes are identified to obtain the feature density. Based on the shallow feature mapping, the difficulty and complexity of the fire scenario are calculated using the feature gradient variance, the feature entropy, and the feature density, resulting in a global complexity score. The expression for calculating the global complexity signal is as follows: ； in, The feature gradient variance reflects the degree of drastic change in local edges. The feature entropy is the pixel distribution information density. To measure the feature density of local texture repeatability and spatial compactness, These are the learnable weight coefficients.

3. The multi-stage lightweight natural fire remote sensing image inference method based on adaptive feature complexity according to claim 2, characterized in that, The multi-level features output by the adaptive inference path are compressed, cached, and adaptively fused to obtain reused features, and the inference results of the natural fire remote sensing image are output, including: Using 1×1 convolution or linear mapping, channel dimensionality reduction and feature summarization are performed on the feature tensors of coarse-stage features, medium-stage features and fine-stage features respectively to obtain coarse-compressed features, medium-compressed features and fine-compressed features. Then, lightweight memory buffer or shared memory cache is used to compress the features. In the next stage, cached features are loaded and fused with newly extracted features using adaptive weights to obtain reused features, outputting the inference results of the natural fire remote sensing image; the calculation expression for the reused features is: ； in, This represents the coarse-stage features output from the previous path. This outputs the fine-stage features for the current path. The learnable fusion coefficient.

Images