CN-121415269-B - AI-based karst mountain land illegal cue intelligent discovery and supervision method

CN121415269BCN 121415269 BCN121415269 BCN 121415269BCN-121415269-B

Abstract

The invention discloses an AI-based karst mountain land illegal cue intelligent discovery and supervision method, which relates to the technical field of land monitoring, and comprises the following steps of S1, acquiring satellite remote sensing images with resolution ratio more than or equal to 2 meters; the method comprises the steps of S2, constructing a terrain awareness positioning network to generate a land probability positioning mask to position a mountain land area, S3, constructing a newly added land illegal change detection model to perform illegal detection on the positioned land area, S4, generating a land illegal probability map and an illegal land suspected area vector boundary based on the output of the newly added land illegal change detection model, S5, uploading the illegal land suspected area vector boundary and a front-back time phase screenshot of the vector boundary to a management platform, performing field investigation through a mobile terminal, and performing continuous tracking supervision on suspected illegal ground wires to realize intelligent supervision of illegal behaviors. The method of the invention improves the regulation efficiency of the land illegal regulations in the karst mountain area.

Inventors

LU JUN
XIONG JING
MA HAOJIE
WANG HONGLEI
LIU SHIQI
ZHANG LANLAN
XU HONG
XIE YIJUAN
Yan Wenpu
ZHANG YE
ZHANG HONG

Assignees

贵州省第三测绘院(贵州省国土资源遥感监测中心)

Dates

Publication Date: 20260512
Application Date: 20251224

Claims (7)

1. The intelligent detection and supervision method for the illegal soil clues in the karst mountain area based on the AI is characterized by comprising the following steps: s1, acquiring satellite remote sensing images with resolution ratio more than or equal to 2 meters; s2, constructing a terrain awareness positioning network to generate a land probability positioning mask, and positioning a mountain land area; S3, constructing a newly-added land illegal change detection model to perform illegal detection on the located land area; s4, generating a land illegal probability map and an illegal land suspected region vector boundary based on the output of the newly-added land illegal change detection model; S5, uploading the suspected regional vector boundary of the illegal land and the front-back time phase screenshot of the vector boundary to a management platform, carrying out field investigation through a mobile terminal, and carrying out continuous tracking supervision on the suspected illegal ground cable to realize intelligent supervision on illegal behaviors; In the step S2, a land probability positioning mask is generated by constructing a terrain awareness positioning network, and the method specifically includes the following steps: Carrying out band superposition on the elevation data, the gradient and the curvature of the DEM to form a terrain feature cube; Extracting the topographic features from the topographic feature cube through the two layers of convolution layers; inputting the satellite remote sensing image into ResNet network to extract spectral image characteristics; generating a land probability positioning mask by adopting a characteristic attention fusion mechanism; The loss function of the terrain-aware positioning network comprises positioning loss, detection loss and terrain consistency loss; The positioning loss is used for guaranteeing the accuracy of land area segmentation, the detection loss is used for guaranteeing the classification accuracy of illegal change samples, the sample imbalance is processed, and the topography consistency loss is used for guaranteeing the consistency of land boundaries and topography characteristics; The loss function of the terrain-aware positioning network is: Wherein, the Representing the joint loss function of the joint, Indicating a loss of positioning and, Indicating that the loss is to be detected, Indicating a loss of consistency of the terrain, 、 And Representing a loss weight; the calculation formula of the positioning loss is as follows: Wherein, the Represents a land location truth value tag, Representing element-wise multiplication i.e. multiplication of elements in the same position of two matrices, Representing a land probability positioning mask; The calculation formula for detecting loss is: Wherein, the Representing the probability that a certain sample is predicted to be an illicit change, Representing the weight coefficient, for balancing the positive and negative samples, Representing a modulation factor for reducing the weight of the easily classified samples; the calculation formula of the topography consistency loss is as follows: Wherein, the Representing the adjustment factor of the gradient of the terrain, The gradient operator is represented by a gradient operator, A gradient map showing the gradient of the i-th sample, Representing the L2 norm, i.e. calculating its euclidean norm for the gradient difference vector for each pixel location, Representing the number of samples to be taken, A gradient map of the land mask representing the i-th sample; In the step S3, the newly added land illegal change detection model specifically includes: Simultaneously inputting a satellite remote sensing image of a time phase t and a satellite remote sensing image of a time phase t+k, wherein the time phase interval k is 30 to 90 days, and the two time phase images have the same space size and the same wave band number; Respectively processing two time phase images through a double-branch convolutional neural network with identical parameters to generate two space-spectrum characteristic diagrams with identical dimensions; expanding the space-spectrum characteristic diagram into a characteristic vector sequence according to the space position, modeling a time sequence nonlinear relation through a three-layer LSTM network, and independently processing each time phase to generate a time sequence enhancement characteristic; Splicing time sequence enhancement features generated by the double-branch LSTM network according to channel dimensions, and extracting change features through a nonlinear transformation layer; and inputting the fusion characteristic into a three-layer full-connection projection network, outputting a central pixel change probability value, and judging that the pixel is illegally changed when the central pixel change probability value is greater than 0.7.
2. The method of claim 1, wherein the extracting topographical features through two convolution layers comprises a first convolution layer of 7 x 7 convolutions, a number of output channels of 64, a ReLU activation function, a second convolution layer of 3 x 3 convolutions, a number of output channels of 256, a ReLU activation function; The method for generating the land probability positioning mask by adopting the feature attention fusion mechanism specifically comprises the steps of adjusting the channel number of the topographic features through a 1X 1 convolution layer to enable the channel number of the topographic features to be the same as that of the spectral image features, generating an attention weight map through a softmax function, carrying out element-by-element multiplication on the attention weight map and the spectral image features to carry out feature fusion, and generating the land probability positioning mask through the fused features through a 1X 1 convolution layer and a Sigmoid activation function.
3. The method of claim 2, wherein the convolutional neural network comprises a 5-layer convolutional structure, the specific hierarchy configured to: layer 1, 5×5 convolution kernel, output 64 channels, step size 1, relu activation; layer 2, 3×3 convolution kernel, output 128 channels, step size 2, relu active; layer 3, 3×3 convolution kernel, output 256 channels, step size 2, relu activation; Layer 4, 3×3 convolution kernel, output 512 channels, step size 1, relu activation; Layer 53 x 3 convolution kernel, output 512 channels, step size 1, relu active.
4. The method of claim 3, wherein a specific hierarchy of the three-tier LSTM network is configured to: The first layer LSTM is used for processing the input characteristic vectors by 64 hidden units; 128 hidden units of the second layer LSTM, which receives the output of the first layer; 128 hidden units of the third layer LSTM, which receives the output of the second layer; Wherein LSTM network parameters are randomly initialized, the learning rate is set to be 0.0001, and the loss function is optimized by using an Adam optimizer.
5. The method of claim 4, wherein a specific hierarchy of the three-tier fully connected projection network is configured to: a first full connection layer, 64 input units, 32 output units, reLU active; a second full connection layer, 32 input units, 16 output units, reLU activation; and the third full connection layer is 16 input units, 1 output unit and Sigmoid activation.
6. The method according to claim 5, wherein in the step S4, the generation of the land violation probability map includes the steps of: arranging pixel-level change probability values output by the newly-added land illegal change detection model according to the original image space positions to form a two-dimensional probability matrix, wherein each matrix element corresponds to the change probability value of the geographic coordinate position; Binding the probability matrix with a geographic coordinate system, performing geographic reference by adopting a geodetic coordinate system, and keeping the same spatial resolution and projection parameters as those of the original satellite image; the land violation visualization scheme is set, and specifically comprises the following steps: The central pixel change probability value of the non-change area is 0.0-0.7, and green rendering is used for representing natural change or legal land; the central pixel of the change area has a change probability value of 0.7-1.0, and red rendering is used for representing the probability illegal land.
7. The method of claim 6, wherein the generating of the illicit plausible region vector boundaries comprises the steps of: setting 0.7 as a probability threshold, marking the pixels which are more than or equal to 0.7 in the land illegal probability map as 1, and marking the rest pixels as 0; sequentially executing a closing operation and an opening operation, eliminating tiny noise and filling holes so that the boundary of an illegal area is continuous and complete; extracting closed polygon boundaries of all the connected areas, and ensuring the correct topological structure; Converting the boundary polygons into vector files and assigning unique identifiers to each polygon; Calculating and storing key attributes, wherein the key attributes comprise polygon area, average probability value, maximum probability value, minimum probability value, perimeter and center point coordinates; Simplifying the boundary by using a Fabry-Perot algorithm, reserving key turning points, and setting the tolerance to be 0.5 m; fine polygons with a filtering area of less than 10 square meters; and generating a filtered illegal land suspected area vector boundary.

Description

AI-based karst mountain land illegal cue intelligent discovery and supervision method Technical Field The invention relates to the technical field of land monitoring, in particular to an AI-based karst mountain land illegal cue intelligent discovery and supervision method. Background With the rapid development of town, land supply cannot meet the requirement of urban scale expansion, land value is higher and higher, and great challenges are brought to land law enforcement and supervision work. The method is an important ring for land law enforcement and supervision work, and is used for dynamically monitoring land illegal behaviors in real time and timely checking and disposing the land illegal behaviors. The traditional manual monitoring has the problems of low efficiency, high omission factor and the like, and cannot meet the requirement of large-scale rapid monitoring. The automatic interpretation based on the remote sensing image can be based on timeliness and wide coverage of the remote sensing image, and the change pattern spots in the monitoring area can be rapidly and automatically extracted. At present, a certain technological breakthrough is made in the field in various plain areas in China, but characteristics of land feature textures, shadows and the like caused by relief of topography are complex due to the fact that mountain areas are numerous and ground features are broken, recognition difficulty is increased, resolution of remote sensing images, spectral characteristics and other factors also influence interpretation accuracy, and therefore the existing automatic interpretation technology is difficult to play a role in the karst mountain areas. The current land illegal activity supervision is managed in a report mode, the informatization degree is low, the data verification is delayed, and the cross-level coordination is low. Most geographic information systems are built by purchasing professional software services, so that the cost is high, the customization degree is low, and the geographic information systems are difficult to adapt to application scenes with changeable requirements. The Chinese patent with the grant of CN112270291B discloses an automatic detection method for illegal construction land development based on multi-source optical remote sensing images, which comprises the steps of respectively obtaining two optical remote sensing images of a region to be detected at different time, calculating land leveling intensity between the two optical remote sensing images at different time to obtain a land leveling intensity image, carrying out cluster analysis on the land leveling intensity image by utilizing a maximum expected algorithm, extracting a construction land development corresponding region of the region to be detected, and comparing the extracted construction land development corresponding region with a land plan of the region to be detected, wherein the construction land development in the strictly forbidden development region of the land plan is illegal construction land development. The method can carry out short-period and full-automatic monitoring on the development of the illegal construction land, is simple in calculation and easy to understand, has high execution efficiency, does not need any training sample, can timely find and prevent various social and environmental problems caused by the illegal construction land, obviously reduces the monitoring cost and improves the monitoring efficiency. The invention discloses a land block violation monitoring system based on satellite images and unmanned aerial vehicle inspection, which comprises a first image acquisition module, a second image acquisition module, a data storage module, a data processing module and a data output module, wherein the data processing module is used for retrieving historical data, constructing a monitoring area data model, receiving real-time overall image data, acquiring abnormal position information and generating an abnormal report, the data processing module is also used for reading real-time local image information of an abnormal position and retrieving the historical data to perform parameter analysis on illegal buildings at the abnormal position so as to estimate land loss caused by the illegal buildings. The problems of the background technology are that the difficulty in identifying land remote sensing illegal due to complex characteristic expression in the karst mountain area is increased, the current pattern spot supervision is managed in a report mode, the informatization degree is low, the data verification is delayed, and the cross-level coordination is low. Disclosure of Invention According to the method, the characteristics of land violations in the karst mountain areas in Guizhou are taken as research objects, a method suitable for automatically changing and detecting newly added land violations clues in the karst mountain areas is researched, a supervision system based on a fu