CN-122024085-A - Improved degraded grassland restoration effect evaluation method and long-term dynamic monitoring system

CN122024085ACN 122024085 ACN122024085 ACN 122024085ACN-122024085-A

Abstract

The invention belongs to the field of degraded grassland restoration evaluation research, and discloses a degraded grassland restoration evaluation research method based on multi-element remote sensing data, which comprises the following steps of observing by adopting a FragMap unmanned aerial vehicle aerial photography analysis system and preprocessing by adopting Pixel Based Manual Classifier software; collecting grassland biomass data, acquiring and preprocessing remote sensing data, constructing a grassland coverage and overground biomass remote sensing estimation model, performing accuracy verification, analyzing grassland coverage and biomass spatial differences before and after the degraded grassland is repaired, and analyzing the grassland coverage and biomass space-time dynamic change. According to the method, a high space-time resolution NDVI data set of a research area is built by combining a high space-time remote sensing data fusion algorithm, a grassland coverage and aboveground biomass remote sensing estimation model is built by combining ground investigation and unmanned aerial vehicle aerial photography, the grassland coverage and aboveground biomass change conditions before and after the degraded grassland restoration measures are implemented are evaluated by the system, and the restoration effect of different restoration measures on the degraded grasslands is evaluated.

Inventors

MENG BAOPING
YAO CHUANG
YI SHUHUA
LI XINYAN
JIANG ZHUOJUN
WANG XIAOWU
QIN YU
SHI YAFEI
XU WENBO
JI GUOHUI

Assignees

甘肃农业大学

Dates

Publication Date: 20260512
Application Date: 20260120

Claims (10)

1. An improved evaluation method for the restoration effect of degraded grassland is used for carrying out high-precision quantitative evaluation on the restoration process of grassland degradation in the range of satellite image pixels in a target area, and is characterized by comprising the following steps: s1, acquiring a vertical image by adopting an unmanned aerial vehicle carrying an RGB+near infrared dual camera or a multispectral sensor according to a grid route, and performing illumination normalization and shadow compensation processing on the image; S2, inputting the normalized image into a deep learning semantic segmentation network for vegetation pixel identification, wherein the semantic segmentation network adopts a coding-decoding structure and performs combined training by using a remote sensing sample set and measured sample side data to output high-precision grassland coverage; S3, distributing a fixed area sample square on the sample land, mowing the living things on the ground, and drying and weighing to obtain a biomass observation value; s4, calculating the grassland biomass value in the range of the satellite image pixel corresponding to the ground sampling sample by combining the grassland vegetation and bare soil plaque space distribution information obtained in the S2; S5, acquiring multi-time phase satellite vegetation index data, screening out low-quality images based on cloud cover, and generating a vegetation index continuous sequence with time resolution not lower than 16 days and spatial resolution not lower than 30m by adopting a space-time fusion algorithm; S6, using grassland coverage and biomass as dependent variables, using vegetation indexes and time sequence characteristics thereof as independent variables to establish a remote sensing estimation model based on a machine learning algorithm, and adopting a leave-one-out cross validation and Bayesian optimization to select an optimal remote sensing inversion model; S7, calculating a grassland coverage and biomass data set with high space-time resolution in the grassland growing season according to the optimal model and the fusion data set, and calculating the annual maximum value and the annual average value of the grassland coverage and biomass data set; S8, judging the degradation recovery rate and the change trend thereof by using the annual sequence grassland coverage and biomass maximum value and combining trend analysis and significance test.
2. The method of claim 1, wherein the semantic segmentation network selects one of U-Net, segFormer or DeepLabV +, and adds landscape texture features to the training process to improve degraded grassland sparse vegetation identification accuracy.
3. The method of claim 1, wherein the illumination normalization employs Retinex enhancement or HSV brightness normalization strategies, and incorporates zone-adaptive shadow compensation to improve vegetation segmentation consistency under different voyage, weather conditions.
4. The method of claim 1, wherein the sample area is 0.5m x 0.5m, the number of samples is not less than 84, and the samples are used for constructing a training set and a verification set to achieve vegetation coverage and biomass calibration.
5. The method of claim 1, wherein the regression inversion model uses one of a random forest, LSTM, or multiple regression model and uses correlation coefficients, root mean square error, and generalization error as optimal indices.
6. The method of claim 1, wherein the recovery trend significance test employs a Mann-Kendall and Sen slope joint evaluation strategy and can output significant recovery zone, stable zone and degradation zone classification results.
7. A deep learning based degraded grassland long-term dynamic monitoring system implementing the improved degraded grassland restoration effect evaluation method according to any one of claims 1 to 6 for generating a cross-year continuous grassland degradation restoration monitoring result, characterized by comprising: A) The unmanned aerial vehicle image acquisition module is used for acquiring multispectral or RGB+NIR images according to grid airlines and executing illumination normalization; B) The semantic segmentation module is used for classifying vegetation and bare land based on the deep learning model and generating a high-precision coverage grid; C) The ground sample side measuring module is used for obtaining the dry weight of the ground and constructing a calibration data set; d) The satellite time sequence fusion module is used for generating a continuous vegetation index sequence of 16 days; E) The regression inversion module is used for inverting coverage and biomass from the time sequence vegetation index and outputting a annual maximum value and an annual average value; f) And the change trend analysis module is used for calculating the slope of the multi-year sequence, evaluating the significance and forming a recovery level partition result graph.
8. The system of claim 7, wherein the semantic segmentation module employs a convolutional encoder fused with a Transformer structure to promote weak vegetation identification.
9. The system of claim 7, wherein the satellite timing fusion module outputs results that directly drive the regression inversion module to implement an automated coverage and biomass estimation link.
10. The system of claim 7, wherein the trend analysis module inverts the spatial distribution of the recovery area and generates a map output based on both the annual maximum vegetation index and the annual average coverage index.

Description

Improved degraded grassland restoration effect evaluation method and long-term dynamic monitoring system Technical Field The invention belongs to the technical field of evaluation and research of degraded grassland restoration effect, and particularly relates to an improved evaluation method of degraded grassland restoration effect and a long-term dynamic inversion system. Background The patents CN114708504A, CN104778451A and CN118781486A disclosed in the prior art can show that the current grassland vegetation coverage or aboveground biomass remote sensing inversion technology still takes 'high-resolution remote sensing in satellites + a small amount of ground sample sides + statistics or machine learning models' as a core technical route, and coverage or biomass estimation of a regional scale can be theoretically realized, but systematic deficiency still exists in the complex application scene of degraded grassland monitoring. Firstly, in terms of space scale and heterogeneity expression, the prior art generally depends on a meter-level satellite image element or even a ten meter-level satellite image element, and an inversion result is essentially 'average state description' on the image element scale, and the prior art lacks sufficient recognition capability for plaque distribution, fragmented weak vegetation and bare soil mixed patterns widely existing in degraded grasslands. Whether CN114708504A with the coverage is inverted only, CN104778451A with the height factor is introduced, or CN118781486A with random forests is adopted, small-scale vegetation structural differences are difficult to describe, and error amplification and insufficient stability are caused in the low coverage or initial recovery stage. In the aspect of a data fusion and calibration mechanism, although weather, DEM, topography or multi-time phase remote sensing is introduced in the existing scheme, the problem of obvious space mismatch commonly exists between a ground actual measurement sample party and satellite pixels, a hierarchy mapping and synchronous calibration mechanism of 'sample party-high-resolution image-satellite image' is lacked, and model training is easily influenced by scale deviation, so that generalization and popularization capability of the model training is restricted. In the aspects of illumination conditions and complex imaging interference, the prior art is insufficient in consideration of strong plateau illumination change, shadows, soil humidity difference and weak vegetation reflection characteristics, lacks systematic illumination normalization, shadow compensation and texture structure correction means, and forms restriction on degraded grassland and inversion reliability at the beginning and the end of a growing season. Finally, in the aspect of monitoring targets and application, the existing patents are concentrated on single index (coverage or biomass) and discrete year estimation, so that it is difficult to construct a continuous closed loop system from ground actual measurement to high-resolution image and then to multi-time phase satellite data, and it is difficult to form stable and reliable annual change rate, recovery trend slope and significance test results, and the actual requirements of long-term, quantitative and dynamic evaluation of the degraded grassland recovery process cannot be met. Therefore, the prior art has not effectively solved key problems such as weak vegetation identification, complex space heterogeneity depiction, illumination interference suppression, long-term trend inversion and the like, and also forms a real technical foundation and innovation space for introducing a high-resolution unmanned aerial vehicle image, depth semantic segmentation, illumination normalization and multisource closed loop fusion analysis scheme. Disclosure of Invention Aiming at the problems existing in the prior art, the degraded grassland is one of the most typical types of ecological degradation, and the recovery process has slowness, instability and space heterogeneity and is often interfered by factors such as artificial grazing, precipitation fluctuation, soil depletion and the like. The existing degraded grassland restoration monitoring is characterized in that an unmanned aerial vehicle RGB image is used for extracting green index EGI and other algorithms to distinguish vegetation from bare soil, and then the grassland coverage is generated through threshold segmentation, but the method is sensitive to illumination change and is easily interfered by factors such as shadow, bare soil background, seasonal dry period color deviation and the like, so that the low coverage and broken-spot vegetation identification error is obvious, the risk of underestimation of the coverage and false noise identification exists, meanwhile, the degradation restoration grassland ecological system also has the problems of unstable system structure, frequent secondary degradation and the like, and the traditio