CN117611993A - A method for estimating vegetation classification based on actual evapotranspiration from remote sensing - Google Patents

Abstract

The invention discloses a method for estimating vegetation classification based on remote sensing actual evapotranspiration, which comprises the steps of collecting hydrology, weather, topography, underlying surface and remote sensing image data in a research flow field; constructing an SEBAL actual evapotranspiration estimation model; calculating the actual evaporation of the research river basin on a multidimensional time scale; generating random sample points, identifying the land coverage types of the sample points by using a visual interpretation method, and editing a sample point attribute table; establishing a classification system and a feature input database; and constructing a classifier. The invention considers the evapotranspiration space-time diversity characteristics of different vegetation types, digs out the influence of vegetation distribution on environmental factors, namely vegetation distribution difference information contained in the evapotranspiration space difference, estimates continuous multi-period instantaneous evapotranspiration by constructing an actual evapotranspiration calculation model based on an energy balance principle and introduces the estimated continuous multi-period instantaneous evapotranspiration into a characteristic input database, thereby improving the classification precision of land coverage, in particular the vegetation classification precision of vegetation mixed regions.

Description

A method to estimate vegetation classification based on actual evapotranspiration from remote sensing

Technical field

The invention relates to the field of remote sensing technology, and specifically relates to a method for estimating vegetation classification based on actual evapotranspiration from remote sensing.

Background technique

Land use/land cover maps are basic data support for scientific research such as ecological environment, hydrological simulation, geographical monitoring, and social economics. Land cover types such as water bodies, cities, and unused land have relatively obvious spectral characteristics and clear texture characteristics, and are easy to identify on satellite images. However, the spectral characteristics of different vegetation types are similar and the spatial distribution is staggered. There is a certain degree of ambiguity and transition between different types, making it difficult to accurately distinguish them during remote sensing classification. Vegetation is one of the main components of land cover and accounts for a large proportion of land surface cover. How to innovate or improve existing methods to improve vegetation classification accuracy is still a research difficulty in the field of remote sensing.

There are many methods for interpreting land cover through remote sensing, among which supervised classification has always been considered a very effective land cover classification method. At present, the main methods to eliminate or reduce the influence of mixed pixels and thereby improve the accuracy of vegetation classification include: fusion mixed pixel decomposition algorithm, classification method based on vegetation index, classification method based on multi-temporal information, classification method with additional auxiliary information, etc. For example, existing literature 1: Zhang Xiaoyu et al. Forest vegetation classification of Landsat-8 remote sensing images based on random forest model [J], Journal of Northeast Forestry University, 2016, 44(6), using three vegetation indices to distinguish forest land and non-forest land, The characteristics of woodland are relatively more obvious and easy to distinguish, but for other types, such as cultivated land and grassland, it is difficult to distinguish using only vegetation index. Existing document 2: Li W. et al. Integrating Google Earth imagery with Landsatdata to improve 30-mresolution land cover mapping [J]. Remote Sensing of Environment, 2020, 237: 111563. (The SCI doi number of this document is /10.1016/ j.rse.2019.111563) uses vegetation index to improve vegetation classification accuracy, but vegetation index is more suitable for distinguishing vegetation from non-vegetation, and the identification effect of different vegetation types is poor. Existing document 3: Sun W. et al. in the document Spatiotemporal vegetation covervariations associated with climate change and ecological restoration in the Loess Plateau[J].Agricultural and Forest Meteorology, 209:87-99. (The SCI doi number of this document is 10.1016/j .agrformet.2015.05.002) mainly adds some environmental factors that affect vegetation distribution, mainly topography, meteorology, and socioeconomic indicators, as additional inputs to improve classification accuracy based on spectral band information of remote sensing images. However, different environmental factors The vegetation may be the same, and the vegetation may be different under the same environmental factors.

Existing literature 4: Sun Na et al. Research on land cover classification method combined with vegetation coverage index [J]. Journal of Beijing Normal University, 2022, 58(6): 917-924, taking Sindh Province of Pakistan as the research area, using The FVC change curve is used to extract vegetation phenological characteristics, and the differences in vegetation phenology are used as the basis for classification. From Figure 2 of the article, we can see that the vegetation area is small, mostly bare land and cultivated land. It can be seen from Figure 4 of the article. , the vegetation coverage change curves of cultivated land and grassland are very different. Existing document 5: Liu Zhen et al. Analysis of spatiotemporal changes in seasonal vegetation coverage in Guanzhong area from 2003 to 2014 [J]. Geospatial Information, 2018, 16(05), through Figure 3 (I-11) of this existing document 5 Corresponding to grassland, I-12 corresponds to cultivated land, and I-15 corresponds to woodland). It can be seen that the vegetation coverage changes of cultivated land and grassland are relatively close. The results obtained by existing literature 5 and existing literature 4 are different. Therefore, Vegetation coverage FVC is essentially a vegetation index that can be used to distinguish between vegetation and non-vegetation areas, but it is difficult to distinguish between different vegetation types.

Contents of the invention

The purpose of the present invention is to overcome the problem that it is difficult to accurately identify and distinguish vegetation types with relatively close spectral characteristics in the existing methods of interpreting land cover from remote sensing images, and there are errors in judging vegetation types based only on environmental factors and vegetation coverage, and provide a method This method of estimating vegetation classification based on actual evapotranspiration from remote sensing uses the spatial difference in evapotranspiration of different vegetation types as feature input to improve the identification accuracy of different vegetation.

The present invention adopts the following technical solutions:

A method for estimating vegetation classification based on actual evapotranspiration from remote sensing, including the following steps:

Step 1: Collect hydrology, meteorology, topography, underlying surface and remote sensing image data in the study watershed. The site data are organized in time series, and the spatial data are cropped according to the vector range of the study watershed and resampled to the same resolution;

Step 2: Based on the data collected in Step 1, build a SEBAL actual evapotranspiration estimation model suitable for the study watershed;

Step 3: Use the SEBAL actual evapotranspiration model built in step 2 to calculate the actual evapotranspiration of the study watershed on a multi-dimensional time scale;

Step 4: Generate random sample points, use visual interpretation method to identify the land cover type of these sample points, edit the sample point attribute table; and divide it into training samples and test samples according to the ratio of 7:3;

Step 5: Establish a classification system and feature input database, reasonably set up the land cover type classification of the study watershed, and organize a database that uses five types of information such as spectral band, spectral index, topographic factors, nighttime lights, and evapotranspiration as feature inputs;

Step 6: Construct a classifier, select an appropriate machine learning algorithm as the classifier, substitute the extracted values of the sample points in the training sample in step 4 on each feature input, and train and optimize the classifier; use the sample points in the test sample in step 4 for verification. By comparing the results calculated by the classifier with the actual land cover types at the sample points in the test samples obtained through visual interpretation in step 4, the classification accuracy can be obtained and its accuracy can be evaluated.

Furthermore, the main calculation of the SEBAL model in step two is based on the energy balance principle:

λE＝ _Rn -GH

In the formula, λ is the latent heat of vaporization, E is the actual evapotranspiration, R _n is the net radiation flux, G is the soil heat flux, and H is the sensible heat flux.

Preferably, the land cover types described in steps four and five include cultivated land, forest land, grassland, urban land, water bodies and unused land.

Preferably, the characteristics described in step five are input into the database, specifically including the 6-band spectral data of Landsat8, namely B2 blue light band, B3 green light band, B4 red light band, B5 near-infrared band, B6 short-wave infrared band 1 and B7 short-wave Infrared band 2, 3 spectral indices, namely NDVI, MNDWI and NDBI, a nighttime light index, 2 terrain factors of elevation and slope, 21 instantaneous evapotranspiration, a total of 33 features.

Preferably, the three spectral index calculation formulas of Landsat8 are as follows:

NDVI＝(B5-B4)/(B5+B4)

MNDWI＝(B3-B6)/(B3+B6)

NDBI=(B6-B5)/(B6+B5).

Preferably, the objective function used to evaluate the classification accuracy in step six includes but is not limited to user accuracy UA, producer accuracy PA, F1 score, overall accuracy OA and Kappa coefficient.

Furthermore, the calculation formulas of the objective function user accuracy UA, producer accuracy PA, F1 score, overall accuracy OA and Kappa coefficient are as follows:

PA＝nn _i /yy _i

Kappa＝(OA-Pe)/(1-Pe)

In the formula, nn _i is the number of pixels in the entire image that are correctly classified into the i-th category, xx _i and yy _i are respectively the total number of pixels in the entire image that are classified into the i-th category by the classifier and the total number of true reference images in the i-th category. The element number, Pe, is the sum of the products of the number of real pixels of various types and the number of predicted samples, divided by the square of the total number of samples.

Preferably, the hydrology, meteorology, topography, underlying surface and remote sensing image data collected in step one specifically include: runoff data from hydrological stations in the study basin, meteorological data from meteorological stations inside and around the basin, and digital elevation. Data sets, vegetation index, surface temperature, surface reflectance data set, terrestrial water storage data set, remote sensing images and nighttime light data sets.

The beneficial effects of the present invention are:

It is difficult to accurately identify and distinguish vegetation types with relatively close spectral characteristics and mixed spatial distribution in existing literature 1 and 2. The present invention considers the spatiotemporal differentiation characteristics of evapotranspiration of different vegetation types, estimates the instantaneous evapotranspiration for multiple consecutive periods by constructing an actual evapotranspiration calculation model based on the energy balance principle, and introduces it into the feature input database, thereby improving the classification accuracy of land cover. , especially the vegetation classification accuracy in mixed vegetation areas.

Existing literature 3 has errors in judging vegetation type only based on environmental factors, because vegetation may be the same under different environmental factors, and vegetation may be different under the same environmental factors. The vegetation phenological characteristics of cultivated land and grassland in existing literature 4 and 5 overlap to a certain extent and are difficult to distinguish. The present invention explores the impact of vegetation distribution on environmental factors, and uses the effect of vegetation on hydrology (the feedback effect of vegetation growth on atmospheric moisture - evapotranspiration) as the basis for classification. The amount of evapotranspiration over time on different vegetation types is and change patterns, that is, the spatial differences in evapotranspiration contain information on vegetation distribution differences. Generally speaking, woodland is the area with the largest evapotranspiration. For the two types of cultivated land and grassland, which are difficult to distinguish, the evapotranspiration of cultivated land will be greater than that of grassland during the crop growth and development period, and the evapotranspiration of grassland will be greater than that of cultivated land during the crop withering period. Through the year-round The amount and change pattern of instantaneous evapotranspiration can well distinguish grassland and cultivated areas, and is applicable to all research types.

Through theory and examples, the present invention demonstrates that introducing evapotranspiration as a feature input improves the accuracy of vegetation classification, and has certain theoretical value and practical significance for realizing and improving the large-scale rapid identification of multiple types of vegetation in mixed vegetation areas or highly heterogeneous areas. .

Description of drawings

Figure 1 shows the geographical location of the Ordos Basin in Embodiment 1;

Figure 2 shows the multi-period instantaneous evapotranspiration results of the Ordos Basin simulated by the SEBAL model in Example 1;

Figure 3 shows the SEBAL simulation results based on water balance verification at the sub-basin scale in Example 1;

Figure 4 is a diagram of the land cover classification results based on actual evapotranspiration estimation in Embodiment 1;

Figure 5 is a histogram of accuracy comparison under different input schemes in Example 1;

Figure 6 shows the comparison between the land cover classification results based on actual evapotranspiration estimation and existing products in Example 1;

Figure 7 shows the feature input difference and importance in Embodiment 1;

Figure 8 shows the difference in instantaneous evapotranspiration under three vegetation types in Example 1.

Detailed ways

The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

Example 1

This embodiment takes the Ordos Basin as an example. The geographical location of the Ordos Basin is shown in Figure 1. This area is located in the upper and middle reaches of the Yellow River Basin, covering an area of approximately 380,000 square kilometers. Vegetation types in the Ordos Basin have a zonal distribution pattern from southeast to northwest. Affected by topography, the regional vegetation landscape pattern shows high spatial heterogeneity of mixed trees, shrubs and grasses. Due to the natural attributes of the Ordos Basin, the terrain is fragmented and vegetation site conditions have high spatial heterogeneity. Coupled with human policy intervention such as ecological engineering construction, existing remote sensing inversion products are difficult to accurately reflect the vegetation spatial pattern under the mixed pattern of trees, shrubs and grasses.

This embodiment is based on the MODIS vegetation index, surface temperature, and surface reflectivity data sets provided by NASA, the GDEMV2 digital elevation data set provided by the Japanese Space Agency JAXA, and the daily meteorological data (temperature, wind speed, and precipitation) provided by the China Meteorological Network. The SEBAL model was constructed and verified using the hydrological station runoff data provided by the China Hydrological Yearbook and the GRACE terrestrial water storage data set provided by the Space Research Center CSR. Using the Sentinel-2 and Landsat8 remote sensing images provided by the global-scale remote sensing cloud computing platform GEE, the VIIRS nighttime light data set provided by the National Oceanic and Atmospheric Administration NOAA, as well as digital elevation data and SEBAL model calculation results, the Ordos Basin land in 2020 Coverage for classification. Compare the classification accuracy under different input schemes, and download existing land cover/land use products to compare the classification accuracy. There are 6 categories of products used for comparison here, including: CNLUCC products (1000m) provided by the Institute of Geographic Sciences and Natural Resources of the Chinese Academy of Sciences, MCD12Q1 products of MODIS (500m), Globe Land products (30m) provided by the National Basic Geographic Information Center, CLCD products (30m) and CRLC products (10m) provided by Wuhan University, and GLC_FCS products (30m) provided by the Institute of Aerospace Information Innovation, Chinese Academy of Sciences. To illustrate that the vegetation classification accuracy improvement method proposed by the present invention based on remote sensing actual evapotranspiration estimation has higher accuracy.

The specific method of this embodiment is:

Step 1: Download MODIS vegetation index MOD13Q1, surface temperature MOD11A2, surface reflectance MCD43A3, GDEMV2 digital elevation data, GRACE terrestrial water storage data, Sentinel-2 images, Landsat8 remote sensing images, VIIRS nighttime light data set, download the meteorology inside and around the basin The meteorological data of the site, including wind speed, temperature, precipitation, etc., are extracted from the hydrological station runoff data provided by the China Hydrological Yearbook. Air temperature and wind speed are meteorological inputs calculated by the SEBAL model, and precipitation and runoff are elements needed to verify the evapotranspiration results calculated by SEBAL using the water balance method. Site data are organized in time series, and spatial data are clipped according to the vector boundaries of the Ordos Basin and resampled to the same resolution.

Step 2: Select the SEBAL actual evapotranspiration estimation model, use the data collected and processed in step 1 to invert surface characteristic parameters such as surface clear sky albedo, surface specific emissivity, surface temperature, etc., and calculate the satellite transit time in single pixel units The instantaneous net radiation flux and soil heat flux are calculated, and the aerodynamic impedance is iteratively corrected by Monin-Obukhov to obtain the sensible heat flux in the stable area. Finally, a SEBAL actual evapotranspiration estimation model suitable for the Ordos Basin is built based on the energy balance equation. .

Step 3: Use the actual evapotranspiration estimation model built in step 2. First, simulate the instantaneous evapotranspiration in the Ordos Basin.

Figure 2 shows the instantaneous evapotranspiration results of the Ordos Basin simulated by the SEBAL model for multiple periods in 2020. Instantaneous evapotranspiration is the actual evapotranspiration. The instantaneous evapotranspiration throughout the year generally shows a trend of first increasing and then decreasing over time, and the instantaneous evapotranspiration from April to September is relatively large. The instantaneous evapotranspiration does not exceed 0.4mm/h from January to March, the average evapotranspiration from April to September is 0.5-0.7mm/h, and can reach more than 1mm/h in some areas, and gradually decreases from October to December. Spatially, instantaneous evapotranspiration decreases from southeast to northwest, and evapotranspiration in the southern loess area is generally greater than in the northern windy sand area.

Secondly, since the calculation of instantaneous evapotranspiration lacks actual measured data for verification, the instantaneous evapotranspiration is scaled to calculate daily evapotranspiration, which is then statistically calculated as monthly evapotranspiration and annual evapotranspiration. The statistics are calculated on a monthly and annual basis for verification, to verify whether the calculated instantaneous evapotranspiration is accurate. If the monthly and annual evapotranspiration are accurate, it means that the instantaneous evapotranspiration is also accurate.

Finally, select multiple small watersheds in the study area, use precipitation, runoff, water storage, etc. above the control station to construct a water balance equation to calculate the theoretical true value of evapotranspiration, and compare it with the monthly evapotranspiration and annual evapotranspiration values calculated previously. Compare and analyze to verify the accuracy of the simulation results, calculate the objective function correlation coefficient r and relative error RE, the formula is as follows:

RE＝(ET _sim -ET _real )/ET _real

In the formula, are the theoretical true value and model simulation value of evapotranspiration in the jth period respectively,/> are the average values of the theoretical true value and model simulation value of evapotranspiration respectively.

The reason for choosing a small watershed is that there is currently no accurate observational data on evapotranspiration. The water balance method mentioned in this article is mostly used, which is based on the watershed scale and uses measured precipitation and runoff data observed in the watershed. Because the measured data of runoff are only available at the outlet of each basin.

Figure 3 shows the verification results of the SEBAL model at the sub-basin scale. (a) is the geographical location of the sub-basin and the control station, (b) is the scatter plot of the theoretical true and simulated values of evapotranspiration at the monthly scale, and (c) is Scatter plot of theoretical true and simulated values of evapotranspiration on an annual scale. (b) and (c) in Figure 3 are scatter plots. This image can be interpreted as: for each point in the figure, the abscissa corresponding to the point represents the theoretical true value, and the corresponding ordinate represents the simulated value. (d) in Figure 3 shows the annual evapotranspiration and relative error distribution of each sub-basin (where Esim represents the simulated value and Eobs represents the theoretical true value). The 44 sub-basins basically cover the range of different regions with different characteristics in the Ordos Basin, and its verification results can represent the simulation accuracy of the entire region. On the monthly scale, the correlation coefficient r between the theoretical true value of evapotranspiration and the simulated evapotranspiration value of the SEBAL model is 0.63 on average across all sub-basins, and the correlation coefficient is 0.78 on the annual scale. In each sub-basin, the correlation coefficient between the theoretical true value of evapotranspiration and the simulated evapotranspiration value of the SEBAL model is mostly above 0.7, except for individual basins, and in terms of magnitude, the evapotranspiration of each sub-basin in 2020 The hair is basically within 350mm-700mm. There is not much difference between the actual evapotranspiration and the simulated evapotranspiration in the sub-basin in 2020, and most of the errors are within 10%. Therefore, the multi-scale and multi-basin verification results show that the evapotranspiration simulated by the SEBAL model is reliable.

Step 4: Randomly select sample points in the Ordos Basin based on the 2m high-resolution Sentinel-2 image based on the stratified sampling method, use visual interpretation methods to identify land cover types, and edit the sample point attribute table. A total of 555 are selected here. There are 121, 107, 126, 63, 61, and 77 sample points in cultivated land, forest land, grassland, urban land, water bodies, and unused land respectively. And divided into training samples and test samples according to the ratio of 7:3;

Step 5: Taking into account the distribution of the main underlying surfaces in the Ordos Basin, set up a six-category classification system for cultivated land, forest land, grassland, urban land, water bodies, and unused land. Construct a feature input database, including five types of information: spectral band, spectral index, terrain factors, night lights, and evapotranspiration. Specifically, it includes 6 bands of spectral data of Landsat8 (B2 blue band, B3 green band, B4 red band, B5 near infrared band, B6 shortwave infrared band 1, B7 shortwave infrared band 2), 3 spectral indices (NDVI, MNDWI , NDBI), 2 terrain factors (elevation and slope), 1 night light index, 21 instantaneous evapotranspiration, a total of 33 features. For Landsat8, the calculation formula of spectral index is as follows:

NDVI＝(B5-B4)/(B5+B4)

MNDWI＝(B3-B6)/(B3+B6)

NDBI＝(B6-B5)/(B6+B5)

The calculation process of evapotranspiration requires the use of aerodynamic methods for iterative calculations. However, due to the influence of weather and clouds, a converged solution cannot be obtained for the instantaneous evapotranspiration for two days (4-26 and 8-28 respectively), as shown in Figure 2 The results of these two days are shown as interpolation results. In essence, no new information enters. In practical applications, only 21 instantaneous evapotranspirations are used to obtain convergence results.

The feature input database constructed here is not only for 555 samples, but for the entire area. That is, these 33 feature inputs, each input has specific data at each position in the entire area. Only when training the classifier in step 6, the values of these feature inputs need to be extracted to the sample points. Finally, when applying the classifier to the entire region, all feature inputs of the entire region need to be used.

Step 6: Construct a classifier, select the random forest algorithm as the classifier, use the sample point data generated in step 4 and the feature input database established in step 5, extract the feature input value at the sample point position and use it as the input of the classifier for training. Specifically, it is the extracted value of each sample point on 33 feature inputs. For 70% of the sample points, the amount of training data for a total of 388 training sample points is (388*33=12804).

Use the remaining 30% of sample points (555*0.3=167 test samples) for verification and accuracy evaluation. These 30% of sample points are not used to train the classifier, so they can be used for accuracy verification. At this time, the land cover classification results of the remaining 30% of the sample points in the area are calculated through the classifier. The results calculated by the classifier are compared with the real land cover types of the 30% sample points obtained through visual interpretation in step 4. By comparison, the classification accuracy can be obtained. To evaluate its accuracy, the objective functions used include user accuracy (UA), producer accuracy (PA), F1 score, overall accuracy (OA), Kappa coefficient, etc. The calculation formula is as follows:

UA＝nn _i /xx _i ,PA＝nn _i /yy _i

Kappa＝(OA-Pe)/(1-Pe)

In the formula, nn _i is the number of pixels in the entire image that are correctly classified into the i-th category, xx _i and yy _i are respectively the total number of pixels in the entire image that are classified into the i-th category by the classifier and the total number of true reference images in the i-th category. The element number, Pe, is the sum of the products of the number of real pixels of various types and the number of predicted samples, divided by the square of the total number of samples.

After using 70% of the sample points (555*0.7=388) to train the classifier, apply the trained classifier and input to the entire area for calculation to obtain the classification result of the entire area, and then pass 30% of the sample points test. As shown in Figure 4, it is the result of vegetation classification using actual evapotranspiration proposed by this method. The classification results of the entire area are compared with the results of visual interpretation in step 4.

Overall, the OA and Kappa of the classification are 0.930 and 0.914 respectively. The PA and UA of the six land cover types are all above 85%, and the F1 is above 90%. The classification accuracy is relatively high. In 2020, the Ordos Basin has the largest proportion of grassland, accounting for approximately 46.1%. Followed by cultivated land and grassland, accounting for 21.3% and 16.5% respectively, and unused land accounting for 11.2%, this part is mainly desert. Urban land accounts for about 3.9%, and water bodies account for less than 1%.

Example 2

Using the classifier in step 6, simulate and calculate the results of land cover classification under the following five input schemes (S1-S5) and compare their accuracy. The input of scheme 1 (S1) is only the 6 spectral band information of Landsat images; scheme In addition to the spectral bands in S1, the input of 2 (S2) adds three spectral indices; the input of option 3 (S3) adds digital elevation information on the basis of S2, including elevation and terrain slope; option 4 (S4) The input of Option 5 is to add nighttime light data on the basis of S3; Scheme 5 (S5) combines all the input information, that is, it adds multiple sets of instantaneous evapotranspiration information in a year on the basis of S4.

Figure 5 is a histogram of accuracy comparison under different input schemes in Embodiment 1, 5a is a comparison of overall accuracy of different input schemes, and 5b is a comparison of Kappa coefficients of different input schemes. For the overall accuracy OA, S1 (0.676) is close to S2 (0.665), S3 (0.730) is close to S4 (0.733), S5 is the highest (0.807), and Kappa is similar to OA. It shows that increasing the spectral index based on S1 or adding night light based on S3 has little impact on the results. After adding DEM, the accuracy of S3 has been greatly improved, which proves that elevation and slope play a large role in the above-mentioned vegetation classification. In particular, after adding instantaneous evapotranspiration, the accuracy of S5 has been significantly improved. S5 combines all the above features as the input of the classifier. S5 improves OA by 7%-14% and Kappa by 9%-17%. For different land cover types, relative to S1-S4, S5 increases F1 by 5%-15%, 0%-16%, 3%-20%, 10%-30%, 10%-13%, 1 respectively. %-10%, respectively, refer to the accuracy improvement on different land cover types, that is, for 6 types of land cover types: cultivated land, woodland, grassland, urban land, water body and unused land. For example, S5’s classification accuracy for cultivated land is relatively high. It is 5-15% higher than the first four solutions (S1-S4). The results prove that S5 not only significantly improves the classification accuracy of cultivated land, woodland, and grassland, but also proves that instantaneous evapotranspiration is an effective input feature.

Example 3

Compare the accuracy of the classification results of this method with existing land cover products of multiple resolutions (CNLUCC products, MCD12Q1, Globe Land products, CLCD products, CRLC products, GLC_FCS products). Through the calculation results of multiple objective functions in step 6, The closer the objective function is to 1, the higher the classification accuracy.

Figure 6 shows the comparison between the land cover classification results based on actual evapotranspiration estimation in Example 1 and existing products. 6a is a CNLUCC product, 6b is an MCD12Q1 product, 6c is a Globe Land product, 6d is a CLCD product, 6e is a GLC_FCS product, 6f is a CRLC product, and 6g is the result of the method proposed by the present invention, AETLULC; from (a) to (f), The spatial resolutions of the six types of data are 1km, 500m, 30m, 30m, 30m, and 10m respectively. The product resolution obtained by the method proposed in the present invention is 30m. The land cover of the CNLUCC product is relatively scattered, while the MCD12Q1 land cover product is relatively concentrated in overall land cover types, especially grassland and cultivated land, which are clustered in a large area and cannot represent the true situation of mixed vegetation. CLCD products, CRLC products, and GLC_FCS products are intuitively closer to the classification results (AETLULC) of the present invention, but are slightly different for unused land dominated by cultivated land, grassland, and deserts. Compared with six existing products, the F1 of AETLULC on grassland increased by 32.7%, 43.5%, 21.0%, 15.8%, 10.9%, 4.9%; the F1 of AETLULC on cultivated land increased by 28.1%, 18.1%, 10.8%, 2.4 %, 5.2%, 2.1%; F1 on forest land increased by 29.2%, 24.1%, 8.2%, 12.0%, 11.2%, 1.0%.

Example 4

In order to verify the rationality and advancement of the method, KL divergence and JS divergence are used to evaluate the difference in feature input under different vegetation types. The greater the difference, the more effective the feature input is in classification. FI importance is used to evaluate the feature input. The greater the importance of FI, the higher the importance of input features in classification. Calculated as follows:

In the formula, F(x) and P(x) represent the probability distribution on the discrete space X, KL(F||P) and JS(F||P) are the KL of F(x) to P(x) respectively. Divergence and JS divergence, ntree is the number of decision trees in the random forest, errOOB _t1 represents the out-of-bag error after the feature input value in the t-th tree changes, errOOB _t2 represents the bag of normal feature input values in the t-th tree external error.

Figure 7 is a histogram of feature input difference and importance in Embodiment 1, where a is the KL divergence and JS divergence of multiple feature inputs under different land cover types, and b is the importance ranking of multiple feature inputs. For different land cover types, the input features that show great differences are: B6, B7, B4, NDVI, NDBI, B2, B3, ET _inst21 , ET _inst3 , ET _inst10 , etc. The importance of the input features is ranked as follows: NDVI, B5, MNDWI, B3, B2, B6, ELEVATION, B7, B4, ET _inst12 , ET _inst21 . It shows that the reflectance information of the spectral band itself can express the differences in vegetation types to a large extent, and the additional information can also show the differences in vegetation types to a certain extent. Therefore, the input of additional information such as instantaneous evapotranspiration can also help improve the classification accuracy.

Figure 8 shows the difference in instantaneous evapotranspiration under the three vegetation types in Example 1. a is the difference in the annual change of instantaneous evapotranspiration, and b is the difference in instantaneous evapotranspiration in different precipitation areas. Regarding the instantaneous evapotranspiration on different dates throughout 2020, the evapotranspiration on the three vegetation types of cultivated land, forestland, and grassland showed a changing pattern of first increasing and then decreasing. In terms of quantity, the evapotranspiration of woodland is generally greater than that of grassland and cultivated land, and the evapotranspiration of grassland and cultivated land is relatively close. The average values of instantaneous evapotranspiration in woodland, cultivated land, and grassland are 0.56mm, 0.42mm, and 0.41mm respectively. In autumn and winter, the evapotranspiration of grassland is slightly greater than that of cultivated land. In summer and late spring, the evapotranspiration of cultivated land is greater than that of grassland. This is consistent with the regional related to the agricultural planting system. Evapotranspiration in different periods shows different characteristics on different vegetation types. Refined to each precipitation area, the evapotranspiration of woodland is the largest and can be significantly distinguished from cultivated land and grassland. The evapotranspiration of cultivated land and grassland is also very close, but as precipitation increases, the degree of differentiation increases. Through verification and evaluation using multiple methods, the accuracy and feasibility of introducing instantaneous evapotranspiration to improve classification accuracy can be proven.

In the context of changing environments, the method proposed in the present invention can provide technical support for rapid and accurate identification of vegetation in highly heterogeneous areas, and provide theoretical basis for future applications in regional land and resources planning, ecological environment protection, etc.

The results obtained by existing literature 4 and existing literature 5 are different, and it is difficult to distinguish different vegetation types. The vegetation phenological characteristics of cultivated land and grassland overlap to a certain extent and are difficult to distinguish. For example, the growth and development periods of once-a-year crops and grassland are very close. They begin to develop in spring, grow rapidly in summer, and wither in autumn and winter. They only rely on vegetation coverage. It is difficult to distinguish, that is, the same vegetation coverage or similar phenological patterns may be different land cover types.

It should be noted that the above specific implementation modes and examples are preferred implementation modes of the present invention and cannot be used to limit the protection scope of the present invention. Any modification made based on the technical solution based on the technical ideas proposed by the present invention Equivalent changes or equivalent modifications still fall within the protection scope of the technical solution of the present invention.

Claims (7)

1. A method for estimating vegetation classification based on remote sensing actual evapotranspiration, which is characterized by comprising the following steps:

step one: collecting hydrologic, meteorological, topographic, underlying surface and remote sensing image data in a research flow field, sorting site data according to time sequence, cutting space data according to a vector range of the research flow field, and resampling to the same resolution;

step two: constructing an SEBAL actual evapotranspiration estimation model suitable for a research basin based on the data collected in the step one;

step three: calculating to obtain the actual evaporation of the research drainage basin on a multidimensional time scale by using the SEBAL actual evaporation model built in the second step;

step four: generating random sample points, identifying the land coverage types of the sample points by using a visual interpretation method, and editing a sample point attribute table; and according to 7:3, dividing the ratio into training samples and test samples;

step five: establishing a classification system and a characteristic input database, reasonably setting land coverage type classification, and arranging a database which takes five types of information of spectral bands, spectral indexes, topography factors, night lights and evapotranspiration as characteristic input;

step six: constructing a classifier, selecting a proper machine learning algorithm as the classifier, substituting the classifier into the extraction value of the training sample point in the fourth step on each characteristic input, and training the optimized classifier; and (3) verifying the sample points in the test sample in the step (IV), and comparing the result calculated by the classifier with the real land coverage type on the test sample points obtained by visual interpretation in the step (IV), so that the classification precision can be obtained, and the precision can be evaluated.

2. The method of estimating vegetation classification based on remote sensing actual evapotranspiration as claimed in claim 1, wherein the characteristic input database in the fifth step specifically includes 6 band spectral data of Landsat8, namely, 2,3, 4, 5, 1 and 7 short wave infrared band, 3 spectral indexes of NDVI, MNDWI and NDBI, a night light index, 2 topography factors of elevation and slope, and 33 total characteristics of 21 instantaneous evapotranspiration.

3. A method for estimating vegetation classification based on remote sensing actual evapotranspiration as claimed in claim 2 wherein 3 spectral indexes of Landsat8 are calculated as follows:

NDVI＝(B5-B4)/(B5+B4)

MNDWI＝(B3-B6)/(B3+B6)

NDBI＝(B6-B5)/(B6+B5)。

4. the method of estimating vegetation classification based on remote sensing actual evapotranspiration as claimed in claim 1 wherein the land cover types in step four and step five include cultivated land, woodland, grassland, town land, water and unutilized land.

5. A method of estimating vegetation classification based on remote sensing actual evapotranspiration as claimed in claim 1 wherein the objective function used in step six to evaluate classification accuracy includes, but is not limited to, user accuracy UA, producer accuracy PA, F1 score, overall accuracy OA and Kappa coefficient.

6. The method for estimating vegetation classification based on remote sensing actual evapotranspiration as claimed in claim 5, wherein the calculation formulas of the objective function user accuracy UA, the producer accuracy PA, the F1 score, the overall accuracy OA and the Kappa coefficient are as follows, respectively:

PA＝nn/yyt

Kappa＝(OA-Pe)/(1-Pe)

wherein nni is the number of pixels of the whole image correctly divided into the i-th type of pixels, xxi and yyi are the total number of pixels of the whole image divided into the i-th type and the i-th type of real reference total number of pixels by the classifier respectively, and Pe is the sum of products of the number of the various real pixels and the number of the predicted samples, and divided by the square of the total number of the samples.

7. The method of estimating vegetation classification based on remote sensing actual evapotranspiration as claimed in claim 1, wherein the hydrologic, meteorological, topographic, underlying surface and remote sensing image data in the research flow area collected in the step one specifically comprises: runoff data of hydrologic stations in a river basin, meteorological data of surrounding meteorological stations in the river basin, a digital elevation data set, a vegetation index, a ground surface temperature, a ground surface reflectivity data set, a land water reserve data set, a remote sensing image and a night light data set are researched.