Feature Extraction Of Vegetation In Non-agricultural Habitats Based On GF-2 Satellite Images And Unmanned Aerial Vehicles Remote Sensing Images

With the development of intensive agriculture,non-agricultural habitats in the agricultural landscape are gradually reduced or disappeared,and the landscape tends to become homogenised,resulting in the loss of biodiversity in farmland ecosystems.At the same time,the uncontrolled and unsustainable exploitation of nature by human beings for their own benefit and the increasing intensity of land resource exploitation have led to an accelerated rate of biodiversity loss and ecosystem degradation,and the phenomenon of"silent fields"has emerged.Food security and the increasing demand for land for construction versus ecological land require us to protect farmland ecosystems and restore their biodiversity from the perspective of restoring and enhancing their service functions.The vegetation structure of non-farm habitats has an important impact on biodiversity in farmland landscapes.Accurately identifying and obtaining information on the vegetation characteristics and spatial location of non-farm habitats is a prerequisite for biodiversity conservation in farmland ecosystems.Previous studies have been limited by data sources and sampling methods,making it difficult to extract vegetation features of non-agricultural habitats at small scales with rapid and high accuracy.Some studies have shown that by increasing the spatial resolution of images and expressing vegetation feature information from a more detailed perspective,the biodiversity estimation capability can be effectively increased.Based on this,this paper selects two different patterns of agricultural landscapes in Shenyei New Area and Changtu County as the study area of the Lower Liaoning River Plain.With the help of GF-2 remote sensing image data,the e Cognition software combined with R language was used to identify and classify the landscape types using a face-to-face approach,and the effects of different feature preference algorithms and different classifiers on the classification results were explored.Based on the best classification results,the evaluation and spatial mapping of vegetation growth characteristics and plant diversity characteristics in non-agricultural habitats were carried out in combination with low-altitude aerial images from UAVs,and the following conclusions were obtained.(1)The multi-layer scale parameter selection tool(ESP2)can effectively achieve the optimal segmentation scale of the GF-2 image.The optimal segmentation scales for the T1 and T2 test areas are 342 and 282 respectively when the weight of the shape and tightness parameters are set to 0.3 and 0.4.(2)The use of visible light images acquired by a light and small UAV,combined with a small amount of field-collected actual measurement data,enables the identification and classification of plant species in the woodland landscape through human-computer interaction,supplementing the number of field samples to meet the needs of model building,while compensating for the limitations and shortcomings of field sampling.(3)The C5.0 algorithm can effectively achieve classification feature preference and reduce information redundancy.Using the C5.0 algorithm,12 and 15 features were optimised in the feature sets of T1 and T2 test areas respectively,which were reduced by 73 and 70compared to the initial set of classification features,while the Kappa coefficient remained at a high level,and only reduced by 0.05 and 0.04 compared to the classification results of the importance function optimised features.(4)Establishing an object-oriented random forest classification model with better accuracy in the classification of agricultural landscapes.Based on the optimal segmentation of the image,using the preferred feature set,feature information extraction and model construction,the landscape in the area was classified into 11 categories:paddy field,watered land,dry land,forest land,grassland,water surface,urban and rural construction land,highway land,hardened road,bare land,and shade,and the object-oriented random forest classification model had higher applicability,and the Kappa coefficients of the T1 and T2 test areas reached 0.85 and0.78,an improvement of 0.54 and 0.05 compared with the support vector machine under the same preferred feature set.(5)The results of vegetation cover estimation using EVI have higher accuracy.EVI introduced the blue band to increase the expression of vegetation information based on the use of near-infrared and light bands,weakening the influence of aerosol and soil background,and compensating for the deficiency of NDVI in vegetation cover in high-coverage landscapes.the R2 of the coefficient of determination between the EVI estimation results and the measured data reached 0.8415(p<0.01),and the root mean square error RMSE was 0.1034.(6)The estimation of plant diversity using the textural features of the near-infrared band of the images was better.There was a significant correlation between plant diversity and the textural features of the image NIR bands,with the Simpson dominance index index of plants in grassland landscapes correlating with synopticism at 0.844(P=0.01)and the best fit between the two,with an adjusted R~2 of 0.696.