Vegetation Land Use/Land Cover Extraction From High-Resolution Satellite Images Based On Adaptive Context Inference

Land cover classification results are commonly used to represent the physical properties of a land's surface,while land use classification results correspond to the activities or functions for which humans utilize land,and activities or functions are inherently related but are nevertheless conceptually distinctive compared with physical properties.Therefore,images are always ambiguously classified base on land use criterion rather than uniquely classified base on land cover criterion during mapping.In this paper,automatic extraction of multi-context and multi-scale land use/land cover vegetation from high-resolution remote sensing images is tackled,aiming to solve typical challenges in classifying remote sensing images at a pixel level.We integrated adaptive context inference(ACI)and focus perception(FP)modules into a semantic segmentation framework to automatically and adaptively extract vegetation in HRRSI,based on the various requirements of land use/land cover vegetation mapping applications.The content of this paper mainly includes the following aspects:(1)We designed an automatic vegetation extraction model for HRRSI based on adaptive context inference(ACI)model and vegetation-feature-sensitive focus perception(FP)module,to automatically extract vegetation in various types of HRRSI vegetation areas end-to-end.Considering that the sementatic segmentation network should retain the spatial detail information as well as high-level semantic information of images,we designed the adaptive context inference model based on the baseline model,which consists of baseline model,vegetation-feature-sensitive focus perception(FP)module and adaptive context inference(ACI)model.(2)We introduced a FP module to extract sensitive features from different types of vegetation and an integrated attention mechanism containing high-level and lowlevel semantic information to solve the problem of small interclass and large intraclass differences between vegetation features and other types of typical features.(3)We established context inference model to represent relationships between the center pixel and its neighbors under semantic constraints,as well as different spatial structures of vegetation features with a diversity in size,shape,and context.To solve the intrinsic property of vegetation objects whose spatial structure cannot be exhaustively expressed in fixed patterns,we established an adaptive context inference model under a supervised setting to satisfy the inference of spatial structure relations based on the data-driven pattern recognition methodology;In the ACI model,we used the pixel-wise cross-entropy loss as a form of unary supervision to enforce semantic labels on individual pixels,and modeled relationships between neighboring pixels in the label space and operated context inference during the training stage.We introduced an adaptive context inference loss function to automatically obtain segments which are respectful of spatial structures and small details.(4)We conducted comparative experiments to analyze the influences of different modules on vegetation extraction results,as well as an ablation study and parameter sensitivity analysis,which finally gave some conclusions about different modules' influences and extracted land use/land cover vegetation in HRRSI automatically and adaptively.We also conducted experiments to integrate the deep-learning-based classification results and pixel-wise classification results of traditional HRRSI classification method,which to effectively combine semantic information of the former sementatic inference stage and precise location information of latter spatial details.We also conducted comprehensive experiments on ZY-3 and GF-2 HRRSI datasets to validate the effectiveness of our proposed approach,including full-image as well as local areas which are difficult to classify.Comparative experiments on the ZY-3 and Gaofen Image Dataset(GID)datasets demonstrate the effectiveness of our proposed automatic vegetation extraction model against the baseline Deeplab v3+ model.Taking precision,kappa coefficient,mean intersection over union(miou),precision rate,and F1-score as the evaluation indexes,the results showed an improvement in the precision by at least 1.44% and miou by2.47%,over the baseline Deeplab v3+ model.In addition,the ACI module improved the precision and miou by 2% and 3.88%,and the FP module improved the precision and miou by 1.13% and 1.65%.These results and statistics of these comprehensive experiments illustrated that our adaptive and effective vegetation extraction model could satisfy different requirements of land use/land cover mapping applications.