A deep learning-based technique for firm classification and domain adaptation in land cover classification using time-series aerial images

【类型】期刊

【作者】Kalita Indrajit  Chakraborty Shounak  Reddy Talla Giridhara Ganesh  Roy Moumita 

【刊名】Earth Science Informatics

【关键词】 Convolutional,Series,images,LSTMs,Attention,adaptation,Time,datasets,mechanism,Domain,temporal,multi

【摘要】In this manuscript, a novel framework has been presented for firm classification of a geographical area based on spatial as well as time-series analysis of multi-temporal very high resolution (VHR) satellite images. For this dual objective, an attention-based deep learning mechanism combined with the capabilities of convolutional-recurrent neural networks has been investigated for this purpose. The proposed classification strategy is introduced as ‘firm’ since it allows the classifier to assign only one class label to a multi-temporal image stack of co-registered images, as opposed to multiple. This technique ascertains the land-cover class by taking into consideration the geophysical changes on a landmass and thus outsmarting the conventional techniques relying on the visual interpretation of a single image. The attention mechanism focuses on the important portions of the image scene while the convolutional long short-term memory neural networks exploit the temporal dependencies on the time-series image scenes. Moreover, an adaptive land cover classification scheme, considering the features extracted from the proposed classification approach has been explored for more robust time-series based firm classification. To assess the performance of the proposed schemes, the experiments have been conducted on the two novels VHR multi-temporal land cover classification datasets. The investigated models have been shown to have the capacity to outperform the other state-of-the-art techniques under non-adaptive as well as adaptive scenarios using the multi-temporal images captured over disjoint geographical locations.

【年份】2023

【作者单位】https://ror.org/05qwgg493,grid.189504.1,0000 0004 1936 7558,Computing & Data Sciences (CDS),Boston University,02215,Boston,Massachusetts,USA;https://ror.org/00gmd7q80,grid.444467.1,Computer Science and Engineering,Indian Institute of Information Technology Design and Manufacturing,518007,Kurnool,Andhra Pradesh,India;grid.273335.3,0000 0004 1936 9887,Computer Science and Engineering,University at Buffalo,14260,Buffalo,NewYork,USA;https://ror.org/00bb9ch64,grid.512560.5,0000 0004 9284 0348,Computer Science and Engineering,Indian Institute of Information Technology Guwahati,781015,Guwahati,Assam,India

【期号】1

【页码】655-678