Authors: Liu He*, University of North Carolina - Chapel Hill, Conghe Song, University of North Carolina - Chapel Hill
Topics: Remote Sensing, Land Use and Land Cover Change
Keywords: deep convolution neural networks (CNNs), spatial resolution, land cover classification
Session Type: Paper
Start / End Time: 1:10 PM / 2:50 PM
Room: Buchanan, Marriott, Mezzanine Level
Presentation File: No File Uploaded
Deep convolution neural networks (CNNs) have reflected superiority over conventional machine learning methods in high-resolution land cover classification work. To intimate common CNNs work environment in computer vision, recent researchers mostly paid attention to small-scene high-resolution experiments, expecting enough centered pixels in each sample were eligible for training. CNNs was utilized as black box; network parameters were configurated and fine-tuned, aiming to obtain state-of-art accuracy of pixel-wise classification outputs in a certain area. In this, the essential tradeoff between spatial resolution and precise segmentation was reluctantly avoided through stingy data selection. Thus large-scale medium-resolution satellite images like Landsat archive were ignored in former researches. However, the power of CNNs to learn deep context feature for fine-grained medium resolution images was proved in this paper. We proposed an experimental benchmark for CNNs configuration and implementation on end-to-end land cover classification in multi-spatial resolutions. For each resolution case ranging from 1 to 30 meters, we optimized CNNs through deep residual blocks and fully convolutional layers to obtain precisely localized and labeled classification results. The influence of striding, pooling, and upsampling processes within networks upon the localizing and labeling accuracy of results was firstly illustrated and analyzed. We also evaluated robustness of our networks under imperfect training dataset to interpret the influence of propagating contaminations, and proposed coarse-labeled conditions to further exploit the capability of CNNs. The initial result indicated potential strength of semi-supervised deep training which corresponded to practical real world applications.