Multi-Temporal-Spectral Land Cover Classification for Remote Sensing Imagery Using Deep Learning

Sustainability research of the environment depends on accurate land cover information over large areas. Even with the increased number of satellite systems and sensors acquiring data with improved spectral, spatial, radiometric and temporal characteristics and the new data distribution policy, most existing global land cover datasets were derived from a single-date multi-spectral remotely sensed image using pixel-based classifiers with low accuracy. To improve the accuracy, the bottleneck is how to develop accurate and effective image classification techniques. By incorporating and utilizing the spatial and multi-temporal information with multi-spectral information of remote sensing images for land cover classification, and considering their spatial and temporal interdependence, I propose three deep network systems tailored for medium-resolution remote sensing data. With a test site from the Florida Everglades area (with a size of 771 square kilometers), the proposed new deep systems have achieved significant improvements in the classification accuracy over most existing pixel-based classifiers. A proposed patch-based recurrent neural network (PB-RNN) system, a proposed pixel-based recurrent neural network system and a proposed patch-based convolutional neural network system achieve 97.21%, 87.65% and 89.26% classification accuracy respectively while a pixel-based single-image neural network (NN) system achieves only 64.74% classification accuracy. By integrating the proposed deep networks and the huge collection of medium-resolution remote sensing data, I believe that much accurate land cover datasets can be produced over large areas.