Land Cover Classification from Multi-temporal, Multi-spectral Remotely Sensed Imagery using Patch-Based Recurrent Neural Networks

摘要

Sustainability of the global environment is dependent on the accurate landcover information over large areas. Even with the increased number of satellitesystems and sensors acquiring data with improved spectral, spatial, radiometricand temporal characteristics and the new data distribution policy, mostexisting land cover datasets were derived from a pixel-based single-datemulti-spectral remotely sensed image with low accuracy. To improve theaccuracy, the bottleneck is how to develop an accurate and effective imageclassification technique. By incorporating and utilizing the completemulti-spectral, multi-temporal and spatial information in remote sensing imagesand considering their inherit spatial and sequential interdependence, wepropose a new patch-based RNN (PB-RNN) system tailored for multi-temporalremote sensing data. The system is designed by incorporating distinctivecharacteristics in multi-temporal remote sensing data. In particular, it usesmulti-temporal-spectral-spatial samples and deals with pixels contaminated byclouds/shadow present in the multi-temporal data series. Using a FloridaEverglades ecosystem study site covering an area of 771 square kilo-meters, theproposed PB-RNN system has achieved a significant improvement in theclassification accuracy over pixel-based RNN system, pixel-based single-imageryNN system, pixel-based multi-images NN system, patch-based single-imagery NNsystem and patch-based multi-images NN system. For example, the proposed systemachieves 97.21% classification accuracy while a pixel-based single-imagery NNsystem achieves 64.74%. By utilizing methods like the proposed PB-RNN one, webelieve that much more accurate land cover datasets can be produced over largeareas efficiently.