Land Cover Classification via Multi-temporal Spatial Data by Recurrent Neural Networks

摘要

Nowadays, modern earth observation programs produce huge volumes of satelliteimages time series (SITS) that can be useful to monitor geographical areasthrough time. How to efficiently analyze such kind of information is still anopen question in the remote sensing field. Recently, deep learning methodsproved suitable to deal with remote sensing data mainly for sceneclassification (i.e. Convolutional Neural Networks - CNNs - on single images)while only very few studies exist involving temporal deep learning approaches(i.e Recurrent Neural Networks - RNNs) to deal with remote sensing time series.In this letter we evaluate the ability of Recurrent Neural Networks, inparticular the Long-Short Term Memory (LSTM) model, to perform land coverclassification considering multi-temporal spatial data derived from a timeseries of satellite images. We carried out experiments on two differentdatasets considering both pixel-based and object-based classification. Theobtained results show that Recurrent Neural Networks are competitive comparedto state-of-the-art classifiers, and may outperform classical approaches inpresence of low represented and/or highly mixed classes. We also show thatusing the alternative feature representation generated by LSTM can improve theperformances of standard classifiers.