Optimizing Spatial-Spectral-Temporal Neural Network Models for Large-Scale Landcover Classification Based on Landsat Data Archive

Abstract:

Large-scale land cover mapping using satellite data is challenging due to a wide range of factors such as spectral mixing, change of spectral reflections over time, weakness of shallow network models, and computing resources. In our research we have developed land cover classification models based on recent advancements in deep networks and machine learning using recurrent LSTM cells to fully employ temporal data on a yearly basis. We performed an extensive selection search on spatial features based on GLCM methods. Our models were trained by random sampling from 84 different ecosystem regions throughout the conterminous United States, by carefully screening USGS's Landcover Trends maps and verifying pixel temporal stability through high resolution historical imagery on Google Earth over a decade.

The deep LSTM model was implemented on Tensorflow/Python and optionally complemented with fully-connected dense layers. We tested multiple configurations/options for network hyperparameter identification such as number of layers and cells, training data size, optimizer type, data batch size and learning rate, and most importantly the spatial features and additional spectral indices. We found a mixture of different GLCM features to help for identifying different land cover classes. The overall achieved performance (accuracy and average F1) were substantially higher than 90% when GLCM features were utilized and enhanced by regularization techniques. We compared the results to conventional classifiers (SVM and Random Forest), and our early results indicate considerable accuracy gains with our proposed recurrent deep neural network.

We are extending the work by moving into a larger dataset and deeper networks with additional dense and convolutional layers.