Neural network models for land cover classification from satellite images

Land cover data represent environmental information for a variety of scientific and policy applications, so its classification from satellite images is important. Since neural networks (NN) do not require a hypothesis about data distribution, they are valuable tools to classify satellite images. The objectives of this work were to develop NN models to classify land cover data from information from satellite images and to evaluate them when different input variables are used. MODIS-MYD13Q1 satellite images and data of 85 plots in Cordoba, Argentina, were used. Five NN models of multi-layer feed-forward perceptron were designed. Four of these received NDVI (Normalized Difference Vegetation Index), EVI (Enhanced Vegetation Index), red (RED) and near infrared (NIR) reflectance values as input patterns, respectively. The fifth NN had RED and NIR reflectances as input values. By comparing the information taken in the field and the classification made during the validation phase, it can be concluded that all models presented good performance in the classification. The model that shows better performance is the one that jointly considers RED and NIR reflectance as input; this model shows an overall classification accuracy of 93% and an excellent Kappa statistic. The networks constructed with NDVI and EVI values have a similar performance (86 and 83% accuracy, respectively). The Kappa statistics correspond to the categories of very good and good, respectively. The networks including only RED or NIR reflectance values get the lowest accuracy results (76 and 81 %, respectively) and Kappa values within fair and good ranks, respectively.