Temporal convolutional neural network for land use and land cover classification using satellite images time series

Satellite image time series provide essential information to closely monitor vegetation dynamics, revealing growth pattern and phenological characteristics of crops. In this study, we explored a one-dimension Temporal Convolutional Neural Network (1D-TempCNN) model for land use and land cover classification from Sentinel-2 image time series. The main goal was to evaluate the generalization ability by comparing 1D-TempCNN with two classical methods, Support Vector Machine (SVM) and Random Forest (RF). The experiments were conducted in three different sites, each one with its own crop cycles and time series compositions, and we performed cross-site and cross-year tests. These cross-testing approaches allow examining the ability of the classification models to generalize temporal (and spectral) patterns. Two data augmentation techniques, sliding window and scaling, were used to increase the amount and diversity of data. Our results show that data augmentation technique was essential to handle the data variability, contributing to the generalization of models and overall performance, resulting in an increase of up to 19.72 percentage points in overall accuracy (OA). Additionally, results show that 1D-TempCNN achieved superior OA (94.34–98.67%) in cross-testing, outperforming SVM (88.32–97.45%) and RF (90.25–98.12%). This demonstrates that the proposed 1D-TempCNN model, combined with data augmentation techniques, is capable of higher generalization for land use and land cover classification.