The combination of multi-temporal images and deep learning is an efficient way to obtain accurate crop distributions and so has drawn increasing attention. However, few studies have compared deep learning models with different architectures, so it remains unclear how a deep learning model should be selected for multi-temporal crop classification, and the best possible accuracy is. To address this issue, the present work compares and analyzes a crop classification application based on deep learning models and different time-series data to exploit the possibility of improving crop classification accuracy. Using Multi-temporal Sentinel-2 images as source data, time-series classification datasets are constructed based on vegetation indexes (VIs) and spectral stacking, respectively, following which we compare and evaluate the crop classification application based on time-series datasets and five deep learning architectures: (1) one-dimensional convolutional neural networks (1D-CNNs), (2) long short-term memory (LSTM), (3) two-dimensional-CNNs (2D-CNNs), (4) three-dimensional-CNNs (3D-CNNs), and (5) two-dimensional convolutional LSTM (ConvLSTM2D). The results show that the accuracy of both 1D-CNN (92.5%) and LSTM (93.25%) is higher than that of random forest (~ 91%) when using a single temporal feature as input. The 2D-CNN model integrates temporal and spatial information and is slightly more accurate (94.76%), but fails to fully utilize its multi-spectral features. The accuracy of 1D-CNN and LSTM models integrated with temporal and multi-spectral features is 96.94% and 96.84%, respectively. However, neither model can extract spatial information. The accuracy of 3D-CNN and ConvLSTM2D models is 97.43% and 97.25%, respectively. The experimental results show limited accuracy for crop classification based on single temporal features, whereas the combination of temporal features with multi-spectral or spatial information significantly improves classification accuracy. The 3D-CNN and ConvLSTM2D models are thus the best deep learning architectures for multi-temporal crop classification. However, the ConvLSTM architecture combining recurrent neural networks and CNNs should be further developed for multi-temporal image crop classification.

The accurate mapping of urban impervious surfaces from remote sensing images is crucial for understanding urban land-cover change and addressing impervious-surface-change-related environment issues. To date, the authors of most studies have built indices to map impervious surfaces based on shortwave infrared (SWIR) or thermal infrared (TIR) bands from middle–low-spatial-resolution remote sensing images. However, this limits the use of high-spatial-resolution remote sensing data (e.g., GaoFen-2, Quickbird, and IKONOS). In addition, the separation of bare soil and impervious surfaces has not been effectively solved. In this article, on the basis of the spectra analysis of impervious surface and non-impervious surface (vegetation, water, soil and non-photosynthetic vegetation (NPV)) data acquired from world-recognized spectral libraries and Sentinel-2 MSI images in different regions and seasons, a novel spectral index named the Normalized Impervious Surface Index (NISI) was proposed for extracting impervious area information by using blue, green, red and near-infrared (NIR) bands. We performed comprehensive assessments for the NISI, and the results demonstrated that the NISI provided the best studied performance in separating the soil and impervious surfaces from Sentinel-2 MSI images. Furthermore, regarding impervious surfaces mapping accuracy, the NISI had an overall accuracy (OA) of 89.28% (±0.258), a producer’s accuracy (PA) of 89.76% (±1.754), and a user’s accuracy (UA) of 90.68% (±1.309), which were higher than those of machine learning algorithms, thus supporting the NISI as an effective measurement for urban impervious surfaces mapping and analysis. The results indicate the NISI has a high robustness and a good applicability.