Satellite image time series classification has become a research focus with the launch of new remote sensing sensors capable of capturing images with high spatial, spectral, and temporal resolutions. In particular, in the field of crop classification, time dimension information is particularly important. Although some advanced machine learning algorithms, such as random forests (RFs), can achieve good results, they often ignore the time series information. To make full use of temporal and spectral information in multitemporal remote sensing images, a channel attention-based temporal convolutional network (CA-TCN) is proposed in this letter. Specifically, the proposed method is composed of two main modules: temporal convolutional network and attention block. The temporal convolutional network can capture long-range dependence by using a hierarchy of temporal convolutional filters. To capture relevant information inside the sequence and enhance the important information, the attention block is used to enhance the important features in the channel dimension since not all bands contain equal information in crop type classification. The proposed CA-TCN can excavate deeper phenological characteristics. Compared to the temporal attention-based temporal convolutional network and other deep learning-based models, the proposed CA-TCN has achieved state-of-the-art performance in the Breizhcrops dataset with fewer parameters.

In this letter, CatBoost was first introduced and investigated for remote sensing (RS) image classification using diverse features. To improve the classification performance by fostering the effective and efficient spatial feature extraction, a new pseudo label features (PLFs) extraction method was proposed via multisize neighboring patches-based multiclustering. Experimental results on two hyperspectral and one PolSAR benchmarks showed that: 1) CatBoost is an advanced ensemble learning (EL) algorithm for classification of RS images using diverse features; 2) CatBoost has better capability of reducing the overfitting issue at large number of boosting iteration; and 3) proposed PLFs can result in compatible and even better classification results than using morphological profiles (MPs) and MPs with partial reconstruction (MPPR) spatial features.

Detailed land cover change in multitemporal images is an important application for earth science. Many techniques have been proposed to solve this problem in different ways. However, accurately identifying changes still remains a challenge due to the difficulties in describing the characteristics of various change categories by using single-level features. In this article, a multilevel feature representation framework was designed to build robust feature set for complex change detection task. First, four different levels of information from low level to high level, including pixel-level, neighborhood-level, object-level, and scene-level features, were extracted. Through the operation of extracting different level features from multitemporal images, the differences between them can be described comprehensively. Second, multilevel features were fused to reduce the dimension and then used as the input for supervised change detector with initial limited labels. Finally, for reducing the labeling cost and improving the change detection results simultaneously, active learning was conducted to select the most informative samples for labeling, and this step together with the previous steps were iteratively conducted to improve the results in each round. Experimental results of three pairs of real remote sensing datasets demonstrated that the proposed framework outperformed the other state-of-the-art methods in terms of accuracy. Moreover, the influences of scene scale for high-level semantic features in the proposed approach on change detection performance were also analyzed and discussed.

Effective features derived from an original hyperspectral image (HSI) are quite important to improve the classification performance. An improved feature set, namely HGFM, is constructed by integrating harmonic analysis (HA) optimized by a multiscale guided filter (GF) with morphological operation for HSI classification. To establish HGFM, HA is first adopted to convert the HSI from spectral space to the frequency domain represented by amplitude, phase, and residual. With the first component of minimum noise fraction obtained from the original HSI as the guidance image, the harmonic components are then processed by the multiscale GF. Finally, the obtained results are then operated via morphological opening by reconstruction and closing by reconstruction to generate an improved feature set for classification. The HGFM features are input to an ensemble learning (EL) based on classification framework, in which EL plays an auxiliary role to enhance the classification stability and reliability. Three commonly used HSIs are used for experiments, and different feature sets are evaluated by comparing EL and rotation forest, support vector machine optimized by particle swarm optimization, random forest, and others. Compared with benchmark feature sets, the proposed HGFM feature set can better depict the details of objects easily, and the experimental results confirm the effectiveness in terms of classification accuracy and generalization ability.

Scene change detection identifies functional changes at the scene level. Compared with pixel-level and object-level change detection, it can provide a higher level understanding of changes on the Earth’s surface. Triple-branch networks that perform scene binary change detection and scene classification tasks simultaneously are competitive in the field of scene change detection, as they consider both single-temporal scene semantic information and cross-temporal change features. However, some problems still exist. First, the temporal change feature extraction is insufficient, and the 1-D feature vector used for scene change detection and classification is lacking in representativeness. Second, the predicted scene binary change detection and classification results are often contradictory at the network prediction stage, leading to the unsatisfactory performance of change trajectory identification. To address these issues, a novel framework that integrates differential aggregation network (DAN) and class probability-based fusion strategy (CPFS) was proposed. The designed DAN can fully capture the temporal change features by using four advanced differential fusion modules to aggregate the multi-level difference information. In addition, it is able to generate more representative 1-D feature vectors by adopting two novel attention-aware adaptive pooling modules. The developed CPFS produces the final consistent scene binary change detection and classification maps by fusing three predicted class probability vectors. The proposed method was validated on two datasets, and the results demonstrated its superiority to the comparison methods.

Accurate cropland mapping is significant for food security and sustainable development. The existing cropland map based on remote sensing mainly focus on moderate to coarse spatial resolution, and these products are generally unsuitable for precision agriculture due to the lack of spatial details. Therefore, there is an urgent need to produce high-resolution (HR) cropland maps to meet current application demands. Recently, the typical classification workflow of HR images employs deep learning models combined with manually annotated samples, and visual interpretation of samples is usually labor-intensive and time-consuming, which is not conducive to large-scale applications. To address this problem, this paper proposes an automated HR cropland extraction solution, namely RRE (Refinement-Reclassification-Extraction), including (i) Refinement of 10 m spatial resolution cropland products, (ii) Reclassifying cropland using the refined product as sample source, and (iii) Extracting HR cropland via designed cross-scale sample transfer. The strength of the proposed framework is that it leverages existing moderate-resolution public products as prior knowledge and provides cross-scale transferable samples for HR images. The whole process does not require manual labeling of samples and is highly automated. Specifically, the experimental results in the three main grain production regions show that, the RRE framework effectively reduces the interference of road networks and ridges, and F1 scores of extracted 1 m HR cropland reaches 87.71 %~94.16 %, which is comparable to the fully supervised cropland extraction method. In addition, the 10 m reclassified cropland, produced by the intermediate process of the RRE, outperforms current cropland product of ESRI Land Cover and ESA World Cover.