Search

The timely and accurate mapping of crops over large areas is essential for alleviating food crises and formulating agricultural policies. However, most existing classical crop mapping methods usually require the whole-year historical time-series data that cannot respond quickly to the current planting information, let alone for future prediction. To address this issue, we propose a novel spatial–temporal feature and deep integration strategy for crop growth pattern prediction and early mapping (STPM). Specifically, the STPM first learns crop spatial–temporal evolving patterns from historical data to generate future remote sensing images based on the current observations. Then, a robust crop type recognition model is applied by combining the current early data with the predicted images for early crop mapping. Compared to existing spatial–temporal prediction models, our proposed model integrates local, global, and temporal multi-modal features comprehensively. Not only does it achieve the capability to predict longer sequence lengths (exceeding 100 days), but it also demonstrates a significant improvement in prediction accuracy for each time step. In addition, this paper analyses the impact of feature dimensionality and initial data length on prediction and early crop mapping accuracy, demonstrating the necessity of multi-modal feature fusion for spatial–temporal prediction of high-resolution remote sensing data and the benefits of longer initial time-series (i.e., longer crop planting time) for crop identification. In general, our method has the potential to carry out early crop mapping on a large scale and provide information to formulate changes in agricultural conditions promptly.

With the development of remote sensing technology, the continuing accumulation of remote sensing data has brought great challenges to the remote sensing field. Although multiple deep-learning-based classification methods have made great progress in scene classification tasks, they are still unable to address the problem of model learning continuously. Facing the constantly updated remote sensing data stream, there is an inevitable problem of forgetting historical information in the model training, which leads to catastrophic forgetting. Therefore, we propose a continual contrastive learning method based on knowledge distillation and contrastive learning in this paper, which is named the Continual Contrastive Learning Network (CCLNet). To overcome the problem of knowledge forgetting, we first designed a knowledge distillation module based on a spatial feature which contains sufficient historical knowledge. The spatial and category-level knowledge distillation enables the model to effectively preserve the already learned knowledge in the current scene classification model. Then, we introduced contrastive learning by leveraging the comparison of augmented samples and minimizing the distance in the feature space to further enhance the extracted feature during the continual learning process. To evaluate the performance of our designed model on streaming remote sensing scene data, we performed three steps of continuous learning experiments on three datasets, the AID, RSI, and NWPU datasets, and simulated the streaming of remote sensing scene data with the aggregate of the three datasets. We also compared other benchmark continual learning models. The experimental results demonstrate that our method achieved superior performance in the continuous scene classification task.

During the past several decades, the carbon budget in the dryland ecosystem has experienced great variation under the joint impact of climate change and anthropogenic interference. How the net ecosystem productivity (NEP) responds to climate change and human interference in the Qilian Mountains (QLM), Northwest China, remains unclear. To fill these gaps, we first estimated the NEP in the QLM and then quantified the independent and interactive influences of natural environment factors, climatic factors, and human activity intensity on the NEP change from 2000 to 2020 by linking the Geodetector and structural equation models. The NEP of the QLM showed a significant increase during the recent 20 years, and 78.93% of the QLM experienced a significant increase in NEP; while only 4.83% of the area in the QLM experienced a decreasing trend, which is dominantly located on the southeast edge, surrounding the Qinghai Lake, and the midland of the QLM. The area percentage of the carbon sink region increased from 47% in 2000 to 62% in 2020. The natural environment factors (e.g., altitude and soil type) and climate factors (e.g., temperature and precipitation) were the dominant factors that determine the spatial distribution of NEP. Compared with a single factor, the interaction of pairs of factors enhanced the influence strength on NEP. The natural environment factors indirectly affected NEP change through influencing human activities intensity and climatic factors. Human activities intensity played a medium indirectly negative effect on NEP, while climatic factors exerted strong direct and indirect positive influences on NEP. The contributions of human activity intensity, climatic factors, and natural environment on NEP change in the QLM were 33.5%, 62.3%, and 38.3%, respectively. Overall, warming and wetting shifts in meteorological conditions offset the negative impact of human activities on NEP in the QLM, and the QLM has acted as a growing carbon sink in the past 20 years.

Dense time series of remote sensing images with high spatio-temporal resolution are critical for monitoring land surface dynamics in heterogeneous landscapes. Spatio-temporal fusion is an effective solution to obtaining such time series images. Many spatio-temporal fusion methods have been developed for producing high spatial resolution images at frequent intervals by blending fine spatial images and coarse spatial resolution images. Previous studies have revealed that the accuracy of fused images depends not only on the fusion algorithm, but also on the input image pairs being used. However, the impact of input images dates on the fusion accuracy for time series with different temporal variation patterns remains unknown. In this paper, the impact of input image pairs on the fusion accuracy for monotonic linear change (MLC), monotonic non-linear change (MNLC), and non-monotonic change (NMC) time periods were evaluated, respectively, and the optimal selection strategies of input image dates for different situations were proposed. The 16-day composited NDVI time series (i.e., Collection 6 MODIS NDVI product) were used to present the temporal variation patterns of land surfaces in the study areas. To obtain sufficient observation dates to evaluate the impact of input image pairs on the spatio-temporal fusion accuracy, we utilized the Harmonized Landsat-8 Sentinel-2 (HLS) data. The ESTARFM was selected as the spatio-temporal fusion method for this study. The results show that the impact of input image date on the accuracy of spatio-temporal fusion varies with the temporal variation patterns of the time periods being fused. For the MLC period, the fusion accuracy at the prediction date (PD) is linearly correlated to the time interval between the change date (CD) of the input image and the PD, but the impact of the input image date on the fusion accuracy at the PD is not very significant. For the MNLC period, the fusion accuracy at the PD is non-linearly correlated to the time interval between the CD and the PD, the impact of the time interval between the CD and the PD on the fusion accuracy is more significant for the MNLC than for the MLC periods. Given the similar change of time intervals between the CD and the PD, the increments of R2 of fusion result for the MNLC is over ten times larger than those for the MLC. For the NMC period, a shorter time interval between the CD and the PD does not lead to higher fusion accuracies. On the contrary, it may lower the fusion accuracy. This study suggests that temporal variation patterns of the data must be taken into account when selecting optimal dates of input images in the fusion model.

Forests play a vital role in combating gradual developmental deficiencies and balancing regional ecosystems, yet they are constantly disturbed by man-made or natural events. Therefore, developing a timely and accurate forest disturbance detection strategy is urgently needed. The accuracy of traditional detection algorithms depends on the selection of thresholds or the formulation of complete rules, which inevitably reduces the accuracy and automation level of detection. In this paper, we propose a new multitemporal convolutional network framework (MT-CNN). It is an integrated method that can realize long-term, large-scale forest interference detection and distinguish the types (forest fire and harvest/deforestation) of disturbances without human intervention. Firstly, it uses the sliding window technique to calculate an adaptive threshold to identify potential interference points, and then a multitemporal CNN network is designed to render the disturbance types with various disturbance duration periods. To illustrate the detection accuracy of MT-CNN, we conducted experiments in a large-scale forest area (about 990 km2) on the west coast of the United States (including northwest California and west Oregon) with long time-series Landsat data from 1986 to 2020. Based on the manually annotated labels, the evaluation results show that the overall accuracies of disturbance point detection and disturbance type recognition reach 90%. Also, this method is able to detect multiple disturbances that continuously occurred in the same pixel. Moreover, we found that forest disturbances that caused forest fire repeatedly appear without a significant coupling effect with annual temporal and precipitation variations. Potentially, our method is able to provide large-scale forest disturbance mapping with detailed disturbance information to support forest inventory management and sustainable development.

Satellite-based models have tremendous potential for monitoring crop production because satellite data can provide temporally and spatially continuous crop growth information at large scale. This study used a satellite-based vegetation production model (i.e., eddy covariance light use efficiency, EC-LUE) to estimate national winter wheat gross primary production, and then combined this model with the harvest index (ratio of aboveground biomass to yield) to convert the estimated winter wheat production to yield. Specifically, considering the spatial differences of the harvest index, we used a cross-validation method to invert the harvest index of winter wheat among counties, municipalities and provinces. Using the field-surveyed and statistical yield data, we evaluated the model performance, and found the model could explain more than 50% of the spatial variations of the yield both in field-surveyed regions and most administrative units. Overall, the mean absolute percentage errors of the yield are less than 20% in most counties, municipalities and provinces, and the mean absolute percentage errors for the production of winter wheat at the national scale is 4.06%. This study demonstrates that a satellite-based model is an alternative method for crop yield estimation on a larger scale.

Although China is the largest producer of rice, accounting for about 25% of global production, there are no high-resolution maps of paddy rice covering the entire country. Using time-weighted dynamic time warping (TWDTW), this study developed a pixel- and phenology-based method to identify planting areas of double-season paddy rice in China, by comparing temporal variations of synthetic aperture radar (SAR) signals of unknown pixels to those of known double-season paddy rice fields. We conducted a comprehensive evaluation of the method’s performance at pixel and regional scales. Based on 145,210 field surveyed samples from 2018 to 2020, the producer’s and user’s accuracy are 88.49% and 87.02%, respectively. Compared to county-level statistical data from 2016 to 2019, the relative mean absolute errors are 34.11%. This study produced distribution maps of double-season rice at 10 m spatial resolution from 2016 to 2020 over nine provinces in South China, which account for more than 99% of the planting areas of double-season paddy rice of China. The maps are expected to contribute to timely monitoring and evaluating rice growth and yield.

Crop rotations, the farming practice of growing crops in sequential seasons, occupy a core position in agriculture management, showing a key influence on food security and agro-ecosystem sustainability. Despite the improvement in accuracy of identifying mono-agricultural crop distribution, crop rotation patterns remain poorly mapped. In this study, a hybrid convolutional neural network (CNN) and long short-term memory (LSTM) architecture, namely crop rotation mapping (CRM), were proposed to synergize the synthetic aperture radar (SAR) and optical time series in a rotational mapping task. The proposed end-to-end architecture had reasonable accuracies (i.e., accuracy > 0.85) in mapping crop rotation, which outperformed other state-of-the-art non-deep or deep-learning solutions. For some confusing rotation types, such as fallow-single rice and crayfish-single rice, CRM showed substantial improvements from traditional methods. Furthermore, the deeply synergistic SAR-optical, time-series data, with a corresponding attention mechanism, were effective in extracting crop rotation features, with an overall gain of accuracy of four points compared with ablation models. Therefore, our proposed method added wisdom to dynamic crop rotation mapping and yields important information for the agro-ecosystem management of the study area.

Human activities have dramatically changed ecosystems. As an irreplaceable ecological barrier in western China, the Qilian Mountains (QLM) provide various ecosystem services for humans. To evaluate the changes in the intensity of human activities in the QLM and their impact on the ecosystem, the human footprint (HF) method was used to conduct a spatial dataset of human activity intensity. In our study, the NDVI was used to characterize the growth of vegetation, and six categories of human pressures were employed to create the HF map in the QLM for 2000–2015 at a 1-km scale. The results showed that the mean NDVI during the growing season showed a significant increasing trend over the entire QLM in the period 2000–2015, while the NDVI showed a significant declining trend of more than 70% concentrated in Qinghai. Human pressure throughout the QLM occurred at a low level during 2000–2015, being greater in the eastern region than the western region, while the Qinghai area had greater human pressure than the Gansu area. Due to the improvement in traffic facilities, tourism, overgrazing, and other illegal activities, grasslands, shrublands, forests, wetlands, and bare land were the vegetation types most affected by human activities (in decreasing order). As the core area of the QLM, the Qilian Mountains National Nature Reserve (NR) has effectively reduced the impact of human activities. However, due to the existence of many ecological historical debts caused by unreasonable management in the past, the national park established in 2017 is facing great challenges to achieve its goals. These data and results will provide reference and guidance for future protection and restoration of the QLM ecosystem.

Timely and accurate agricultural information is essential for food security assessment and agricultural management. Synthetic aperture radar (SAR) systems are increasingly available in crop mapping, as they provide all-weather imagery. In particular, the Sentinel-1 sensor provides dense time-series data, thus offering a unique opportunity for crop mapping. However, in most studies, the Sentinel-1 complex backscatter coefficient was used directly which limits the potential of the Sentinel-1 in crop mapping. Meanwhile, most of the existing methods may not be tailored for the task of crop classification in time-series polarimetric SAR data. To solve the above problem, we present a novel deep learning strategy in this research. To be specific, we collected Sentinel-1 time-series data in two study areas. The Sentinel-1 image covariance matrix is used as an input to maintain the integrity of polarimetric information. Then, a depthwise separable convolution recurrent neural network (DSCRNN) architecture is proposed to characterize crop types from multiple perspectives and achieve better classification results. The experimental results indicate that the proposed method achieves better accuracy in complex agricultural areas than other classical methods. Additionally, the variable importance provided by the random forest (RF) illustrated that the covariance vector has a far greater influence than the backscatter coefficient. Consequently, the strategy proposed in this research is effective and promising for crop mapping.

Hyperspectral image analysis plays an important role in agriculture, mineral industry, and for military purposes. However, it is quite challenging when classifying high-dimensional hyperspectral data with few labeled samples. Currently, generative adversarial networks (GANs) have been widely used for sample generation, but it is difficult to acquire high-quality samples with unwanted noises and uncontrolled divergences. To generate high-quality hyperspectral samples, a self-attention generative adversarial adaptation network (SaGAAN) is proposed in this work. It aims to increase the number and quality of training samples to avoid the impact of over-fitting. Compared to the traditional GANs, the proposed method has two contributions: (1) it includes a domain adaptation term to constrain generated samples to be more realistic to the original ones; and (2) it uses the self-attention mechanism to capture the long-range dependencies across the spectral bands and further improve the quality of generated samples. To demonstrate the effectiveness of the proposed SaGAAN, we tested it on two well-known hyperspectral datasets: Pavia University and Indian Pines. The experiment results illustrate that the proposed method can greatly improve the classification accuracy, even with a small number of initial labeled samples.

Deep learning has become a standard processing procedure in land cover mapping for remote sensing images. Instead of relying on hand-crafted features, deep learning algorithms, such as Convolutional Neural Networks (CNN) can automatically generate effective feature representations, in order to recognize objects with complex image patterns. However, the rich spatial information still remains unexploited, since most of the deep learning algorithms only focus on small image patches that overlook the contextual information at larger scales. To utilize these contextual information and improve the classification performance for high-resolution imagery, we propose a graph-based model in order to capture the contextual information over semantic segments of the image. First, we explore semantic segments which build on the top of deep features and obtain the initial classification result. Then, we further improve the initial classification results with a higher-order co-occurrence model by extending the existing conditional random field (HCO-CRF) algorithm. Compared to the pixel- and object-based CNN methods, the proposed model achieved better performance in terms of classification accuracy.