Search

Plantation forests provide critical ecosystem services and have experienced worldwide expansion during the past few decades. Accurate mapping of tree species through remote sensing is critical for managing plantation forests. The typical temporal behaviors and traits of tree species in satellite image time series (SITS) generate temporal and spectral features in multiple phenological stages that are critical to improve tree species mapping. However, the diverse input features, sequential relations and complex structures in SITS drastically increase the dimension and difficulty of spectral-temporal feature extraction, which challenges the capacity of many general classifiers not explicitly adapted for spectral-temporal learning. As a result, there is still a lack of a method that could automatically extract spectral-temporal features with high separability and regional adaptability from high-dimensional SITS for tree species mapping of plantation forests. Moreover, the effects of varying temporal resolution and feature combination on the plantation tree species mapping are under-explored. Here, we developed a multi-head attention-based method for automatically extracting spectral-temporal features with high separability based on a modified Transformer network (Transformer4SITS) for improved plantation tree species mapping. The end-to-end network model consists of a feature extraction module to learn deep spectral-temporal features from SITS and a fusion module to combine multiple features for improving mapping accuracy. We applied this method to two representative plantation forests in southern and northern China for tree species mapping. The results show that: (1) Transformer4SITS method could self-adaptively extract typical spectral-temporal features of key phenological stages (e.g., greenness rising and falling) from SITS, and achieved significantly improved accuracies by at most 15% in comparison with all four baseline methods (i.e., long short-term memory, harmonic analysis, time-weighted dynamic time warping, linear discriminant analysis); (2) time series with higher temporal resolution tended to produce more accurate species maps consistently across two sites, with their overall accuracies (OA) respectively increasing from 91.05% and 84.33% (60-day) to 94.88% and 88.72% (5-day), but the effect of high temporal resolution respectively leveled off around 90-day and 50-day resolution across two sites; (3) the mapping results using all available bands and two-band spectral indices outperformed the results using a subset of them, but with only modest increase in the accuracy (i.e., OA increased from 93.63% and 86.01% to 94.88% and 88.72%. This study thus provides a state-of-the-art deep learning-based method for improved tree species mapping, which is critical for sustainable management and biodiversity monitoring of plantation forests across large scales. [ABSTRACT FROM AUTHOR]

Abstract This study aims to develop a deep learning based classification framework for remotely sensed time series. The experiment was carried out in Yolo County, California, which has a very diverse irrigated agricultural system dominated by economic crops. For the challenging task of classifying summer crops using Landsat Enhanced Vegetation Index (EVI) time series, two types of deep learning models were designed: one is based on Long Short-Term Memory (LSTM), and the other is based on one-dimensional convolutional (Conv1D) layers. Three widely-used classifiers were also tested for comparison, including a gradient boosting machine called XGBoost, Random Forest, and Support Vector Machine. Although LSTM is widely used for sequential data representation, in this study its accuracy (82.41%) and F1 score (0.67) were the lowest among all the classifiers. Among non-deep-learning classifiers, XGBoost achieved the best result with 84.17% accuracy and an F1 score of 0.69. The highest accuracy (85.54%) and F1 score (0.73) were achieved by the Conv1D-based model, which mainly consists of a stack of Conv1D layers and an inception module. The behavior of the Conv1D-based model was inspected by visualizing the activation on different layers. The model employs EVI time series by examining shapes at various scales in a hierarchical manner. Lower Conv1D layers of the optimized model capture small scale temporal variations, while upper layers focus on overall seasonal patterns. Conv1D layers were used as an embedded multi-level feature extractor in the classification model which automatically extracts features from input time series during training. The automated feature extraction reduces the dependency on manual feature engineering and pre-defined equations of crop growing cycles. This study shows that the Conv1D-based deep learning framework provides an effective and efficient method of time series representation in multi-temporal classification tasks. Highlights • Deep neural networks were developed for crop classification. • Deep neural network achieved 85.54% accuracy and an F1 score of 0.73. • The best non-deep-learning classifier achieved 84.17% accuracy and an F1 score of 0.69. • One-dimensional convolutional neural network was used as automated temporal feature extractor. • One-dimensional convolutional neural network identifies complex seasonal dynamics of economic crops. [ABSTRACT FROM AUTHOR]

For the two of the most important agricultural commodities, soybean and corn, remote sensing plays a substantial role in delivering timely information on the crop area for economic, environmental and policy studies. Traditional long-term mapping of soybean and corn is challenging as a result of the high cost of repeated training data collection, the inconsistency in image process and interpretation, and the difficulty of handling the inter-annual variability of weather and crop progress. In this study, we developed an automated approach to map soybean and corn in the state of Paraná, Brazil for crop years 2010–2015. The core of the approach is a decision tree classifier with rules manually built based on expert interaction for repeated use. The automated approach is advantageous for its capacity of multi-year mapping without the need to re-train or re-calibrate the classifier. Time series MODerate-resolution Imaging Spectroradiometer (MODIS) reflectance product (MCD43A4) were employed to derive vegetation phenology to identify soybean and corn based on crop calendar. To deal with the phenological similarity between soybean and corn, the surface reflectance of the shortwave infrared band scaled to a phenological stage was used to fully separate the two crops. Results suggested that the mapped areas of soybean and corn agreed with official statistics at the municipal level. The resultant map in the crop year 2012 was evaluated using an independent reference data set, and the overall accuracy and Kappa coefficient were 87.2% and 0.804 respectively. As a result of mixed pixel effect at the 500 m resolution, classification results were biased depending on topography. In the flat, broad and highly-cropped areas, uncultivated lands were likely to be identified as soybean or corn, causing over-estimation of cropland area. By contrast, scattered crop fields in mountainous regions with dense natural vegetation tend to be overlooked. For future mapping efforts, it has great potential to apply the automated mapping algorithm to other image series at various scales especially high-resolution images. [ABSTRACT FROM AUTHOR]

Land cover classification in remote sensing is often faced with the challenge of limited ground truth labels. Incorporating historical ground information has the potential to significantly lower the expensive cost associated with collecting ground truth and, more importantly, enable early- and in-season mapping that is helpful to many pre-harvest decisions. In this study, we propose a new approach that can effectively transfer knowledge about the topology (i.e. relative position) of different crop types in the spectral feature space (e.g. the histogram of SWIR1 vs RDEG1 bands) to generate labels, thereby supporting crop classification in a different year. Importantly, our approach does not attempt to transfer classification decision boundaries that are susceptible to inter-annual variations of weather and management, but relies on the more robust and shift-invariant topology information. We tested this approach for mapping corn/soybeans in the US Midwest, paddy rice/corn/soybeans in Northeast China and multiple crops in Northern France using Landsat-8 and Sentinel-2 data. Results show that our approach automatically generates high-quality labels for crops in the target year immediately after each image becomes available. Based on these generated labels from our approach, the subsequent crop type mapping using a random forest classifier can reach the F1 score as high as 0.887 for corn as early as the silking stage and 0.851 for soybean as early as the flowering stage and the overall accuracy of 0.873 in the test state of Iowa. In Northeast China, F1 scores of paddy rice, corn and soybeans and the overall accuracy can exceed 0.85 two and half months ahead of harvest. In the Hauts-de-France region, the OA of multiple crop mapping could reach 0.837 based on labels generated from our approach. Overall, these results highlight the unique advantages of our approach in transferring historical knowledge and maximizing the timeliness of crop maps. Our approach supports a general paradigm shift towards learning transferrable and generalizable knowledge to facilitate land cover classification. • Topology is shift-invariant and more stable over time than hard decision boundaries. • High-quality labels were generated by transferring historical topology knowledge. • Crop type mapping based on generated labels reached good performance. • Accurate crop type maps were generated in the US and China 2.5 months before harvest. • Reduced needs for labels to obtain moderately accurate crop type maps in France. [ABSTRACT FROM AUTHOR]

Abstract: An innovative phenology-based classification method was developed to map corn and soybean in multiple years using training data limited to a single year. Unlike traditional mapping efforts mainly based on multi-spectral image data, the classifier employed by this method takes phenological metrics as the major input. Phenological metrics represent crop characteristics related to crop calendar and progress such as the timing of emergence, maturity, harvest, etc. While considerable inter-annual variability exists among remotely sensed images from different years, phenological characteristics of each crop type are relatively consistent for a long period of time. Therefore, it is assumed that phenological metrics can be used to classify crop types in multiple years with the same rules, which is a valuable feature not possessed by traditional classification inputs. The classification method was tested in mapping corn and soybean, which are two dominant summer crop types in the central United States, and the experiment was carried out for Doniphan County, Kansas during years 2006–2010. Over 100 Landsat TM and ETM+images in this period were utilized and phenological metrics were calculated from Enhanced Vegetation Index time series using techniques including image segmentation and curve-fitting. Several sets of input variables, ranging from multi-spectral features of selected dates, which are widely used in traditional mapping efforts, to phenological metrics and derived measurements such as accumulated temperature, were tested using a random forest classifier. When the classifier was trained by reference data collected in the same year as that of remotely sensed data, most sets of input variables yielded accuracies higher than 88%. However, when the training data used by the classifier were obtained in a year different from the mapping years, only input sets containing phenological metrics were able to achieve acceptable accuracies greater than 80%. The use of phenological metrics as classification inputs avoids the restrictive requirements of a large ground reference dataset, enabling frequent and routine crop mapping without repeated collection of reference data. [Copyright &y& Elsevier]

An accurate and timely crop-type map is essential in water planning in California. So far, no effort has been made to effectively and efficiently identify specific crop types on an annual basis in this area. We have explored the potential of Moderate Resolution Imaging Spectroradiometer (MODIS) reflectance images to annually map major crop types in the San Joaquin Valley, California. A phenology-based classification approach has been employed, which has extracted phenological metrics from normalized difference vegetation index (NDVI) profiles and identified crop types based on these metrics using decision trees. According to a comparison with traditional maximum-likelihood classification, this phenology-based approach has shown great advantages when the size of the training set was limited by ground-truth availability and when the central tendency was absent in agricultural systems heavily influenced by human activities. [ABSTRACT FROM AUTHOR]

Accurately estimating vegetation biomass in urban forested areas is of great interest to researchers as it is a key indicator of the carbon sequestration capacity necessary for cities to achieve carbon neutrality. The emerging vegetation biomass estimation methods that use AI technologies with remote sensing images often suffer from arge estimating errors due to the diversity of vegetation and the complex three-dimensional terrain environment in urban ares. However, the high resolution of Light Detection and Ranging (i.e., LiDAR) data provides an opportunity to accurately describe the complex 3D scenes of urban forests, thereby improving estimation accuracy. Additionally, deep earning foundation models have widely succeeded in the industry, and show great potential promise to estimate vegetation biomass through processing complex and arge amounts of urban LiDAR data efficiently and accurately. In this study, we propose an efficient and accurate method called 3D-CiLBE (3DCity Long-term Biomass Estimation) to estimate urban vegetation biomass by utilizing advanced deep earning foundation models. In the 3D-CiLBE method, the Segment Anything Model (i.e., SAM) was used to segment single wood information from a arge amount of complex urban LiDAR data. Then, we modified the Contrastive Language–Image Pre-training (i.e., CLIP) model to identify the species of the wood so that the classic anisotropic growth equation can be used to estimate biomass. Finally, we utilized the Informer model to predict the biomass in the ong term. We evaluate it in eight urban areas across the United States. In the task of identifying urban greening areas, the 3D-CiLBE achieves optimal performance with a mean Intersection over Union (i.e., mIoU) of 0.94. Additionally, for vegetation classification, 3D-CiLBE achieves an optimal recognition accuracy of 92.72%. The estimation of urban vegetation biomass using 3D-CiLBE achieves a Mean Square Error of 0.045 kg/m2, reducing the error by up to 8.2% compared to 2D methods. The MSE for biomass prediction by 3D-CiLBE was 0.06kg/m2 smaller on average than the inear regression model. Therefore, the experimental results indicate that the 3D-CiLBE method can accurately estimate urban vegetation biomass and has potential for practical application. [ABSTRACT FROM AUTHOR]

Winter wheat is a major staple crop and it is critical to monitor winter wheat production using efficient and automated means. This study proposed a novel approach to produce winter wheat maps using statistics as the training targets of supervised classification. Deep neural network architectures were built to link remotely sensed image series to the harvested areas of individual administrative units. After training, the resultant maps were generated using the activations on a middle layer of the deep model. The direct use of statistical data to some extent alleviates the shortage of ground samples in classification tasks and provides an opportunity to utilize a wealth of statistical records to improve land use mapping. The experiments were carried out in Kansas and Northern Texas during 2001–2017. For each study area the goal was to create winter maps that are consistent with USDA county-level statistics of harvested areas. The trained deep models automatically identified the seasonal pattern of winter wheat pixels without using pixel-level reference data. The winter wheat maps were compared with the Cropland Data Layer (CDL) for years when the CDL is available. In Kansas where the winter wheat extent of the CDL has high reported accuracy and agrees well with county statistics, the maps produced from the deep model was evaluated using the CDL as an independent test set. Northern Texas was selected as an example where the winter wheat area of the CDL is very different from official statistics, and the maps by the deep model enabled a map-to-map comparison with the CDL to highlight the areas of discrepancy. Visual representation of the deep model behaviors and recognized patterns show that deep learning is an automated and robust means to handle the variability of winter wheat seasonality without the need of manual feature engineering and intensive ground data collection. Showing the possibility of generating maps solely from regional statistics, the proposed deep learning approach has great potential to fill the historical gaps of conventional sample-based classification and extend applications to areas where only regional statistics are available. The flexible deep network architecture can be fused with various statistical datasets to fully employ existing sources of data and knowledge. • Winter wheat maps were produced by deep learning models and per-county statistics. • The approach reduced the dependency on ground sample collection in crop mapping. • The fully convolutional architecture connects pixel-level data to statistics. • Pixel-level crop coverage was from a subset of the model trained with statistics. • The model automatically identified winter wheat seasonality without using samples. [ABSTRACT FROM AUTHOR]

Regional carbon sink estimation and local forest management require spatially explicit maps of bamboo distribution. However, accurate bamboo mapping is challenging due to the similarity of bamboo's optical spectral with those of other vegetation. To gain a high-precision bamboo forest distribution map circa year 2020, we developed a novel classification framework that integrated in-situ measurements, GEDI, and time-series multispectral Sentinel-2 images in a whale-optimized dual-channel Dense Convolutional Network (DenseNet) classification algorithm. The new method shows that vegetation structure features from GEDI can improve the performance of bamboo mapping by more than 5 percent compared with results obtained using only spectral information. Additionally, ablation studies show our proposed dual-channel DenseNet with whale-optimized DropBlock regularization outperforms traditional deep learning algorithms. The proposed method demonstrates a high level of accuracy in identifying bamboo, with precision, recall (producer's and user's accuracies), and F1-score reaching 90.81%, 91.86%, and 91.33%, respectively, which can provide more detailed and accurate results than previous studies, especially in regions where the terrain is complex or bamboo is intermixed with other vegetation. [ABSTRACT FROM AUTHOR]

The construction of contour levees for rice irrigation represents a major landscape management activity with impacts on irrigation water use efficiency, crop management decisions, and food production. However, levee distribution information traditionally relies on local field surveys because remote sensing approaches are complicated by irregular spacing, shape, and landscape variability within the field. In this paper the authors develop a deep learning approach capable of identifying rice fields with contour style levee irrigation practices from open-source aerial imagery. To generate a levee-identification scheme, a hybrid ResNet/Unet model is built from the commonly known Residual Network (ResNet) architecture for multi-layer deep learning strategies. The model takes a 320 × 320 RGB aerial landscape image from the US National Agricultural Imagery Program as input along with label data to then generate a probability map of the distribution of farm fields that use contour levees within the image. In performing this task, the model generates a 0.991 receiver operating characteristic curve score. The model continues to perform well under the introduction of clouds, data augmentation, or minor reductions in spatial resolution. Throughout these tests, the model performed within 0.2 of its original score, except for when the image quality was reduced to 60 m wherein the model score dropped to 0.691. Via these tests the model demonstrates potential to function well given different spatial extents or potential satellite remote sensing with moderate (10 m) resolutions. This model provides a proof-of-concept for the use of aerial imagery and a deep learning strategy for irrigation-type mapping practices. • Contour levee identification method using aerial images yields ROC score of 0.991. • A ResNet/UNet hybrid provides ample performance for remote sensing applications. • The method is insensitive to simulated cloud cover up to 40% opacity. • Performance degrades with a decrease in resolution from 1 m to less than 10 m. • The introduction of dense static noise could present a challenge for the model. [ABSTRACT FROM AUTHOR]

The goal of this study was to promptly map the extent of corn and soybeans early in the growing season. A classification experiment was conducted for the US Corn Belt and neighboring states, which is the most important production area of corn and soybeans in the world. To improve the timeliness of the classification algorithm, training was completely based on reference data and images from other years, circumventing the need to finish reference data collection in the current season. To account for interannual variability in crop development in the cross-year classification scenario, several innovative strategies were used. A random forest classifier was used in all tests, and MODIS surface reflectance products from the years 2008-2014 were used for training and cross-year validation. It is concluded that the fuzzy classification approach is necessary to achieve satisfactory results with R-squared ~0.9 (compared with the USDA Cropland Data Layer). The year of training data is an important factor, and it is recommended to select a year with similar crop phenology as the mapping year. With this phenology-based and cross-year-training method, in 2015 we mapped the cropping proportion of corn and soybeans around mid-August, when the two crops just reached peak growth. [ABSTRACT FROM AUTHOR]

It is not easy in remote sensing field to distinguish corn and soybean mapping for the similarity of the mixed summer crops. To understand the variations better and generate corn and soybean maps more accurately, more accurate mapping of these two crops is required. However, classifying different crops with remote sensing is not easy. Finding the discriminating features to use in a mapping classifier can lead to higher-precision mapping. In this paper, we used feature selection of random forests to analyse the most important phenological features for mapping corn and soybeans in the US Corn Belt (also called the Extended Corn Belt, ECB), and identified the spatial and temporal patterns of these features in the ECB. Results show that all the states in the ECB have very similar patterns: Temporal pattern showed that different years always reflected the same pattern. Date-related and phenology curve-related features were the most important. Also, the whole phenology curve was a prerequisite for more accurate mapping. These patterns make the variations more understandable and also support the further research to consider the most important variables for detailed classification. [ABSTRACT FROM AUTHOR]

Bamboo forest is a unique forest landscape that is mainly composed of herbal plants. It has a stronger capability to increase terrestrial carbon sinks than woody forests in the same environment, thus playing a special role in absorbing atmospheric CO2. Accurate and timely bamboo forest maps are necessary to better understand and quantify their contribution to the carbon and hydrological cycles. Previous studies have reported that the unique phenology pattern of bamboo forests, i.e., the on- and off-year cycle, can be detected with time-series high spatial resolution remote sensing (RS) images. Nevertheless, this information has not yet been applied in large-scale bamboo mapping. In this study, we innovatively incorporate newly designed phenology features reflecting the aforementioned on- and off-year cycles into a typical end-to-end classification workflow, including two features describing growing efficiency during the green-up season and two features describing the difference between annual peak greenness. Additionally, two horizonal morphology features and one tree height feature were employed, simultaneously. An experiment in southeast China was carried out to test the method's performance, in which seven categories were focused. A total of 987 field samples were used for training and validation (70% and 30%, respectively). The results show that combining the time-series features based on spectral bands and vegetation indices and newly designed phenology and morphology patterns can differentiate bamboo forests from other vegetation categories. Based on these features, the classification results exhibit a reasonable spatial distribution and a satisfactory overall accuracy (0.89). The detected bamboo area proportion in 82 counties agrees with the statistics from China's Third National Land Survey, which was produced based on high resolution images from commercial satellites and human interpretation (correlation coefficient = 0.69, and root mean squared error = 5.1%). This study demonstrates that the new scheme incorporating phenology features helps to map bamboo forests accurately while reducing the sample size requirement. [ABSTRACT FROM AUTHOR]

Mapping the corn dynamics at a large scale and multiple years is essential for global food security. Traditional mapping approaches by collecting training samples from field surveys are labor-intensive, challenging large-scale mapping of corn dynamics over the long term. This study developed an efficient approach to map large-scale corn dynamics in the main corn districts of the United States (US) using adaptive strategies for collecting high-quality training samples. First, this study proposed an automatic approach to collect stable and representative corn samples from the crop data layers (CDL) product. Then, this study improved the mapping performance of corn at a large scale and in earlier years by using adaptive strategies to collect limited (less than 500) but high-quality training samples. Finally, this study assessed and discussed the model performance across spaces and years using multi-source datasets from 2001 to 2020. Our results indicate the proposed approach with adaptive strategies can generate robust classification models with good performance in mapping large-scale corn dynamics. In our study area, the mean overall accuracy (OA) of corn is about 88% if using the CDL data as a reference. Besides, the R2 of corn areas at the county scale between our result and the surveyed acreage statistics is above 0.9, suggesting the proposed strategy is helpful for mapping corn dynamics across different years. The proposed approach demonstrated that limited but representative samples could map corn dynamics at a large scale with good performance, showing significant improvement compared to the traditional approach. It is feasible to map crop dynamics with multiple types by combining existing crop products with collected field samples, particularly in China, where crop products at the national scale are still lacking. [ABSTRACT FROM AUTHOR]

Compared with disturbance maps produced at annual or multi-year time steps, monthly mapping of forest harvesting can provide more temporal details needed for studying the socio-economic drivers (e.g., differentiating salvage logging and slash-and-burn from other timber harvesting) of harvesting and characterizing the associated intra-annual carbon and hydrological dynamics. Frequent cloud cover limits the application of optical remote sensing in timely mapping of forest changes. The freely available Sentinel-1 synthetic aperture radar (SAR) sensor provides an unprecedented opportunity to achieve more frequent mapping of forest harvesting than ever before (i.e., at monthly interval). The unique landscape pattern of forest harvesting from Sentienl-1 data (i.e., how a harvested patch contrasts to surrounding intact forests) holds critical information for harvesting mapping but have not been fully explored. In this study, we propose a deep learning-based (i.e., U-Net) approach using the landscape pattern from Sentinel-1 data to produce monthly maps of forest harvesting in two deforestation hotspots - California, USA and Rondônia, Brazil – for as long as three years. Our results show that (1) our proposed approach is reliable (mean F1 scores (the geometric mean of user's and producer's accuracies) 0.74–0.78; mean IoU (the area of intersection over union between the prediction part and target part) 0.59–0.65) for monthly forest harvesting mapping with Sentinel-1 data, outperforming the traditional object-based approach (0.38–0.43 in IoU). The varying harvesting pattern from Sentinel-1 data can be recognized by the U-Net bottleneck block as whole entities, which is the key advantage of our proposed approach; (2) multi-temporal SAR filtering is helpful for improving the accuracies of our proposed approach (increased F1 and IoU for 0.04 and 0.06, respectively); (3) our proposed model can be trained using samples collected during a particular time period over one location and be fine-tuned using sparse local samples from a new area to achieve optimal performance, and hence can greatly reduce training data collection effort when applied to new study sites; (4) forest harvesting maps produced using our approach revealed substantial variations in monthly harvesting activities: in Rondônia, most of the forest harvest occurred in July/August (the dry season) and about 14% of the dry season harvesting were followed by fires (i.e., slash-and-burn); in California, the rates of forest harvesting were relatively stable, but abnormally high values could occur due to salvage logging after big fires. Our novel approach for mapping forest harvesting at monthly interval represents an important step towards timely monitoring of forest harvesting and assisting stakeholders in developing sustainable strategy of forest management, especially for regions with frequent cloud cover. • A novel approach for monthly forest harvesting mapping from Sentinle-1 data. • A deep learning-based approach using the landscape pattern from Sentinel-1 data. • The proposed approach with high transferability between two study sites. • Implemented in the cloud computing environment as a full pipeline. • Produced forest harvesting maps revealed dynamics in monthly harvesting activities. [ABSTRACT FROM AUTHOR]

Land use and land cover maps provide fundamental information that has been used in different types of studies, ranging from public health to carbon cycling. However, the existing remote sensing image classification methods thus far suffer from the insufficient usage of multiple modalities, underconsideration of prior domain knowledge, and poor performance on minority classes. To alleviate these problems, we propose a novel domain knowledge-guided deep collaborative fusion network (DKDFN) with performance boosting for minority categories for land cover classification. More specifically, the DKDFN adopts a multihead encoder and a multibranch decoder structure. The architecture of the encoder probablizes sufficient mining of complementary information from multiple modalities, which are Sentinel-2, Sentinel-1, and SRTM Digital Elevation Data (SRTM) in our case. The multibranch decoder enables land cover classification in a multitask learning setup, performing semantic segmentation and reconstructing multimodal remote sensing indices, which are selected as representatives of domain knowledge. This design incorporates domain knowledge in an effective end-to-end manner. The training stage of our DKDFN is supervised by our proposed asymmetry loss function (ALF), which boosts performance on nearly all categories, especially the categories with a low frequency of occurrence. Ablation studies of the network suggest that our design logic is worth testing in any network with an encoder-decoder structure. The study is conducted in Hunan, China and is verified using a self-labeled multimodal unitemporal remote sensing image dataset. The comparative experiments between DKDFN and 6 state-of-the-art models (U-Net, SegNet, PSPNet, DeepLab, HRNet, MP-ResNet) testify to the superiority of our method and suggest its potential to be applied more widely to map land cover in other geographical areas given the availability of Sentinel-2, Sentinel-1, and SRTM data. The dataset can be downloaded by https://github.com/LauraChow/HunanMultimodalDataset. [ABSTRACT FROM AUTHOR]

Modeling in computer vision has long been dominated by convolutional neural networks (CNNs). Recently, in light of the excellent performance of self-attention mechanism in the language field, transformers tailored for visual data have drawn significant attention and triumphed over CNNs in various vision tasks. These vision transformers heavily rely on large-scale pre-training to achieve competitive accuracy, which not only hinders the freedom of architectural design in downstream tasks like object detection, but also causes learning bias and domain mismatch in the fine-tuning stages. To this end, we aim to get rid of the “pre-train and fine-tune” paradigm of vision transformer and train transformer based object detector from scratch. Some earlier works in the CNNs era have successfully trained CNNs based detectors without pre-training, unfortunately, their findings do not generalize well when the backbone is switched from CNNs to a vision transformer. Instead of proposing a specific vision transformer based detector, in this work, our goal is to reveal the insights of training vision transformer based detectors from scratch. In particular, we expect those insights to help other researchers and practitioners, and inspire more interesting research in other fields, such as remote sensing, visual-linguistic pre-training, etc. One of the key findings is that both architectural changes and more epochs play critical roles in training vision transformer based detectors from scratch. Experiments on the MS COCO dataset demonstrate that vision transformer based detectors trained from scratch can also achieve similar performance to their counterparts with ImageNet pre-training. [ABSTRACT FROM AUTHOR]

• The developed ACNN model outperforms RF and 1DCNN models in identifying spatial heterogeneity of wheat yields. • Coupling of 1D CNNs and multi-channel attention mechanism to improve feature learning. • Investigated the importance of key phenological stages and spectral bands for yield estimation. Accurately expressing the spatial pattern of crop yield at a sizeable regional field scale is paramount for precision agriculture. However, the current methodologies consistently face numerous challenges, comprising inadequate yield samples and algorithmic tendencies to underestimate higher yield values and overestimate lower ones. Data-driven deep learning algorithms effectively transform remotely sensed data into high-dimensional feature representations. Developing a cost-efficient model that minimizes the impact of insufficient field samples while maximizing the representation of crop yield differences on plots is worthwhile. The study developed an ACNN (Attention-based One-dimensional Convolutional Neural Network) model to efficiently extract and optimize the spatiotemporal features of winter wheat yield from Sentinel-2 MSI Level-2A data while relatively reducing the model's reliance on training samples. Winter wheat yield at both county and field scales in Henan Province was estimated utilizing Random Forest (RF), Convolutional Neural Network (CNN), and the ACNN models. The proposed ACNN model effectively identified key phenological stages and discerned noteworthy spectral bands essential for field-scale winter wheat yield estimation. It underscores the heading stage as a pivotal period influencing yield formation, emphasizing the indispensable contributions of the B8 (835.1 nm) and B12 (2202.4 nm) spectral bands in yield estimation. Moreover, this model adeptly overcame the challenge of underestimating higher values and overestimating lower values in crop yield estimation while showcasing robust generalization ability. The study presents a practical method for field-scale crop yield estimation over a large area with insufficient yield samples. [ABSTRACT FROM AUTHOR]

Long-term time series of spatially explicit cropland maps are essential for global crop modelling and climate change studies. The spatial resolution and temporal continuity of global cropland maps have been improving and several global data sets are released recently. Here, we calculated country-level cropland areas from the annual land-cover (LC) maps produced by the European Space Agency Climate Change Initiative (ESA-CCI) project and from the Food and Agricultural Organization of the United Nations statistical data (FAOSTAT) from 1992 to 2014. Because these two data sets used different approaches for generating the cropland data, we further quantified the consistency/difference in cropland areas and temporal changes between both products. Using log-transformed the time-averaged country-level cropland areas, a good linear relationship was found between these two products across different countries. However, only 8% of countries (mostly Organization for Economic Co-operation and Development countries) showed cropland area difference smaller than 1% between ESA-CCI and FAOSTAT. The cropland areas (without mosaic cropland types) from ESA-CCI are lower than the areas from FAOSTAT in 26% of countries but higher in 66% of countries. The magnitude of the latter difference (i.e. higher estimates of ESA-CCI than FAOSTAT) would be further amplified if crop intensity was taken into account. In addition, opposite temporal trends of cropland areas were found between these two data sets in 41% of countries. Although there are uncertainties in ESA-CCI LC maps, resulting from remote-sensing techniques such as mixed pixels, spectral similar objects, and same subject with different spectrum, the long time series and relatively high resolution of this product help us to understand the differences between satellite-based and inventory-based data sets and thus identify the possible strategies to improve the accuracy of satellite-based LC products. [ABSTRACT FROM AUTHOR]

The relationship between vegetation index (VI) and climatic variables such as temperature (TEP) and precipitation (PRE) at local, regional and global scales are conventionally analysed to understand the responses of vegetation to climate change. Those unique responses also afford opportunities for using climate variables to discriminate vegetation types. This paper presents a data-driven analysis to explore correlations between ten monthly climatic variables (temperature, precipitation, potential evapotranspiration (PET), vapour pressure (VAP), wet days (WET), and others) and monthly VIs of four different crop types (maize, rice, soybeans, and wheat) at global scale. The purpose is to show the VI–climate correlations in a spatially explicit way, laying the foundation for better crop type mapping by integrating climatic variables and remote sensing. The results show large variations in VI–climate correlation for different crop types and regions. Most cropland areas in the world show strong correlations between VI and VAP, and other variables such as WET, PET, and monthly average daily minimum temperature (TMN). This result encourages future studies using additional climate variables (in addition to TMP and PRE) for detailed vegetation/crop-type mapping. [ABSTRACT FROM AUTHOR]

Periodically inundated wetlands with high short-term surface variation require special approaches to assess their composition and long-term change. To circumvent high uncertainty in single-date analyses of such areas, we propose to characterize them as dynamic cover types (DCTs), or sequences of wetland states and transitions informed by physically and ecologically plausible surface processes. This study delineated DCTs for one 2007–2008 flood cycle at Poyang Lake, the largest freshwater wetland in China, using spatial and temporal orientation modes of extended principal components analysis (EPCA) and supervised object-based classification of multi-spectral and radar image series. Classification accuracy was compared among three sets of attributes selected by machine-learning optimization from object-level mean and standard deviations of: 1) image time series alone; 2) the most informative EPCA outputs alone and 3) image time series and EPCA results together. Classification uncertainty was additionally assessed as low values of object's maximum class membership (< 0.5). The highest accuracy was achieved with a larger set of 33 attributes selected from combined time series and EPCA results (overall accuracy 95.0%, kappa 0.94); however, accuracies with smaller sets of variables from input image series or EPCA results alone were comparably high (93.1% and 94.7%, respectively). All three selected attribute sets included standard deviations of image and/or EPCA values, suggesting the utility of object texture in dynamic class discrimination. The highest classification uncertainty was observed primarily along the mapped class boundaries, in some cases indicating minor change trajectories for which prior reference data were not available. Results indicate that DCTs provide a reasonable classification framework for complex and variable Poyang Lake wetlands that can be facilitated by EPCA transformation of complementary remote sensing time series. Future work should test this approach over multiple change cycles and assess sensitivity of results to temporal frequency of input image series, alternative variable selection algorithms and other remote sensors. [ABSTRACT FROM AUTHOR]

We report on a global cropland extent product at 30-m spatial resolution developed with two 30-m global land cover maps (i.e. FROM-GLC, Finer Resolution Observation and Monitoring, Global Land Cover; FROM-GLC-agg) and a 250-m cropland probability map. A common land cover validation sample database was used to determine optimal thresholds of cropland probability in different parts of the world to generate a cropland/noncropland mask according to the classification accuracies for cropland samples. A decision tree was then applied to combine two 250-m cropland masks: one existing mask from the literature and the other produced in this study, with the 30-m global land cover map FROM-GLC-agg. For the smallest difference with country-level cropland area in Food and Agriculture Organization Corporate Statistical (FAOSTAT) database, a final global cropland extent map was composited from the FROM-GLC, FROM-GLC-agg, and two masked cropland layers. From this map FROM-GC (Global Cropland), we estimated the global cropland areas to be 1533.83 million hectares (Mha) in 2010, which is 6.95 Mha (0.45%) less than the area reported by the Food and Agriculture Organization (FAO) of the United Nations for the year 2010. A country-by-country comparison between the map and the FAOSTAT data showed a linear relationship (FROM-GC = 1.05*FAOSTAT −1.2 (Mha) withR2= 0.97). Africa, South America, Southeastern Asia, and Oceania are the regions with large discrepancies with the FAO survey. [ABSTRACT FROM PUBLISHER]