Timely and accurate estimation of the area and distribution of crops is vital for food security. Optical remote sensing has been a key technique for acquiring crop area and conditions on regional to global scales, but great challenges arise due to frequent cloudy days in southern China. This makes optical remote sensing images usually unavailable. Synthetic aperture radar (SAR) could bridge this gap since it is less affected by clouds. The recent availability of Sentinel-1A (S1A) SAR imagery with a 12-day revisit period at a high spatial resolution of about 10 m makes it possible to fully utilize phenological information to improve early crop classification. In deep learning methods, one-dimensional convolutional neural networks (1D CNNs), long short-term memory recurrent neural networks (LSTM RNNs), and gated recurrent unit RNNs (GRU RNNs) have been shown to efficiently extract temporal features for classification tasks. However, due to the complexity of training, these three deep learning methods have been less used in early crop classification. In this work, we attempted to combine them with an incremental classification method to avoid the need for training optimal architectures and hyper-parameters for data from each time series. First, we trained 1D CNNs, LSTM RNNs, and GRU RNNs based on the full images’ time series to attain three classifiers with optimal architectures and hyper-parameters. Then, starting at the first time point, we performed an incremental classification process to train each classifier using all of the previous data, and obtained a classification network with all parameter values (including the hyper-parameters) at each time point. Finally, test accuracies of each time point were assessed for each crop type to determine the optimal time series length. A case study was conducted in Suixi and Leizhou counties of Zhanjiang City, China. To verify the effectiveness of this method, we also implemented the classic random forest (RF) approach. The results were as follows: (i) 1D CNNs achieved the highest Kappa coefficient (0.942) of the four classifiers, and the highest value (0.934) in the GRU RNNs time series was attained earlier than with other classifiers; (ii) all three deep learning methods and the RF achieved F measures above 0.900 before the end of growth seasons of banana, eucalyptus, second-season paddy rice, and sugarcane; while, the 1D CNN classifier was the only one that could obtain an F-measure above 0.900 for pineapple before harvest. All results indicated the effectiveness of the solution combining the deep learning models with the incremental classification approach for early crop classification. This method is expected to provide new perspectives for early mapping of croplands in cloudy areas.

In recent decades, plastic-mulched farmland has expanded rapidly in China as well as in the rest of the world because it results in marked increases of crop production. However, plastic-mulched farmland significantly influences the environment and has so far been inadequately investigated. Accurately monitoring and mapping plastic-mulched farmland is crucial for agricultural production, environmental protection, resource management, and so on. Monitoring plastic-mulched farmland using moderate-resolution remote sensing data is technically challenging because of spatial mixing and spectral confusion with other ground objects. This paper proposed a new scheme that combines spectral and textural features for monitoring the plastic-mulched farmland and evaluates the performance of a Support Vector Machine (SVM) classifier with different kernel functions using Landsat-8 Operational Land Imager (OLI) imagery. The textural features were extracted from multi-bands OLI data using a Grey Level Co-occurrence Matrix (GLCM) algorithm. Then, six combined feature sets were developed for classification. The results indicated that Landsat-8 OLI data are well suitable for monitoring plastic-mulched farmland; the SVM classifier with a linear kernel function is superior both to other kernel functions and to two other widely used supervised classifiers: Maximum Likelihood Classifier (MLC) and Minimum Distance Classifier (MDC). For the SVM classifier with a linear kernel function, the highest overall accuracy was derived from combined spectral and textural features in the 90° direction (94.14%, kappa 0.92), followed by the combined spectral and textural features in the 45° (93.84%, kappa 0.92), 135° (93.73%, kappa 0.92), 0° (93.71%, kappa 0.92) directions, and the spectral features alone (93.57%, kappa 0.91). Spectral features make a more significant contribution to monitoring the plastic-mulched farmland; adding textural features from medium resolution imagery provide only limited improvement in accuracy.

It is well known that timely crop growth monitoring and accurate crop yield estimation at a fine scale is of vital importance for agricultural monitoring and crop management. Crop growth models have been widely used for crop growth process description and yield prediction. In particular, the accurate simulation of important state variables, such as leaf area index (LAI) and root zone soil moisture (SM), is of great importance for yield estimation. Data assimilation is a useful tool that combines a crop model and external observations (often derived from remote sensing data) to improve the simulated crop state variables and consequently model outputs like crop total biomass, water use and grain yield. In spite of its effectiveness, applying data assimilation for monitoring crop growth at the regional scale in China remains challenging, due to the lack of high spatiotemporal resolution satellite data that can match the small field sizes which are typical for agriculture in China. With the accessibility of freely available images acquired by Sentinel satellites, it becomes possible to acquire data at high spatiotemporal resolution (10–30 m, 5–6 days), which offers attractive opportunities to characterize crop growth. In this study, we assimilated remotely sensed LAI and SM into the Word Food Studies (WOFOST) model to estimate winter wheat yield using an ensemble Kalman filter (EnKF) algorithm. The LAI was calculated from Sentinel-2 using a lookup table method, and the SM was calculated from Sentinel-1 and Sentinel-2 based on a change detection approach. Through validation with field data, the inverse error was 10% and 35% for LAI and SM, respectively. The open-loop wheat yield estimation, independent assimilations of LAI and SM, and a joint assimilation of LAI + SM were tested and validated using field measurement observation in the city of Hengshui, China, during the 2016–2017 winter wheat growing season. The results indicated that the accuracy of wheat yield simulated by WOFOST was significantly improved after joint assimilation at the field scale. Compared to the open-loop estimation, the yield root mean square error (RMSE) with field observations was decreased by 69 kg/ha for the LAI assimilation, 39 kg/ha for the SM assimilation and 167 kg/ha for the joint LAI + SM assimilation. Yield coefficients of determination (R²) of 0.41, 0.65, 0.50, and 0.76 and mean relative errors (MRE) of 4.87%, 4.32%, 4.45% and 3.17% were obtained for open-loop, LAI assimilation alone, SM assimilation alone and joint LAI + SM assimilation, respectively. The results suggest that LAI was the first-choice variable for crop data assimilation over SM, and when both LAI and SM satellite data are available, the joint data assimilation has a better performance because LAI and SM have interacting effects. Hence, joint assimilation of LAI and SM from Sentinel-1 and Sentinel-2 at a 20 m resolution into the WOFOST provides a robust method to improve crop yield estimations. However, there is still bias between the key soil moisture in the root zone and the Sentinel-1 C band retrieved SM, especially when the vegetation cover is high. By active and passive microwave data fusion, it may be possible to offer a higher accuracy SM for crop yield prediction.

Using plastic mulching for farmland is booming around the world. Despite its benefit of protecting crops from unfavorable conditions and increasing crop yield, the massive use of the plastic-mulching technique causes many environmental problems. Therefore, timely and effective mapping of plastic-mulched farmland (PMF) is of great interest to policy-makers to leverage the trade-off between economic profit and adverse environmental impacts. However, it is still challenging to implement remote-sensing-based PMF mapping due to its changing spectral characteristics with the growing seasons of crops and geographic regions. In this study, we examined the potential of multi-temporal Landsat-8 imagery for mapping PMF. To this end, we gathered the information of spectra, textures, indices, and thermal features into random forest (RF) and support vector machine (SVM) algorithms in order to select the common characteristics for distinguishing PMF from other land cover types. The experiment was conducted in Jizhou, Hebei Province. The results demonstrated that the spectral features and indices features of NDVI (normalized difference vegetation index), GI (greenness index), and textural features of mean are more important than the other features for mapping PMF in Jizhou. With that, the optimal period for mapping PMF is in April, followed by May. A combination of these two times (April and May) is better than later in the season. The highest overall, producer’s, and user’s accuracies achieved were 97.01%, 92.48%, and 96.40% in Jizhou, respectively.

To improve the accuracy of winter wheat yield estimation, the Crop Environment Resource Synthesis for Wheat (CERES-Wheat) model with an assimilation strategy was performed by assimilating measured or remotely-sensed leaf area index (LAI) values. The performances of the crop model for two different assimilation methods were compared by employing particle filters (PF) and the proper orthogonal decomposition-based ensemble four-dimensional variational (POD4DVar) strategies. The uncertainties of wheat yield estimates due to different assimilation temporal scales (phenological stages and temporal frequencies) and spatial scale were also analyzed. The results showed that, compared with the crop model without assimilation and with PF-based assimilation, a better yield estimate performance resulted when the POD4DVar-based strategy was used at the field scale. When using this strategy, root mean square errors (RMSE) of 523 kg·ha−1, 543 kg·ha−1 and 172 kg·ha−1 and relative errors (RE) of 5.65%, 5.91% and 7.77% were obtained at the field plot scale, a pixel scale of 1 km and the county scale, respectively. Although the best yield estimates were obtained when all of the observed LAIs were assimilated into the crop model, an acceptable estimate of crop yield could also be achieved by assimilating fewer observations between jointing and anthesis periods of the crop growth season. With decreasing assimilation frequency and pixel resolution, the accuracy of the crop yield estimates decreased; however, the computation time decreased. Thus, it is important to consider reasonable spatio-temporal scales to obtain tradeoffs between accuracy and effectiveness in regional wheat estimates.

Plastic mulching is an important technology in agricultural production both in China and the rest of the world. In spite of its benefit of increasing crop yields, the booming expansion of the plastic mulching area has been changing the landscape patterns and affecting the environment. Accurate and effective mapping of Plastic-Mulched Farmland (PMF) can provide useful information for leveraging its advantages and disadvantages. However, mapping the PMF with remote sensing is still challenging owing to its varying spectral characteristics with the crop growth and geographic spatial division. In this paper, we investigated the potential of Radarsat-2 data for mapping PMF. We obtained the backscattering intensity of different polarizations and multiple polarimetric decomposition descriptors. These remotely-sensed information was used as input features for Random Forest (RF) and Support Vector Machine (SVM) classifiers. The results indicated that the features from Radarsat-2 data have great potential for mapping PMF. The overall accuracies of PMF mapping with Radarsat-2 data were close to 75%. Although the classification accuracy with the back-scattering intensity information alone was relatively lower owing to the inherent speckle noise in SAR data, it has been improved significantly by introducing the polarimetric decomposition descriptors. The accuracy was nearly 75%. In addition, the features derived from the Entropy/Anisotropy/Alpha (H/A/Alpha) polarimetric decomposition, such as Alpha, entropy, and so on, made a greater contribution to PMF mapping than the Freeman decomposition, Krogager decomposition and the Yamaguchi4 decomposition. The performances of different classifiers were also compared. In this study, the RF classifier performed better than the SVM classifier. However, it is expected that the classification accuracy of PMF with SAR remote sensing data can be improved by combining SAR remote sensing data with optical remote sensing data.