To address insufficient feature extraction due to individual deep learning models’ limitations in capturing local and global features, and the tendency to underestimate high yields and overestimate low yields, a new deep learning model called convolutional neural network (CNN)-transformer with serial connection (CNN-transformers) was introduced to estimate winter wheat yield by combining the local feature extraction strengths of CNNs with the global information extraction abilities of transformer networks utilizing self-attention mechanisms. The remote sensing technology included temperature and spectral response indicators; the vegetation temperature condition index (VTCI), leaf area index (LAI), and fraction of photosynthetically active radiation (FPAR) aggregated over ten-day periods. Compared to the CNN model, the transformer model, the CNN-transformer model with parallel connection (CNN-transformerp), and the transformer-CNN model with serial connection (transformer-CNNs), the CNN-transformers achieved higher accuracy in estimating winter wheat yield (${R} ^{2} =0.70$, RMSE =420.39 kg/ha, MAPE =7.65%), which was capable of extracting more information related to yield from various remotely sensed parameters and addressing the problems of high yield underestimation and low yield overestimation observed. The robustness and generalization of the CNN-transformers were further assessed through the fivefold cross-validation and the leave-one-year-out methods. In addition, utilizing the CNN-transformers, the study uncovered the cumulative impact of the winter wheat growth period, examined how incrementally adding data at ten-day intervals affects yield estimation, and assessed the model’s proficiency in depicting growth accumulation during the growth process. The findings indicated that the model well identified the crucial growth phase of winter wheat between late March and early May.

Accurate crop yield estimation in advance is of paramount importance for making appropriate decisions in adjusting food price and formulating food policy. Most traditional statistical yield estimation models directly capture nonlinear relationship between the original time series remotely sensed variables and yields, ignoring the complexity and non-stationarity in the remotely sensed variables. This study integrated variational mode decomposition (VMD), correlation analysis and neural network architecture to construct a Decomposition-Screening-Reconstruction-Estimation paradigm for wheat yield estimation. Firstly, VMD optimized by whale optimization algorithm (WOA) was applied to decompose the time series remotely sensed variables into intrinsic mode function (IMF) components. Secondly, effective IMF components with strong correlation to the original time series remotely sensed variables were identified and reconstructed as the key time series features. Finally, the performances under the proposed paradigm of Transformer model and baseline models including support vector regression (SVR), random forest (RF) and long-short term memory (LSTM) were compared. The results showed that WOA-VMD-Transformer (WVTF) (R2= 0.66, RMSE = 459.09 kg/ha, MRE = 8.22 %) outperformed other models, and deep learning models (LSTM and Transformer) based on the new paradigm exhibited superior performance compared with traditional models (SVR and RF). In addition, WVTF (R2=0.54, RMSE=468.19 kg/ha, MRE=6.13 %) provided better performance at the sampling sites compared with previous proposed model (R2= 0.45, RMSE = 738.63 kg/ha, MRE = 8.21 %). The estimated wheat yields of 2019-2024 based on the optimal model were identical to the distribution of actual yields. In conclusion, the framework for yield estimation under novel paradigm extracts the original remotely sensed variables into key time series features, which provides an important reference for regional crop yield estimation and agricultural development.

Precise crop yield estimation is crucial for maintaining national food security. Crop growth models, statistical regression methods, and deep learning methods are among the numerous methods available for estimating crop yields. Deep learning, with its robust feature extraction capability and data-driven features, has propelled agriculture forward by leaps and bounds, and one of the applications oriented to national needs is the agricultural application of deep learning-based AI technology. However, as research and applications deepen and grow, the limitations of data-driven methodologies have emerged, such as difficulties in migrating reuse, reliance on samples, and poor interpretability. The organic integration of knowledge in various forms of expression and deep learning is the way to realize the dual drive of knowledge and data, and the integration of the two has not been well investigated. Therefore, we proposed a knowledge-guided deep learning (KGDL) framework that integrated knowledge with CNN and BiLSTM networks from both feature-level and network-level perspectives for winter wheat yield estimate in the Guanzhong Plain from 2007 to 2021. Our results showed that incorporating feature-level and network-level knowledge significantly enhanced wheat yield estimates, achieving R2= 0.73, RMSE = 447.12 kg/ha, and MAPE = 8.72%, reducing uncertainty by 109.88 kg/ha for the high-yield dataset, 81.73 kg/ha for the medium-yield dataset and 55.34 kg/ha for the low-yield dataset. Furthermore, the KGDL model, driven by both knowledge and data, resulted in a gradual increase in the coefficients between the output and the target yield on a layer-by-layer basis, with the average coefficients increasing from 0.35 to 0.58. Overall, the proposed framework shows promising application prospects for deep learning-based crop yield estimation, which can provide important technical support for promoting precision production in regional agriculture.

Due to climate change, the impact of drought on field crop production is extremely important. This study focuses on the vegetation temperature condition index (VTCI), an index-based drought monitoring index that can characterize drought conditions in near real time (at ten-day intervals), and explores the applicability of different spatial and temporal data fusion schemes to it. It also proposes a field-scale VTCI fusion framework based on the Sentinel-3 VTCI calculation and the land surface temperature (LST) downscaling. First, based on analyzing the computational characteristics of VTCI, multiyear VTCI based on Sentinel data sources was obtained, which further expands the diversity of data sources for VTCI. On this basis, a combination of qualitative and quantitative methods was used to compare the applicability of two schemes: Scheme 1, based on the “blend-then-index” (BI) strategy, which first fuses normalized difference vegetation index (NDVI) and LST, and then calculated the fused VTCIs; and Scheme 2, based on the “index-then-blend” (IB) strategy, which directly fuses the VTCIs based on the calculated VTCIs. It was found that all the fused VTCIs remained highly correlated with the ten-day cumulative precipitation. Compared with the fused VTCIs obtained by Scheme 2, the VTCIs obtained by Scheme 1 were able to display more spatial details. In addition, the VTCIs of Scheme 1 were more consistent with the Sentinel-3 VTCIs, and the accuracy of field yield estimation using the fused VTCIs was higher ($r$of 0.58 and root-mean-square error (RMSE) of 783.27 kg/ha).

The Sea and Land Surface Temperature Radiometer (SLSTR) onboard the two Sentinel-3 satellites provides daily global coverage observation at daytime and nighttime. The split-window (SW) algorithm is currently used to retrieve the land surface temperature (LST) from SLSTR images; however, this algorithm has to utilize visible and near-infrared (VNIR) images and land cover to determine pixel emissivity. For nighttime observation, VNIR cannot be observed, and this limitation complicates the LST retrieval from nighttime images using the SW algorithm. This article proposed a three-channel temperature-emissivity separation (TES) algorithm that estimates the nighttime LST and emissivity from one middle-infrared (MIR) and two thermal-infrared (TIR) nighttime Sentinel-3 SLSTR images. The sensitive analysis showed that the algorithm could theoretically retrieve the LST and emissivity with errors less than 0.8 K and 0.015, respectively. Ground validation showed that the nighttime LST retrieval error was approximately 1.84 K and the bias was approximately -0.33 K. Finally, the TES algorithm was applied to obtain the LST and emissivity images over northern China as an example. The emissivity retrieved from the nighttime observation can be used in the daytime SW algorithm to improve its feasibility in the LST retrieval process.

TanDEM-X forms together with TerraSAR-X the first single-pass polarimetric interferometer in space. This allows for the first time the acquisition and analysis of Single-, Dual-, and Quad-Pol-InSAR data without the disturbing effect of temporal decorrelation globally. For this reason, the exploration of TanDEM-X data for forestry is constantly increasing especially concerning forest height estimation, biomass classification and structure characterization. This paper reports the results of recent experiments aimed at investigating the potentials of TanDEM-X in characterizing quantitatively the spatial variability of the canopy top and phase center height, which is a proxy to horizontal structure. It is shown that such characterization can allow to differentiate among e.g. different successional and / disturbance stages in tropical forests.

The ESA Fire Essential Climate Variable provides a 1km pixel and 0.5 degree grid burned area product by merging several burned area data sets to a set of user requirements for the climate, atmospheric and ecological modeling communities. A description of the merging is given from preliminary results during the first phase of the project. Allowing for errors, some contributed to by ongoing algorithm development, results are reported for a test site in northern Australia. In the pixel product 50-60% of the burned areas corresponded in at least two data sets allowing for an improvement on the earliest detection date and the confidence level of a burn. The grid product shows good spatial and temporal consistency in burned area whilst providing further information on measures of confidence, cloudiness, burned area patchiness and dominant vegetation burned, useful information for modeling burning dynamics.

Understanding the spatial and temporal variation in carbon fluxes is essential to constrain models that predict climate change. However, our current knowledge of spatial and temporal patterns is uncertain, particularly over land. The ESA GLOBCARBON project aims to generate estimates of at-land products quasi-independent of the original Earth Observation source for use in Dynamic Global Vegetation Models, a central component of the ESSP Global Carbon Project. The service features global estimates of: burned area, fAPARS, LAI and vegetation growth cycle. The demonstrator focused on six complete years, from 1998 to 2003 when overlap exists between ESA Earth Observation sensors (ATSR-2, AATSR and MERIS) and VEGETATION but has recently been extended to 2007. This paper presents early results of the first re-processing in the GLOBCARBON project, which was undertaken after comments from users involved in beta testing.

Daily global mosaics of data acquired by the VEGETATION instrument onboard the SPOT 4 between Nov. 1st, 1999 and Dec. 31st, 2000 have been used to produce a reference high quality global land cover map (GLC 2000) and the global survey of burned surfaces (GBA 2000). These products have been generated through an international partnership coordinated by the Joint Research Centre and are now freely available to users. This paper reviews briefly the results achieved and provides an assessment of the specific properties of the VEGETATION data for land cover and burn scar mapping at continental to global scales.

An algorithm has been developed to map burnt areas over the Australian continent using SPOT-VEGETATION (VGT) S1 satellite images. The algorithm is composed of a set of thresholds applied to each pixel's value of the VGT spectral channels, two spectral indices and their temporal difference. The threshold values have been derived by means of a supervised classification methodology based on the classification and regression trees algorithm. A procedure has also been developed specifically for preprocessing the daily S1 images for burnt area mapping purposes. The final product is composed of ten-day and monthly burnt area maps over Australia for the full year 2000.