Assessing canopy nitrogen content (CNC) and canopy carbon content (CCC) of maize by hyperspectral remote sensing data permits estimating cropland productivity, protecting farmland ecology, and investigating the nitrogen and carbon cycles in the atmosphere. This study aimed to assess maize CNC and CCC using canopy hyperspectral information and uninformative variable elimination (UVE). Vegetation indices (VIs) and wavelet functions were adopted for estimating CNC and CCC under varying water and nitrogen regimes. Linear, nonlinear, and partial least squares (PLS) regression models were fitted to VIs and wavelet functions to estimate CNC and CCC, and were evaluated for their prediction accuracy. UVE was used to eliminate uninformative variables, improve the prediction accuracy of the models, and simplify the PLS regression models (UVE-PLS). For estimating CNC and CCC, the normalized difference vegetation index (NDVI, based on red edge and NIR wavebands) yielded the highest correlation coefficients (r > 0.88). PLS regression models showed the lowest root mean square error (RMSE) among all models. However, PLS regression models required nine VIs and four wavelet functions, increasing their complexity. UVE was used to retain valid spectral parameters and optimize the PLS regression models. UVE-PLS regression models improved validation accuracy and resulted in more accurate CNC and CCC than the PLS regression models. Thus, canopy spectral reflectance integrated with UVE-PLS can accurately reflect maize leaf nitrogen and carbon status.

High-throughput maize phenotyping at both organ and plant levels plays a key role in molecular breeding for increasing crop yields. Although the rapid development of light detection and ranging (LiDAR) provides a new way to characterize three-dimensional (3D) plant structure, there is a need to develop robust algorithms for extracting 3D phenotypic traits from LiDAR data to assist in gene identification and selection. Accurate 3D phenotyping in field environments remains challenging, owing to difficulties in segmentation of organs and individual plants in field terrestrial LiDAR data. We describe a two-stage method that combines both convolutional neural networks (CNNs) and morphological characteristics to segment stems and leaves of individual maize plants in field environments. It initially extracts stem points using the PointCNN model and obtains stem instances by fitting 3D cylinders to the points. It then segments the field LiDAR point cloud into individual plants using local point densities and 3D morphological structures of maize plants. The method was tested using 40 samples from field observations and showed high accuracy in the segmentation of both organs (F-score =0.8207) and plants (F-score =0.9909). The effectiveness of terrestrial LiDAR for phenotyping at organ (including leaf area and stem position) and individual plant (including individual height and crown width) levels in field environments was evaluated. The accuracies of derived stem position (position error =0.0141 m), plant height (R² >0.99), crown width (R² >0.90), and leaf area (R² >0.85) allow investigating plant structural and functional phenotypes in a high-throughput way. This CNN-based solution overcomes the major challenges in organ-level phenotypic trait extraction associated with the organ segmentation, and potentially contributes to studies of plant phenomics and precision agriculture.

Leaf pigments are critical indicators of plant photosynthesis, stress, and physiological conditions. Inversion of radiative transfer models (RTMs) is a promising method for robustly retrieving leaf biochemical traits from canopy observations, and adding prior information has been effective in alleviating the “ill-posed" problem, a major challenge in model inversion. Canopy structure parameters, such as leaf area index (LAI) and average leaf inclination angle (ALA), can serve as prior information for leaf pigment retrieval. Using canopy spectra simulated from the PROSAIL model, we estimated the effects of uncertainty in LAI and ALA used as prior information for lookup table-based inversions of leaf chlorophyll (C_ab) and carotenoid (C_ar). The retrieval accuracies of the two pigments were increased by use of the priors of LAI (RMSE of C_ab from 7.67 to 6.32 μg cm⁻², C_ar from 2.41 to 2.28 μg cm⁻²) and ALA (RMSE of C_ab from 7.67 to 5.72 μg cm⁻², C_ar from 2.41 to 2.23 μg cm⁻²). However, this improvement deteriorated with an increase of additive and multiplicative uncertainties, and when 40% and 20% noise was added to LAI and ALA respectively, these priors ceased to increase retrieval accuracy. Validation using an experimental winter wheat dataset also showed that compared with C_ar, the estimation accuracy of C_ab increased more or deteriorated less with uncertainty in prior canopy structure. This study demonstrates possible limitations of using prior information in RTM inversions for retrieval of leaf biochemistry, when large uncertainties are present.

Spectroscopy can be used for detecting crop characteristics. A goal of crop spectrum analysis is to extract effective features from spectral data for establishing a detection model. An ideal spectral feature set should have high sensitivity to target parameters but low information redundancy among features. However, feature-selection methods that satisfy both requirements are lacking. To address this issue, in this study, a novel method, the continuous wavelet projections algorithm (CWPA), was developed, which has advantages of both continuous wavelet analysis (CWA) and the successive projections algorithm (SPA) for generating optimal spectral feature set for crop detection. Three datasets collected for crop stress detection and retrieval of biochemical properties were used to validate the CWPA under both classification and regression scenarios. The CWPA generated a feature set with fewer features yet achieving accuracy comparable to or even higher than those of CWA and SPA. With only two to three features identified by CWPA, an overall accuracy of 98% in classifying tea plant stresses was achieved, and high coefficients of determination were obtained in retrieving corn leaf chlorophyll content (R² = 0.8521) and equivalent water thickness (R² = 0.9508). The mechanism of the CWPA ensures that the novel algorithm discovers the most sensitive features while retaining complementarity among features. Its ability to reduce the data dimension suggests its potential for crop monitoring and phenotyping with hyperspectral data.

Plant height can be used for assessing plant vigor and predicting biomass and yield. Manual measurement of plant height is time-consuming and labor-intensive. We describe a method for measuring maize plant height using an RGB-D camera that captures a color image and depth information of plants under field conditions. The color image was first processed to locate its central area using the S component in HSV color space and the Density-Based Spatial Clustering of Applications with Noise algorithm. Testing showed that the central areas of plants could be accurately located. The point cloud data were then clustered and the plant was extracted based on the located central area. The point cloud data were further processed to generate skeletons, whose end points were detected and used to extract the highest points of the central leaves. Finally, the height differences between the ground and the highest points of the central leaves were calculated to determine plant heights. The coefficients of determination for plant heights manually measured and estimated by the proposed approach were all greater than 0.95. The method can effectively extract the plant from overlapping leaves and estimate its plant height. The proposed method may facilitate maize height measurement and monitoring under field conditions.

The wheat canopy reflectance spectrum is affected by many internal and external factors such as diseases and growth stage. Separating the effects of disease stress on the crop from the observed mixed signals is crucial for increasing the precision of remote sensing monitoring of wheat stripe rust. The canopy spectrum of winter wheat infected by stripe rust was processed with the difference-in-differences (DID) algorithm used in econometrics. The monitoring accuracies of wheat stripe rust before and after processing with the DID algorithm were compared in the presence of various external factors, disease severity, and several simulated satellite sensors. The correlation between the normalized difference vegetation index processed by the DID algorithm (NDVI-_DID) and the disease severity level (SL) increased in comparison with the NDVI before processing. The increase in precision in the natural disease area in the field in the presence of large differences in growth stage, growth, planting, and management of the crop was greater than that in the controlled experiment. For low disease levels (SL < 20%), the R² of the regression of NDVI-_DID on SL was 38.8% higher than that of the NDVI and the root mean square error (RMSE) was reduced by 11.1%. The increase in precision was greater than that for the severe level (SL > 40%). According to the measured hyperspectral data, the spectral reflectance of three satellite sensor levels was simulated. The wide-band NDVI was calculated. Compared with the wide-band NDVI and vegetation indexes (VI_S) before DID processing, there were increases in the correlation between SL and the various types of VI_S-_DID, as well as in the correlation between SL and NDVI-_DID. It is feasible to apply the DID algorithm to multispectral satellite data and diverse types of VI_S for monitoring wheat stripe rust. Our results improve the quantification of independent effects of stripe rust infection on canopy reflectance spectrum, increase the precision of remote sensing monitoring of wheat stripe rust, and provide a reference for remote sensing monitoring of other crop diseases.

Estimation of leaf chlorophyll content (LCC) by proximal sensing is an important tool for photosynthesis evaluation in high-throughput phenotyping. The temporal variability of crop biochemical properties and canopy structure across different growth stages has great impacts on wheat LCC estimation, known as growth stage effects. It will result in the heterogeneity of crop canopy at different growth stages, which would mask subtle spectral response of biochemistry variations. This study aims to explore spectral responses on the growth stage effects and establish LCC models suited for different growth stages. A total number of 864 pairwise samples of wheat canopy spectra and LCC values with 216 observations of each stage were sampled at the tillering, jointing, booting and heading stages in 2021. Firstly, statistical analysis of LCC and spectral response presented different distribution traits and typical spectral variations peak at 470, 520 and 680 nm. Correlation analysis between LCC and reflectance showed typical red edge shifts. Secondly, the testing model of partial least square (PLS) established by the entire datasets to validate the predictive performance at each stage yielded poor LCC estimation accuracy. The spectral wavelengths of red edge (RE) and blue edge (BE) shifts and the poor estimation capability motivated us to further explore the growth stage effects by establishing LCC models at respective growth periods. Finally, competitive adaptive reweighted sampling PLS (CARS-PLS), decision tree (DT) and random forest (RF) were used to select sensitive bands and establish LCC models at specific stages. Bayes optimisation was used to tune the hyperparameters of DT and RF regression. The modelling results indicated that CARS-PLS and DT did not extract specific wavelengths that could decrease the influences of growth stage effects. From the RF out-of-bag (OOB) evaluation, the sensitive wavelengths displayed consistent spectral shifts from BE to GP and from RE to RV from tillering to heading stages. Compared with CARS-PLS and DT, results of RF modelling yielded an estimation accuracy with deviation to performance (RPD) of 2.11, 2.02, 3.21 and 3.02, which can accommodate the growth stage effects. Thus, this study explores spectral response on growth stage effects and provides models for chlorophyll content estimation to satisfy the requirement of high-throughput phenotyping.

Spike number (SN) per unit area is one of the major determinants of grain yield in wheat. Development of high-throughput techniques to count SN from large populations enables rapid and cost-effective selection and facilitates genetic studies. In the present study, we used a deep-learning algorithm, i.e., Faster Region-based Convolutional Neural Networks (Faster R-CNN) on Red-Green-Blue (RGB) images to explore the possibility of image-based detection of SN and its application to identify the loci underlying SN. A doubled haploid population of 101 lines derived from the Yangmai 16/Zhongmai 895 cross was grown at two sites for SN phenotyping and genotyped using the high-density wheat 660K SNP array. Analysis of manual spike number (MSN) in the field, image-based spike number (ISN), and verification of spike number (VSN) by Faster R-CNN revealed significant variation (P < 0.001) among genotypes, with high heritability ranged from 0.71 to 0.96. The coefficients of determination (R²) between ISN and VSN was 0.83, which was higher than that between ISN and MSN (R² = 0.51), and between VSN and MSN (R² = 0.50). Results showed that VSN data can effectively predict wheat spikes with an average accuracy of 86.7% when validated using MSN data. Three QTL Qsnyz.caas-4DS, Qsnyz.caas-7DS, and QSnyz.caas-7DL were identified based on MSN, ISN and VSN data, while QSnyz.caas-7DS was detected in all the three data sets. These results indicate that using Faster R-CNN model for image-based identification of SN per unit area is a precise and rapid phenotyping method, which can be used for genetic studies of SN in wheat.

Powdery mildew is a disease that threatens wheat production and causes severe economic losses worldwide. Its timely diagnosis is imperative for preventing and controlling its spread. In this study, the multi-angle canopy spectra and disease severity of wheat were investigated at several developmental stages and degrees of disease severity. Four wavelength variable-selected algorithms: successive projection (SPA), competitive adaptive reweighted sampling (CARS), feature selection learning (Relief-F), and genetic algorithm (GA), were used to identify bands sensitive to powdery mildew. The wavelength variables selected were used as input variables for partial least squares (PLS), extreme learning machine (ELM), random forest (RF), and support vector machine (SVM) algorithms, to construct a suitable prediction model for powdery mildew. Spectral reflectance and conventional vegetation indices (VIs) displayed angle effects under several disease severity indices (DIs). The CARS method selected relatively few wavelength variables and showed a relatively homogeneous distribution across the 13 viewing zenith angles. Overall accuracies of the four modeling algorithms were ranked as follows: ELM (0.70-0.82) > PLS (0.63-0.79) > SVM (0.49-0.69) > RF (0.43-0.69). Combinations of features and algorithms generated varied accuracies, with coefficients of determination (R²) single-peaked at different observation angles. The constructed CARS-ELM model extracted a predictable bivariate relationship between the multi-angle canopy spectrum and disease severity, yielding an R² > 0.8 at each measured angle. Especially for larger angles, monitoring accuracies were increased relative to the optimal VI model (40% at −60°, 33% at +60°), indicating that the CARS-ELM model is suitable for extreme angles of −60° and +60°. The results are proposed to provide a technical basis for rapid and large-scale monitoring of wheat powdery mildew.

Spike development directly affects the yield and quality of rice. We describe an algorithm for automatically identifying multiple developmental stages of rice spikes (AI-MDSRS) that transforms the automatic identification of multiple developmental stages of rice spikes into the detection of rice spikes of diverse maturity levels. The scales vary greatly in different growth and development stages because rice spikes are dense and small, posing challenges for their effective and accurate detection. We describe a rice spike detection model based on an improved faster regions with convolutional neural network (Faster R-CNN). The model incorporates the following optimization strategies: first, Inception_ResNet-v2 replaces VGG16 as a feature extraction network; second, a feature pyramid network (FPN) replaces single-scale feature maps to fuse with region proposal network (RPN); third, region of interest (RoI) alignment replaces RoI pooling, and distance-intersection over union (DIoU) is used as a standard for non-maximum suppression (NMS). The performance of the proposed model was compared with that of the original Faster R-CNN and YOLOv4 models. The mean average precision (mAP) of the rice spike detection model was 92.47%, a substantial improvement on the original Faster R-CNN model (with 40.96% mAP) and 3.4% higher than that of the YOLOv4 model, experimentally indicating that the model is more accurate and reliable. The identification results of the model for the heading-flowering, milky maturity, and full maturity stages were within two days of the results of manual observation, fully meeting the needs of agricultural activities.

Accurate, efficient, and timely yield estimation is critical for crop variety breeding and management optimization. However, the contributions of proximal sensing data characteristics (spectral, temporal, and spatial) to yield estimation have not been systematically evaluated. We collected long-term, hyper-temporal, and large-volume light detection and ranging (LiDAR) and multispectral data to (i) identify the best machine learning method and prediction stage for wheat yield estimation, (ii) characterize the contribution of multisource data fusion and the dynamic importance of structural and spectral traits to yield estimation, and (iii) elucidate the contribution of time-series data fusion and 3D spatial information to yield estimation. Wheat yield could be accurately (R² = 0.891) and timely (approximately-two months before harvest) estimated from fused LiDAR and multispectral data. The artificial neural network model and the flowering stage were always the best method and prediction stage, respectively. Spectral traits (such as CIgreen) dominated yield estimation, especially in the early stage, whereas the contribution of structural traits (such as height) was more stable in the late stage. Fusing spectral and structural traits increased estimation accuracy at all growth stages. Better yield estimation was realized from traits derived from complete 3D points than from canopy surface points and from integrated multi-stage (especially from jointing to heading and flowering stages) data than from single-stage data. We suggest that this study offers a novel perspective on deciphering the contributions of spectral, structural, and time-series information to wheat yield estimation and can guide accurate, efficient, and timely estimation of wheat yield.

Most existing agronomic trait models of winter wheat vary across growing seasons, and it is an open question whether a unified statistical model can be developed to predict agronomic traits using a vegetation index (VI) across multiple growing seasons. In this study, we constructed a hierarchical linear model (HLM) to automatically adapt the relationship between VIs and agronomic traits across growing seasons and tested the model’s performance by sensitivity analysis. Results demonstrated that (1) optical VIs give poor performance in predicting AGB and PNC across all growth stages, whereas VIs perform well for LAI, LGB, LNC, and SPAD. (2) The sensitivity indices of the phenological information in the AGB and PNC prediction models were 0.81-0.86 and 0.66-0.73, whereas LAI, LGB, LNC, and SPAD prediction models produced sensitivity indexes of 0.01-0.02, 0.01-0.02, 0.01-0.02, and 0.02-0.08, respectively. (3) The AGB and PNC prediction models considering ZS were more accurate than the prediction models based on VI. Whether or not phenological information is used, there was no difference in model accuracy for LGB, LNC, SPAD, and LAI. This study may provide a guideline for deciding whether phenological correction is required for estimation of agronomic traits across multiple growing seasons.

Accurate information about phenological stages is essential for canola field management practices such as irrigation, fertilization, and harvesting. Previous studies in canola phenology monitoring focused mainly on the flowering stage, using its apparent structure features and colors. Additional phenological stages have been largely overlooked. The objective of this study was to improve a shape-model method (SMM) for extracting winter canola phenological stages from time-series top-of-canopy reflectance images collected by an unmanned aerial vehicle (UAV). The transformation equation of the SMM was refined to account for the multi-peak features of the temporal dynamics of three vegetation indices (VIs) (NDVI, EVI, and CI_red-edge). An experiment with various seeding scenarios was conducted, including four different seeding dates and three seeding densities. Three mathematical functions: asymmetric Gaussian function (AGF), Fourier function, and double logistic function, were employed to fit time-series vegetation indices to extract information about phenological stages. The refined SMM effectively estimated the phenological stages of canola, with a minimum root mean square error (RMSE) of 3.7 days for all phenological stages. The AGF function provided the best fitting performance, as it captured multiple peaks in the growth dynamics characteristics for all seeding date scenarios using four scaling parameters. For the three selected VIs, CI_red-edge achieved the greatest accuracy in estimating the phenological stage dates. This study demonstrates the high potential of the refined SMM for estimating winter canola phenology.

Screening for drought tolerance is critical to ensure high biomass production of bioenergy sorghum in arid or semi-arid environments. The bottleneck in drought tolerance selection is the challenge of accurately predicting biomass for a large number of genotypes. Although biomass prediction by low-altitude remote sensing has been widely investigated on various crops, the performance of the predictions are not consistent, especially when applied in a breeding context with hundreds of genotypes. In some cases, biomass prediction of a large group of genotypes benefited from multimodal remote sensing data; while in other cases, the benefits were not obvious. In this study, we evaluated the performance of single and multimodal data (thermal, RGB, and multispectral) derived from an unmanned aerial vehicle (UAV) for biomass prediction for drought tolerance assessments within a context of bioenergy sorghum breeding. The biomass of 360 sorghum genotypes grown under well-watered and water-stressed regimes was predicted with a series of UAV-derived canopy features, including canopy structure, spectral reflectance, and thermal radiation features. Biomass predictions using canopy features derived from the multimodal data showed comparable performance with the best results obtained with the single modal data with coefficients of determination (R²) ranging from 0.40 to 0.53 under water-stressed environment and 0.11 to 0.35 under well-watered environment. The significance in biomass prediction was highest with multispectral followed by RGB and lowest with the thermal sensor. Finally, two well-recognized yield-based drought tolerance indices were calculated from ground truth biomass data and UAV predicted biomass, respectively. Results showed that the geometric mean productivity index outperformed the yield stability index in terms of the potential for reliable predictions by the remotely sensed data. Collectively, this study demonstrated a promising strategy for the use of different UAV-based imaging sensors to quantify yield-based drought tolerance.

Estimating spatial variation in crop transpiration coefficients (CT_c) and aboveground biomass (AGB) rapidly and accurately by remote sensing can facilitate precision irrigation management in semiarid regions. This study developed and assessed a novel machine learning (ML) method for estimating CT_c and AGB using time-series unmanned aerial vehicle (UAV)-based multispectral vegetation indices (VIs) of maize under several irrigation treatments at the field scale. Four ML regression methods: multiple linear regression (MLR), support vector regression (SVR), random forest regression (RFR), and adaptive boosting regression (ABR), were used to address the complex relationship between CT_c and VIs. AGB was then estimated using exponential, logistic, sigmoid, and linear equations because of their clear mathematical formulations based on the optimal CT_c estimation model. The UAV VIs-derived CT_c using the RFR estimation model yielded the highest accuracy (R² = 0.91, RMSE = 0.0526, and nRMSE = 9.07%). The normalized difference red-edge index, transformed chlorophyll absorption in reflectance index, and simple ratio contributed significantly to the RFR-based CT_c model. The accuracy of AGB estimation using nonlinear methods was higher than that using the linear method. The exponential method yielded the highest accuracy (R² = 0.76, RMSE = 282.8 g m⁻², and nRMSE = 39.24%) in both the 2018 and 2019 growing seasons. The study confirms that AGB estimation models based on cumulative CT_c performed well under several irrigation treatments using high-resolution time-series UAV multispectral VIs and can support irrigation management with high spatial precision at a field scale.

Self-occlusions are common in rice canopy images and strongly influence the calculation accuracies of panicle traits. Such interference can be largely eliminated if panicles are phenotyped at the 3D level. Research on 3D panicle phenotyping has been limited. Given that existing 3D modeling techniques do not focus on specified parts of a target object, an efficient method for panicle modeling of large numbers of rice plants is lacking. This paper presents an automatic and nondestructive method for 3D panicle modeling. The proposed method integrates shoot rice reconstruction with shape from silhouette, 2D panicle segmentation with a deep convolutional neural network, and 3D panicle segmentation with ray tracing and supervoxel clustering. A multiview imaging system was built to acquire image sequences of rice canopies with an efficiency of approximately 4 min per rice plant. The execution time of panicle modeling per rice plant using 90 images was approximately 26 min. The outputs of the algorithm for a single rice plant are a shoot rice model, surface shoot rice model, panicle model, and surface panicle model, all represented by a list of spatial coordinates. The efficiency and performance were evaluated and compared with the classical structure-from-motion algorithm. The results demonstrated that the proposed method is well qualified to recover the 3D shapes of rice panicles from multiview images and is readily adaptable to rice plants of diverse accessions and growth stages. The proposed algorithm is superior to the structure-from-motion method in terms of texture preservation and computational efficiency. The sample images and implementation of the algorithm are available online. This automatic, cost-efficient, and nondestructive method of 3D panicle modeling may be applied to high-throughput 3D phenotyping of large rice populations.

This study aimed to set a computer-integrated multichannel spectral imaging system as a high-throughput phenotyping tool for the analysis of individual cowpea seeds harvested at different developmental stages. The changes in germination capacity and variations in moisture, protein and different sugars during twelve stages of seed development from 10 to 32 days after anthesis were non-destructively monitored. Multispectral data at 20 discrete wavelengths in the ultraviolet, visible and near infrared regions were extracted from individual seeds and then modelled using partial least squares regression and linear discriminant analysis (LDA) models. The developed multivariate models were accurate enough for monitoring all possible changes occurred in moisture, protein and sugar contents with coefficients of determination in prediction Rp2 of 0.93, 0.80 and 0.78 and root mean square errors in prediction (RMSEP) of 6.045%, 2.236% and 0.890%, respectively. The accuracy of PLS models in predicting individual sugars such as verbascose and stachyose was reasonable with Rp2 of 0.87 and 0.87 and RMSEP of 0.071% and 0.485%, respectively; but for the prediction of sucrose and raffinose the accuracy was relatively limited with Rp2 of 0.24 and 0.66 and RMSEP of 0.567% and 0.045%, respectively. The developed LDA model was robust in classifying the seeds based on their germination capacity with overall correct classification of 96.33% and 95.67% in the training and validation datasets, respectively. With these levels of accuracy, the proposed multichannel spectral imaging system designed for single seeds could be an effective choice as a rapid screening and non-destructive technique for identifying the ideal harvesting time of cowpea seeds based on their chemical composition and germination capacity. Moreover, the development of chemical images of the major constituents along with classification images confirmed the usefulness of the proposed technique as a non-destructive tool for estimating the concentrations and spatial distributions of moisture, protein and sugars during different developmental stages of cowpea seeds.

Mature soybean phenotyping is an important process in soybean breeding; however, the manual process is time-consuming and labor-intensive. Therefore, a novel approach that is rapid, accurate and highly precise is required to obtain the phenotypic data of soybean stems, pods and seeds. In this research, we propose a mature soybean phenotype measurement algorithm called Soybean Phenotype Measure-instance Segmentation (SPM-IS). SPM-IS is based on a feature pyramid network, Principal Component Analysis (PCA) and instance segmentation. We also propose a new method that uses PCA to locate and measure the length and width of a target object via image instance segmentation. After 60,000 iterations, the maximum mean Average Precision (mAP) of the mask and box was able to reach 95.7%. The correlation coefficients R2 of the manual measurement and SPM-IS measurement of the pod length, pod width, stem length, complete main stem length, seed length and seed width were 0.9755, 0.9872, 0.9692, 0.9803, 0.9656, and 0.9716, respectively. The correlation coefficients R2 of the manual counting and SPM-IS counting of pods, stems and seeds were 0.9733, 0.9872, and 0.9851, respectively. The above results show that SPM-IS is a robust measurement and counting algorithm that can reduce labor intensity, improve efficiency and speed up the soybean breeding process.

Plant vascular bundles are responsible for water and material transportation, and their quantitative and functional evaluation is desirable in plant research. At the single-plant level, the number, size, and distribution of vascular bundles vary widely, posing a challenge to automatically and accurately identifying and quantifying them. In this study, a deep learning-integrated phenotyping pipeline was developed to robustly and accurately detect vascular bundles in Computed Tomography (CT) images of stem internodes. Two semantic indicators were used to evaluate and identify a suitable feature extraction network for semantic segmentation models. The epidermis thickness of maize stem was evaluated for the first time and adjacent vascular bundles were improved using an adaptive watershed-based approach. The counting accuracy (R²) of vascular bundles was 0.997 for all types of stem internodes, and the measured accuracy of size traits was over 0.98. Combining sap flow experiments, multiscale traits of vascular bundles were evaluated at the single-plant level, which provided an insight into the water use efficiency of the maize plant.

High-resolution deep-learning-based remote-sensing imagery analysis has been widely used in land-use and crop-classification mapping. However, the influence of composite feature bands, including complex feature indices arising from different sensors on the backbone, patch size, and predictions in transferable deep models require further testing. The experiments were conducted in six sites in Henan province from 2019 to 2021. This study sought to enable the transfer of classification models across regions and years for Sentinel-2A (10-m resolution) and Gaofen PMS (2-m resolution) imagery. With feature selection and up-sampling of small samples, the performance of UNet++ architecture on five backbones and four patch sizes was examined. Joint loss, mean Intersection over Union (mIoU), and epoch time were analyzed, and the optimal backbone and patch size for both sensors were Timm-RegNetY-320 and 256 × 256, respectively. The overall accuracy and F₁ scores of the Sentinel-2A predictions ranged from 96.86% to 97.72% and 71.29% to 80.75%, respectively, compared to 75.34%-97.72% and 54.89%-73.25% for the Gaofen predictions. The accuracies of each site indicated that patch size exerted a greater influence on model performance than the backbone. The feature-selection-based predictions with UNet++ architecture and up-sampling of minor classes demonstrated the capabilities of deep-learning generalization for classifying complex ground objects, offering improved performance compared to the UNet, Deeplab V3+, Random Forest, and Object-Oriented Classification models. In addition to the overall accuracy, confusion matrices, precision, recall, and F₁ scores should be evaluated for minor land-cover types. This study contributes to large-scale, dynamic, and near-real-time land-use and crop mapping by integrating deep learning and multi-source remote-sensing imagery

Forecasting crop yields based on remote sensing data is one of the most important tasks in agriculture. Soybean is the main crop in the Russian Far East. It is desirable to forecast soybean yield as early as possible while maintaining high accuracy. This study aimed to investigate seasonal time series of the normalized difference vegetation index (NDVI) to achieve early forecasting of soybean yield. This research used data from the Moderate Resolution Image Spectroradiometer (MODIS), an arable-land mask obtained from the VEGA-Science web service, and soybean yield data for 2008-2017 for the Jewish Autonomous Region (JAR) districts. Four approximating functions were fitted to model the NDVI time series: Gaussian, double logistic (DL), and quadratic and cubic polynomials. In the period from calendar weeks 22-42 (end of May to mid-October), averaged over two districts, the model using the DL function showed the highest accuracy (mean absolute percentage error -4.0%, root mean square error (RMSE) -0.029, P < 0.01). The yield forecast accuracy of prediction in the period of weeks 25-30 in JAR municipalities using the parameters of the Gaussian function was higher (P < 0.05) than that using the other functions. The mean forecast error for the Gaussian function was 14.9% in week 25 (RMSE was 0.21 t ha⁻¹) and 5.1%−12.9% in weeks 26-30 (RMSE varied from 0.06 to 0.15 t ha⁻¹) according to the 2013-2017 data. In weeks 31-32, the error was 5.0%−5.4% (RMSE was 0.07 t ha⁻¹) using the Gaussian parameters and 7.4%−7.7% (RMSE was 0.09-0.11 t ha⁻¹) for the DL function. When the method was applied to municipal districts of other soy-producing regions of the Russian Far East. RMSE was 0.14-0.32 t ha⁻¹ in weeks 25-26 and did not exceed 0.20 t ha⁻¹ in subsequent weeks.

Spectral and spatial features in remotely sensed data play an irreplaceable role in classifying crop types for precision agriculture. Despite the thriving establishment of the handcrafted features, designing or selecting such features valid for specific crop types requires prior knowledge and thus remains an open challenge. Convolutional neural networks (CNNs) can effectively overcome this issue with their advanced ability to generate high-level features automatically but are still inadequate in mining spectral features compared to mining spatial features. This study proposed an enhanced spectral feature called Stacked Spectral Feature Space Patch (SSFSP) for CNN-based crop classification. SSFSP is a stack of two-dimensional (2D) gridded spectral feature images that record various crop types’ spatial and intensity distribution characteristics in a 2D feature space consisting of two spectral bands. SSFSP can be input into 2D-CNNs to support the simultaneous mining of spectral and spatial features, as the spectral features are successfully converted to 2D images that can be processed by CNN. We tested the performance of SSFSP by using it as the input to seven CNN models and one multilayer perceptron model for crop type classification compared to using conventional spectral features as input. Using high spatial resolution hyperspectral datasets at three sites, the comparative study demonstrated that SSFSP outperforms conventional spectral features regarding classification accuracy, robustness, and training efficiency. The theoretical analysis summarizes three reasons for its excellent performance. First, SSFSP mines the spectral interrelationship with feature generality, which reduces the required number of training samples. Second, the intra-class variance can be largely reduced by grid partitioning. Third, SSFSP is a highly sparse feature, which reduces the dependence on the CNN model structure and enables early and fast convergence in model training. In conclusion, SSFSP has great potential for practical crop classification in precision agriculture.

Accurate estimation of regional-scale crop yield under drought conditions allows farmers and agricultural agencies to make well-informed decisions and guide agronomic management. However, few studies have focused on using the crop model data assimilation (CMDA) method for regional-scale winter wheat yield estimation under drought stress and partial-irrigation conditions. In this study, we developed a CMDA framework to integrate remotely sensed water stress factor (MOD16 ET PET⁻¹) with the WOFOST model using an ensemble Kalman filter (EnKF) for winter wheat yield estimation at the regional scale in the North China Plain (NCP) during 2008-2018. According to our results, integration of MOD16 ET PET⁻¹ with the WOFOST model produced more accurate estimates of regional winter wheat yield than open-loop simulation. The correlation coefficient of simulated yield with statistical yield increased for each year and error decreased in most years, with r ranging from 0.28 to 0.65 and RMSE ranging from 700.08 to 1966.12 kg ha⁻¹. Yield estimation using the CMDA method was more suitable in drought years (r = 0.47, RMSE = 919.04 kg ha⁻¹) than in normal years (r = 0.30, RMSE = 1215.51 kg ha⁻¹). Our approach performed better in yield estimation under drought conditions than the conventional empirical correlation method using vegetation condition index (VCI). This research highlighted the potential of assimilating remotely sensed water stress factor, which can account for irrigation benefit, into crop model for improving the accuracy of winter wheat yield estimation at the regional scale especially under drought conditions, and this approach can be easily adapted to other regions and crops.

The timely and rapid mapping of rapeseed planting areas is desirable for national food security. Most current rapeseed mapping methods depend strongly on images with good observations obtained during the flowering stages. Although vegetation indices have been proposed to identify the rapeseed flowering stage in some areas, automatically mapping rapeseed planting areas in large regions is still challenging. We developed an automatic phenology- and pixel-based algorithm (APPA) by integrating Landsat 8 and Sentinel-1 satellite data. We found that the Normalized Rapeseed Flowering Index shows unique spectral characteristics during the flowering and post-flowering periods, which distinguish rapeseed parcels from other land-use types (urban, water, forest, grass, maize, wheat, barley, and soybean). To verify the robustness of APPA, we applied APPA to seven areas in five rapeseed-producing countries with flowering images unavailable. The rapeseed maps by APPA showed consistently high accuracies with producer accuracies of (0.87-0.93 and F-scores of 0.92-0.95 based on 4503 verification samples. They showed high spatial consistency at the pixel level with the land cover Scientific Expertise Centres (SEC) map in France, Crop Map of England in United Kingdom, national-scale crop- and land-cover map of Germany, and Annual Crop Inventory in Canada at the pixel level. We propose APPA as a highly promising method for automatically and efficiently mapping rapeseed areas.

Evaluating actual crop evapotranspiration (ET_c) variations and their determining factors under changing climates is crucial for agricultural irrigation management and crop productivity improvement in non-humid regions. This study analyzed the spatiotemporal characteristics and detected the determining factors of ET_c for winter wheat and summer maize rotation system from 2000 to 2017 in the North China Plain (NCP), by combining the FAO-56 dual crop coefficient approach with remotely sensed vegetation indices (VIs). The results indicated that daily air temperature increased in varying degrees while wind speed and sunshine hours decreased slightly during the growing season of winter wheat and summer maize over the study period. The trends of relative humidity and effective precipitation varied in crop growing seasons. Based on the validated relationship of dual crop coefficients and VIs, the estimated multi-year average ET_c of winter wheat (370.29 ± 31.28 mm) was much higher than summer maize (281.85 ± 20.14 mm), and the rotation cycle was 652.43 ± 27.67 mm. Annual ET_c of winter wheat and the rotation cycle increased by 2.96 mm a⁻¹ and 1.77 mm a⁻¹, respectively. However, the ET_c of summer maize decreased with distinct spatial variation. Spatially, winter wheat ET_c increased significantly in the northeast NCP, covering the Beijing-Tianjin-Hebei areas. Meanwhile, significant increases in summer maize ET_c were detected in the southwest NCP. The sensitivity and contribution analysis showed that ET_c of winter wheat and summer maize was positively sensitive to temperature, wind speed, and sunshine hours while negatively to relative humidity. Moreover, wind speed and sunshine hours contributed most to changes in ET_c (around 20%-40%).

Accurate crop distribution mapping is required for crop yield prediction and field management. Due to rapid progress in remote sensing technology, fine spatial resolution (FSR) remotely sensed imagery now offers great opportunities for mapping crop types in great detail. However, within-class variance can hamper attempts to discriminate crop classes at fine resolutions. Multi-temporal FSR remotely sensed imagery provides a means of increasing crop classification from FSR imagery, although current methods do not exploit the available information fully. In this research, a novel Temporal Sequence Object-based Convolutional Neural Network (TS-OCNN) was proposed to classify agricultural crop type from FSR image time-series. An object-based CNN (OCNN) model was adopted in the TS-OCNN to classify images at the object level (i.e., segmented objects or crop parcels), thus, maintaining the precise boundary information of crop parcels. The combination of image time-series was first utilized as the input to the OCNN model to produce an ‘original’ or baseline classification. Then the single-date images were fed automatically into the deep learning model scene-by-scene in order of image acquisition date to increase successively the crop classification accuracy. By doing so, the joint information in the FSR multi-temporal observations and the unique individual information from the single-date images were exploited comprehensively for crop classification. The effectiveness of the proposed approach was investigated using multitemporal SAR and optical imagery, respectively, over two heterogeneous agricultural areas. The experimental results demonstrated that the newly proposed TS-OCNN approach consistently increased crop classification accuracy, and achieved the greatest accuracies (82.68% and 87.40%) in comparison with state-of-the-art benchmark methods, including the object-based CNN (OCNN) (81.63% and 85.88%), object-based image analysis (OBIA) (78.21% and 84.83%), and standard pixel-wise CNN (79.18% and 82.90%). The proposed approach is the first known attempt to explore simultaneously the joint information from image time-series with the unique information from single-date images for crop classification using a deep learning framework. The TS-OCNN, therefore, represents a new approach for agricultural landscape classification from multi-temporal FSR imagery. Besides, it is readily generalizable to other landscapes (e.g., forest landscapes), with a wide application prospect.