Unsupervised Plot-Scale LAI Phenotyping via UAV-Based Imaging, Modelling, and Machine Learning

UAV-Assisted Dynamic Monitoring of Wheat Uniformity toward Yield and Biomass Estimation

Crop uniformity is a comprehensive indicator used to describe crop growth and is important for assessing crop yield and biomass potential. However, there is still a lack of continuous monitoring of uniformity throughout the growing season to explain their effects on yield and biomass. Therefore, this paper proposed a wheat uniformity quantification method based on unmanned aerial vehicle imaging technology to monitor and analyze the dynamic changes in wheat uniformity. The leaf area index (LAI), soil plant analysis development (SPAD), and fractional vegetation cover were estimated from hyperspectral images, while plant height was estimated by a point cloud model from RGB images. Based on these 4 agronomic parameters, a total of 20 uniformity indices covering multiple growing stages were calculated. The changing trends in the uniformity indices were consistent with the results of visual interpretation. The uniformity indices strongly correlated with yield and biomass were selected to construct multiple linear regression models for estimating yield and biomass. The results showed that Pielou's index of LAI had the strongest correlation with yield and biomass, with correlation coefficients of -0.760 and -0.801, respectively. The accuracies of the yield (coefficient of determination [R 2] = 0.616, root mean square error [RMSE] = 1.189 Mg/ha) and biomass estimation model (R 2 = 0.798, RMSE = 1.952 Mg/ha) using uniformity indices were better than those of the models using the mean values of the 4 agronomic parameters. Therefore, the proposed uniformity monitoring method can be used to effectively evaluate the temporal and spatial variations in wheat uniformity and can provide new insights into the prediction of yield and biomass.

Simulation of Automatically Annotated Visible and Multi-/Hyperspectral Images Using the Helios 3D Plant and Radiative Transfer Modeling Framework

Deep learning and multimodal remote and proximal sensing are widely used for analyzing plant and crop traits, but many of these deep learning models are supervised and necessitate reference datasets with image annotations. Acquiring these datasets often demands experiments that are both labor-intensive and time-consuming. Furthermore, extracting traits from remote sensing data beyond simple geometric features remains a challenge. To address these challenges, we proposed a radiative transfer modeling framework based on the Helios 3-dimensional (3D) plant modeling software designed for plant remote and proximal sensing image simulation. The framework has the capability to simulate RGB, multi-/hyperspectral, thermal, and depth cameras, and produce associated plant images with fully resolved reference labels such as plant physical traits, leaf chemical concentrations, and leaf physiological traits. Helios offers a simulated environment that enables generation of 3D geometric models of plants and soil with random variation, and specification or simulation of their properties and function. This approach differs from traditional computer graphics rendering by explicitly modeling radiation transfer physics, which provides a critical link to underlying plant biophysical processes. Results indicate that the framework is capable of generating high-quality, labeled synthetic plant images under given lighting scenarios, which can lessen or remove the need for manually collected and annotated data. Two example applications are presented that demonstrate the feasibility of using the model to enable unsupervised learning by training deep learning models exclusively with simulated images and performing prediction tasks using real images.

Handling the Challenges of Small-Scale Labeled Data and Class Imbalances in Classifying the N and K Statuses of Rubber Leaves Using Hyperspectroscopy Techniques

The nutritional status of rubber trees (Hevea brasiliensis) is inseparable from the production of natural rubber. Nitrogen (N) and potassium (K) levels in rubber leaves are 2 crucial criteria that reflect the nutritional status of the rubber tree. Advanced hyperspectral technology can evaluate N and K statuses in leaves rapidly. However, high bias and uncertain results will be generated when using a small size and imbalance dataset to train a spectral estimaion model. A typical solution of laborious long-term nutrient stress and high-intensive data collection deviates from rapid and flexible advantages of hyperspectral tech. Therefore, a less intensive and streamlined method, remining information from hyperspectral image data, was assessed. From this new perspective, a semisupervised learning (SSL) method and resampling techniques were employed for generating pseudo-labeling data and class rebalancing. Subsequently, a 5-classification spectral model of the N and K statuses of rubber leaves was established. The SSL model based on random forest classifiers and mean sampling techniques yielded optimal classification results both on imbalance/balance dataset (weighted average precision 67.8/78.6%, macro averaged precision 61.2/74.4%, and weighted recall 65.7/78.5% for the N status). All data and code could be viewed on the:Github https://github.com/WeehowTang/SSL-rebalancingtest. Ultimately, we proposed an efficient way to rapidly and accurately monitor the N and K levels in rubber leaves, especially in the scenario of small annotation and imbalance categories ratios.

Soybean leaf estimation based on RGB images and machine learning methods

BACKGROUND

RGB photographs are a powerful tool for dynamically estimating crop growth. Leaves are related to crop photosynthesis, transpiration, and nutrient uptake. Traditional blade parameter measurements were labor-intensive and time-consuming. Therefore, based on the phenotypic features extracted from RGB images, it is essential to choose the best model for soybean leaf parameter estimation. This research was carried out to speed up the breeding procedure and provide a novel technique for precisely estimating soybean leaf parameters.

RESULTS

The findings demonstrate that using an Unet neural network, the IOU, PA, and Recall values for soybean image segmentation can achieve 0.98, 0.99, and 0.98, respectively. Overall, the average testing prediction accuracy (ATPA) of the three regression models is Random forest > Cat Boost > Simple nonlinear regression. The Random forest ATPAs for leaf number (LN), leaf fresh weight (LFW), and leaf area index (LAI) reached 73.45%, 74.96%, and 85.09%, respectively, which were 6.93%, 3.98%, and 8.01%, respectively, higher than those of the optimal Cat Boost model and 18.78%, 19.08%, and 10.88%, respectively, higher than those of the optimal SNR model.

CONCLUSION

The results show that the Unet neural network can separate soybeans accurately from an RGB image. The Random forest model has a strong ability for generalization and high accuracy for the estimation of leaf parameters. Combining cutting-edge machine learning methods with digital images improves the estimation of soybean leaf characteristics.

Estimation of Rice Aboveground Biomass by UAV Imagery with Photosynthetic Accumulation Models

The effective and accurate aboveground biomass (AGB) estimation facilitates evaluating crop growth and site-specific crop management. Considering that rice accumulates AGB mainly through green leaf photosynthesis, we proposed the photosynthetic accumulation model (PAM) and its simplified version and compared them for estimating AGB. These methods estimate the AGB of various rice cultivars throughout the growing season by integrating vegetation index (VI) and canopy height based on images acquired by unmanned aerial vehicles (UAV). The results indicated that the correlation of VI and AGB was weak for the whole growing season of rice and the accuracy of the height model was also limited for the whole growing season. In comparison with the NDVI-based rice AGB estimation model in 2019 data (R² = 0.03, RMSE = 603.33 g/m²) and canopy height (R² = 0.79, RMSE = 283.33 g/m²), the PAM calculated by NDVI and canopy height could provide a better estimate of AGB of rice (R² = 0.95, RMSE = 136.81 g/m²). Then, based on the time-series analysis of the accumulative model, a simplified photosynthetic accumulation model (SPAM) was proposed that only needs limited observations to achieve R² above 0.8. The PAM and SPAM models built by using 2 years of samples successfully predicted the third year of samples and also demonstrated the robustness and generalization ability of the models. In conclusion, these methods can be easily and efficiently applied to the UAV estimation of rice AGB over the entire growing season, which has great potential to serve for large-scale field management and also for breeding.

A Generic Model to Estimate Wheat LAI over Growing Season Regardless of the Soil-Type Background

It is valuable to develop a generic model that can accurately estimate the leaf area index (LAI) of wheat from unmanned aerial vehicle-based multispectral data for diverse soil backgrounds without any ground calibration. To achieve this objective, 2 strategies were investigated to improve our existing random forest regression (RFR) model, which was trained with simulations from a radiative transfer model (PROSAIL). The 2 strategies consisted of (a) broadening the reflectance domain of soil background to generate training data and (b) finding an appropriate set of indicators (band reflectance and/or vegetation indices) as inputs of the RFR model. The RFR models were tested in diverse soils representing varying soil types in Australia. Simulation analysis indicated that adopting both strategies resulted in a generic model that can provide accurate estimation for wheat LAI and is resistant to changes in soil background. From validation on 2 years of field trials, this model achieved high prediction accuracy for LAI over the entire crop cycle (LAI up to 7 m² m^-2) (root mean square error (RMSE): 0.23 to 0.89 m² m^-2), including for sparse canopy (LAI less than 0.3 m² m^-2) grown on different soil types (RMSE: 0.02 to 0.25 m² m^-2). The model reliably captured the seasonal pattern of LAI dynamics for different treatments in terms of genotypes, plant densities, and water-nitrogen managements (correlation coefficient: 0.82 to 0.98). With appropriate adaptations, this framework can be adjusted to any type of sensors to estimate various traits for various species (including but not limited to LAI of wheat) in associated disciplines, e.g., crop breeding, precision agriculture, etc.

Retrieving rice (Oryza sativa L.) net photosynthetic rate from UAV multispectral images based on machine learning methods

Photosynthesis is the key physiological activity in the process of crop growth and plays an irreplaceable role in carbon assimilation and yield formation. This study extracted rice (Oryza sativa L.) canopy reflectance based on the UAV multispectral images and analyzed the correlation between 25 vegetation indices (VIs), three textural indices (TIs), and net photosynthetic rate (Pn) at different growth stages. Linear regression (LR), support vector regression (SVR), gradient boosting decision tree (GBDT), random forest (RF), and multilayer perceptron neural network (MLP) models were employed for Pn estimation, and the modeling accuracy was compared under the input condition of VIs, VIs combined with TIs, and fusion of VIs and TIs with plant height (PH) and SPAD. The results showed that VIs and TIs generally had the relatively best correlation with Pn at the jointing-booting stage and the number of VIs with significant correlation (p< 0.05) was the largest. Therefore, the employed models could achieve the highest overall accuracy [coefficient of determination (R ²) of 0.383-0.938]. However, as the growth stage progressed, the correlation gradually weakened and resulted in accuracy decrease (R ² of 0.258-0.928 and 0.125-0.863 at the heading-flowering and ripening stages, respectively). Among the tested models, GBDT and RF models could attain the best performance based on only VIs input (with R ² ranging from 0.863 to 0.938 and from 0.815 to 0.872, respectively). Furthermore, the fusion input of VIs, TIs with PH, and SPAD could more effectively improve the model accuracy (R ² increased by 0.049-0.249, 0.063-0.470, and 0.113-0.471, respectively, for three growth stages) compared with the input combination of VIs and TIs (R ² increased by 0.015-0.090, 0.001-0.139, and 0.023-0.114). Therefore, the GBDT and RF model with fused input could be highly recommended for rice Pn estimation and the methods could also provide reference for Pn monitoring and further yield prediction at field scale.

Integrating a crop growth model and radiative transfer model to improve estimation of crop traits based on deep learning

A major challenge for the estimation of crop traits (biophysical variables) from canopy reflectance is the creation of a high-quality training dataset. To address this problem, this research investigated a conceptual framework by integrating a crop growth model with a radiative transfer model to introduce biological constraints in a synthetic training dataset. In addition to the comparison of two datasets without and with biological constraints, we also investigated the effects of observation geometry, retrieval method, and wavelength range on estimation accuracy of four crop traits (leaf area index, leaf chlorophyll content, leaf dry matter, and leaf water content) of wheat. The theoretical analysis demonstrated potential advantages of adding biological constraints in synthetic training datasets as well as the capability of deep learning. Additionally, the predictive models were validated on real unmanned aerial vehicle-based multispectral images collected from wheat plots contrasting in canopy structure. The predictive model trained over a synthetic dataset with biological constraints enabled the prediction of leaf water content from using wavelengths in the visible to near infrared range based on the correlations between crop traits. Our findings presented the potential of the proposed conceptual framework in simultaneously retrieving multiple crop traits from canopy reflectance for applications in precision agriculture and plant breeding.