Search

Non-invasive crop analysis through image-based methods holds great promise for applications in plant research, yet accurate and robust trait inference from images remains a critical challenge. Our study investigates the potential of AI model ensembling and hybridization approaches to infer sorghum crop traits from RGB images generated via unmanned aerial vehicle (UAV). In our study, we cultivated 21 sorghum cultivars in two independent seasons (2021 and 2022) with a gradient of fertilizer and water inputs. We collected 470 ground-truth N measurements and captured corresponding RGB images with a drone-mounted camera. We computed five RGB vegetation indices, employed several ML models such as MLR, MLP, and various CNN architectures (season 2021), and compared their prediction accuracy for N-inference on the independent test set (season 2022). We assessed strategies that leveraged both deep and handcrafted features, namely hybridized and ensembled AI architectures. Our approach considered two different datasets collected during the two seasons (2021 and 2022), with the training set from the first season only. This allowed for testing of the models’ robustness, particularly their sensitivity to concept drifts, in the independent season (2022), which is fundamental for practical agriculture applications. Our findings underscore the superiority of hybrid and ensembled AI algorithms in these experiments. The MLP + CNN-VGG16 combination achieved the best accuracy (R2 = 0.733, MAE = 0.264 N% on an independent dataset). This study emphasized that carefully crafted AI-based models applied to RGB images can achieve robust trait prediction with accuracies comparable to the similar phenotyping tasks using more complex (multi- and hyper-spectral) sensors presented in the current literature.

Canopy height is an important crop biophysical parameter. It provides information about the crop growth as well as act as an input parameter for biomass and crop yield models. Considering the importance of this parameter, a novel semi-automatic canopy height estimation model has been developed which can work with both georeferenced or non-georeferenced top-of-canopy aerial images. The model employs a Structure-from-Motion algorithm followed by dense point cloud reconstruction and polygon triangulation to obtain polygon meshes which are used for height estimation. The process has been tested on drone-based data collected from a maize crop over the 2018-19 Rabi season from a semi-arid area in central-south India. The ground truth canopy height was measured by manually measuring height of plants using a meter scale. The ground elevation has been modelled using a linear best fit plane and the estimated canopy height was found to have the best R2 value of 0.85 and RMSE values of 14.17 cm.

The biophysical properties of a crop are a good indicator of potential crop stress conditions. However, these visible properties cannot indicate areas exhibiting non-visible stress, e.g., early water or nutrient stress. In this research, maize crop biophysical properties including canopy height and Leaf Area Index (LAI), estimated using drone-based RGB images, were used to identify stressed areas in the farm. First, the APSIM process-based model was used to simulate temporal variation in LAI and canopy height under optimal management conditions, and thus used as a reference for estimating healthy crop parameters. The simulated LAI and canopy height were then compared with the ground-truth information to generate synthetic data for training a linear and a random forest model to identify stressed and healthy areas in the farm using drone-based data products. A Healthiness Index was developed using linear as well as random forest models for indicating the health of the crop, with a maximum correlation coefficient of 0.67 obtained between Healthiness Index during the dough stage of the crop and crop yield. Although these methods are effective in identifying stressed and non-stressed areas, they currently do not offer direct insights into the underlying causes of stress. However, this presents an opportunity for further research and improvement of the approach.

Remote estimation of leaf nitrogen content is a critical requirement for precision farm management. Precise knowledge of nitrogen distribution in the crop enables farmers to decide the fertilisation amount required at specific locations on the farm. Importantly, nitrogen related molecules in plants are transported using water molecules, and water molecules surround the amide bonds (a plant protein created from nitrogen). Consequently, the nitrogen in various crop parts loses its activity in the absence of sufficient water molecules. The association of water molecules around plant proteins makes the optical remote estimation of plant nitrogen challenging as nitrogen and water molecules simultaneously affect the reflectance data. Moreover, the coarse spatial resolution of satellite data and sparse canopy coverage at early growth stages of the crop make it challenging to estimate leaf-level nitrogen contents. Accordingly, this research developed a leaf nitrogen content estimation model using drone-based top-of-canopy 400–1000 nm pure pixel hyperspectral images collected from a maize research farm treated with different water and nitrogen levels. Leaf level spectral signatures were also collected using a field spectroradiometer and used to identify indices more sensitive to nitrogen than water. The leaves were also destructively sampled for obtaining ground truth leaf water and nitrogen content. Red-edge region bands of electromagnetic spectra were identified to be sensitive to leaf nitrogen content. A synthetic data was created using maximum and minimum values of these indices and crop growth stage information, which was further used for training a gradient-boosting machine model to estimate leaf nitrogen content from drone-based hyperspectral images. The estimated leaf nitrogen content values from drone observations were critically analysed with respect to leaf water content values. For water-stressed areas, the model gave an R2 and RMSE of 0.63 and 2.74 mg/g, respectively. However, the model did not perform adequately for well irrigated areas, having an R2 and RMSE of 0.26 and 4.54 mg/g, respectively.

Remotely sensed estimation of leaf water content (LWC) using optical data at early crop growth stage is important for identification of water-stressed plants. However, its accurate estimation is currently a major challenge due to the coarse spatial and spectral resolution of the available optical data, and the atmospheric impact on satellite-based remotely sensed data. Moreover, during early growth stages the canopy coverage is low, increasing the effect of the bare soil background on low spatial resolution data. Consequently, broadband optical data is insensitive to overtone frequencies of O-H stretching bonds of water molecules. Accordingly, this research developed a new model for estimating LWC based on newly identified, pure-pixel, water sensitive indices from high spatial resolution hyperspectral data. A hand-held field spectroradiometer and drone-based hyperspectral imager were used to collect temporal high spectral resolution hyperspectral data (Range: 400–1000 nm; Bandwidth: ~2.1 nm) at leaf level, together with destructively sampled leaves to measure their LWC using the oven-drying method. The spectroradiometer data were used to explore the wavelengths sensitive to vibrational overtone frequencies of O-H bonds of water molecules present in leaves. A total of seven water-sensitive wavelengths were identified, and corresponding normalised indices created for use with pure pixel narrowband hyperspectral data from vegetation. Farm scale maps of LWC were then created using drone-based hyperspectral data, based on minimum and maximum values of the above indices and ‘days after sowing’ information, through a gradient boost machine (GBM) model. The early growth stage maps of LWC were able to distinguish between water-stressed and well-irrigated plots with an R2 of 0.93 and RMSE of 1.6% (g/g).

Leaf Area Index (LAI) is one of the most important biophysical properties of a crop, used in detecting long-term water stress, estimating biomass, and identifying crop growth stage. Remote sensing based LAI estimation techniques perform well for early growth stages but tend to produce high error during the crop reproductive stage due to canopy closure. Moreover, estimation of the true LAI from individual leaf measurements remains a challenge. Consequently, two alternate methods have been developed and compared for estimating the LAI of a maize crop using top-of-canopy RGB images collected throughout the growing season using a hexacopter. Both methods used the RGB images to estimate the canopy height and the green-canopy cover together with a ‘vertical leaf area distribution factor’ (VLADF) from allometric relations (using crop height from RBG images and days after sowing). The first method used an empirical approach to estimate the LAI from training a linear function of the above inputs to Licor canopy analyser values of LAI. The method was trialled for a research farm located in a semi-arid area of southern peninsula India and found to have validation results with an R2 of 0.84 and RMSE of 0.36 for the unused portion of the Rabi (post-monsoon) season data of 2018–19, and R2 of 0.77 and RMSE of 0.45 for the Rabi 2019–20 season data when compared with Licor LAI values. While seemingly acceptable, the Licor canopy analyser gives a foliage area index and so the accuracy of this model was very low (R2 of 0.56 and RMSE of 1.34) when evaluated with true LAI from manual measurements of the leaf area. Consequently, a refinement was introduced using only VLADF, green-canopy cover estimates from the RBG images, and a field measured top leaf angle. The output derived from this conceptual model had an R2 of ~0.6 and RMSE of 0.73 when compared with the true LAI values. Importantly, the LAI from this conceptual model was found to be unaffected by canopy closure during the reproductive stage of the crop.

High-frequency monitoring of agrometeorological parameters is quintessential in the domain of Precision Agriculture (PA), where timeliness of collected observations and the ability to generate ahead-of-time predictions can substantially impact the crop yield. In this context, state-of-the-art internet-of-things (IoT)-based sensing platforms are often employed to generate, pre-process and assimilate real-time data from heterogeneous sensors and streaming data sources. Simultaneously, Time-Series Forecasting Algorithms (TSFAs) are responsible for generating reliable forecasts with a pre-defined forecast horizon and confidence. These TSFAs often rely on modelling the correlation between endogenous variables, the impact of exogenous variables on latent form and structural properties of data such as autocorrelation, periodicity, trend, pattern, and causality to approximate the model parameters. Traditionally, TSFAs such as the Holt–Winters (HW) and Autoregressive family of models (ARIMA) apply a linear and parametric approach towards model approximation, whilst models like Support Vector Regression (SVRs) and Neural Networks (NNs) adhere to a non-linear, non-parametric approach for modelling the historical data. Recently, Deep-Learning-based TSFAs such as Recurrent Neural Networks (RNNs), and Long-Short-Term-Memory (LSTMS) have gained popularity due to their capability to model long sequences of highly non-linear and stochastic data effectively. However, the evolution of TSFAs for predicting agrometeorological parameters pivots around one-step-ahead forecasting, which often overestimates the performance metrics defined for validating forecast capabilities of potential TSFAs. Hence, this paper attempts to evaluate and compare the performance of different machine learning (ML) and deep learning (DL) based TSFAs under one-step and multi-step-ahead forecast scenarios, thereby estimating the generalization capabilities of TSFA models over unseen data. The data used in this study are collected from an Automatic Weather Station (AWS), sampled at an interval of 15 min, and range over one month. Temperature (T) and Humidity (H) observations from the AWS are further converted into univariate, supervised time-series diurnal data profiles. Finally, walk-forward validation is used to evaluate recursive one-step-ahead forecasts until the desired prediction horizon is achieved. The results show that the Seasonal Auto-Regressive Integrated Moving Average (SARIMA) and SVR models outperform their DL-based counterparts in one-step and multi-step ahead settings with a fixed forecast horizon. This work aims to present a baseline comparison between different TSFAs to assist the process of model selection and facilitate rapid ahead-of-time forecasting for end-user applications.