How Leaf Spectroscopy in Agriculture Optimizes Precision Farming: Five Studies From 2024

• Leaf visual and near-infrared spectroscopy is used to predict new parameters, like leaf macronutrients, micronutrients, and water content, which can have applications in advising precision agriculture decisions.
• Leaf spectral application in phenotyping varieties covers more crop species for crop breeding.
• Disease detection and leaf degradation of cut flowers through leaf spectroscopy are other critical applications.
• Deep-learning chemometric models are the best for analyzing large volumes and hundreds of features generated in spectral data for accurate prediction.

Near-infrared and visual light are exceptionally suited for studying plants, as plants absorb, transmit, and reflect these wavelengths. Each plant compound interacts with different wavelengths, and the light interaction will also vary to provide a unique spectral signature of the compounds and the entire plant. Spectroscopy is the technology that measures these light interactions to estimate plant growth, health, and stress. Leaf spectroscopic applications are vital in precision agriculture, and research in this area is increasing. In this article, we summarize five crucial research findings from 2024.

1. Genotyping Wheat Plants

Scientists are looking for variations in the gene pool to climate-proof wheat species, Triticum aestivum (bread wheat) and T. durum (durum wheat). However, wheat has developed a genetic bottleneck due to domestication over centuries, so existing cultivars are very similar. In addition, efficient phenotyping and genotyping techniques for related traits are few, destructive, resource-intensive, time-consuming, and expensive. Visual and near-infrared (Vis-NIR) spectroscopy is emerging as a fast and reliable technique since genotype influences crops’ biochemical and structural aspects.

Scientists Salehi, Miree, Jafari, Hemmat, and Majidi, also wanted to consider the changes in leaf optical properties at different crop phases.

The experiment

The scientists decided to use Vis-NIR, as spectroscopy can analyze hundreds of features if chemometric models are used. They chose to test the efficiency of PCA-based “soft independent modeling of class analogy (SIMCA)” vs. deep learning model artificial neural networks (ANN). They collected field leaf reflectance spectra for five wheat genotypes- durum, monococcum, roshan,  synthetic, and turanicum at three crop stages of leaf opening, tillering, and flowering. A deep learning technique, the stacked autoencoder (SAE), was used to extract features from the spectra.

The results showed that though the spectra were very similar, there were variations among the genotypes and the growth stages.

The linear SIMCA models predicted wheat genotypes accurately at the tillering stage. However, the ANN method using the features from SAE performed better for spectra taken at leaf opening and tillering. The ANN models predicted the wheat genotypes with 100% accuracy and during flowering with 98.02% accuracy.

Takeaway: Deep learning chemometric models, especially those that cover feature extraction from Vis-NIR data, are well-suited to identify wheat genotypes at various stages.

2. Estimating Fruit Tree Leaf Water Content

Figure 1: “Changes in spectral at different LWC in walnut leaves,” Cui et al. 2024. (Image credits: https://doi.org/10.3390/agronomy14081664).

Leaf water content (LWC) is vital for maintaining physiological processes. However, as leaves are the largest organ, they are most susceptible to environmental conditions. Hence, the leaf water content is considered an indicator of crop water status and is monitored to decide the level of irrigation needed. Traditional water estimation methods are not suitable for covering large areas. Hyperspectral data has been used to study fruit trees. So far, most applications of spectral data have been for determining nitrogen levels and chlorophyll content in fruit tree leaves.

Chinese scientists Cui, Sawut, Ailijiang, Manlike, and Hu decided to use only a few wavelength bands from hyperspectral data to calculate leaf water status with vegetation indices.

Experiment

The scientists collected leaf spectroscopy data in the midday for three fruits, apricot, walnut, and jujube trees, using hyperspectral data consisting of visual, NIR, and short wave infrared (SWIR) light to estimate LWC. They sampled the bands 350–1050 nm at 1.4 nm intervals and 1001–2500 nm at 1.1 nm intervals. They preprocessed the spectral data to get a better signal-to-noise ratio, using Standard normal variate (SNV), continuous wavelet transform (CWT), and fractional order derivative (FOD) methods.

The Pearson correlation coefficient method screened the sensitive bands, reducing data volume and improving model accuracy. The results showed that preprocessing with CWT and FOD improved the correlation of spectral data and LWC. FOD gave better results than CWT methods.

The processed spectral data was used to develop deep learning convolutional neural network (CNN) models to predict LWC in the three fruit species individually and in a mixed sample set.

The CNN CWT3 and FOD1.2 models were validated for apricot, jujube, and mixed sample sets. The FOD1.2-CNN model was stable and had the highest accuracy of R2 > 0.95 in estimating LWC for the fruit trees.

For higher reflectance in leaves with low water content, see Figure 1. The water-sensitive light bands in this study were 600-700 nm associated with chlorophyll absorption. The scientists explained this as the changes that occur due to alterations in chlorophyll light absorption due to changes in LWC.

Takeaway: The CNN deep learning models were more accurate in predicting LWC than traditional machine learning models like partial least squares regression (PLSR), random forest regression (RFR), and support vector regression (SVR).

3. Early Diseases Diagnosis In Poplar

Diseases are a major cause of crop yield reduction, affecting growth and development. An early diagnosis can help growers isolate and treat affected plants to prevent disease spread and contain yield loss in poplar (Populus L.), used to produce pharmaceuticals, biomass energy, lubricants, and polymers. The species is also used for carbon fixation as it grows fast and has a short rotation period. However, the tree is plagued by anthracnose caused by Colletotrichum gloeosporioides, which can affect its growth rate and the formation of essential compounds used to make medicines. Moreover, its symptoms are similar to black spot disease, complicating its diagnosis and leading to economic losses in China.

Experiment

Forest scientists Jia, Duan, Wang, Wu, and Jiang wanted to find a means of early detection for improved prevention and control of diseases using a model based on hyperspectral data. Their aim was a model that could detect anthracnose using spectral disease indices (SDIs).

Figure 2: “Four types of poplar leaves: (a) healthy; (b) black spot disease; (c) early-stage anthracnose; (d) late-stage anthracnose. The numbers on the lower right side show the scale of the infected area on the leaf “ Jia et al. 2024. (Image credits: https://www.mdpi.com/1999-4907/15/8/1309)

They collected spectral data within the 400-2400 nm range from four sets of poplar-healthy trees affected by early-stage anthracnose, late-stage anthracnose, and black spot disease (see Figure 2). The spectra were collected in three bands- visual, NIR, and all to produce the models.

They first preprocessed the data to construct a model based on deep learning (LSTM and 1DCNN) and machine learning classification (RF and SVM). For this, the scientists used the “Design of Experiments (DoE) method” to find the best combination of preprocessing techniques.

The data was analyzed using typical feature extraction methods like principal component analysis (PCA), successive projection algorithm (SPA), and variable combination population analysis (VCPA), and the results were compared with those from SDI models. The SDI model had an accuracy of 89.86% but was inferior to the performance of typical feature extraction methods. However, the SDI models had 100% success in detecting black spots and early-stage anthracnose.

Takeaway: SDI-based models can be used for low-cost, rapid, non-destructive diagnostic tools for early poplar anthracnose detection. The model can be tested for large-scale aerial data to help manage poplar forestry.

4. Estimating Macro And Micro-Nutrients In Oil Palm

Figure 3.:” Relationship of Ash, N, P, K, Mg, B, Cu and Zn elements between FOSS NIRS DS2500 (predicted) and conventional chemical laboratory (measured/reference) values for leaf sample,” Thandapani et al. 2024. (Image credits: DOI 10.1088/1755-1315/1308/1/012036)

The oil palm industry is focused on improving efficiency by reducing cost per unit output. Fertilizers are a significant cost factor, and to optimize their use, the industry spends considerable sums on chemicals to analyze leaf, soil, and fertilizer samples in laboratories. NIR spectroscopy data analyzed by models has provided indirect yet accurate estimations without using expensive chemicals in other cash crops like sugarcane, oranges, etc.

Experiment

Thandapani, Harahap, and Nadaraj wanted to extend this method to oil palms. They tested the capacity of NIR spectroscopy-based models to determine fertilizers, leaf, and soil nutrient content. They collected data in the 800-2500 nm range and analyzed it with the FOSS NIRS DS2500 model in laboratories. The device gave results in under a minute with a single scan. Since a wide bandwidth was used, it covered the interaction of many elements and compounds with light.

Using this method, the scientists could accurately analyze the nitrogen, phosphorus, potassium, magnesium, calcium, boron, zinc, and copper levels in the leaf, soil, and fertilizer samples. However, the FOSS NIRS DS2500 model has to be calibrated periodically with data from a traditional laboratory; see Figure 3 for the calibration relationship for leaf data.

Unification of statistics, plant biochemistry, sampling, and analysis could eliminate nearly 90% of chemicals without affecting result accuracy.

Takeaway: NIR spectroscopy analyzes multiple nutrient statuses with a single scan while helping reduce time and costs for the oil palm industry.

5. Leaf Degradation Analysis in Floriculture

Floriculture is a significant revenue source for many countries. The postharvest vase-life that determines the ornamental value is short-lived. Several methods are used to extend vase life, including preservative solutions.

Experiment

Barbosa et al. measured the efficacy of some solutions in preventing leaf degradation by tracking leaf pigments using hyperspectral data for lisianthus (Eustoma grandiflorum) cut flowers. The scientists expected to gain insights into postharvest physiology of pigment levels,  dehydration, and internal structure changes in the cut flower stems.

The preservative solutions used to treat the cut flowers were phytohormones, sucrose, glucose, and deionized water. The scientists collected leaves from cut flower stems every 4 days for 12 days for hyperspectral analysis. The hyperspectral data in the range of 350-2500 nm collected by the sensor was analyzed by statistical tests. Pigment prediction in leaves was done by estimating the ratio between carotenoids to chlorophyll (CAR/CLF) using PLSR and RRMSE.

The best preservation of leaf quality was obtained from glucose treatments, especially at dosages of 180 g/L, followed by similar performance by sucrose and phytohormone treatments. Glucose replaced the loss of sugar content in flowers, which led to their senescence.

Discriminant analysis results showed the difference in spectral due to the dosages tested, proving that hyperspectral data is suitable for measuring leaf pigment and water stress in cut flowers. The prediction by CAR/CLF had R² values of 0.6, and the RRMSE was below 6.99%.

Takeaway: Hyperspectral analyses are a potential means of leaf degradation evaluation in lisianthus flower stems undergoing different treatments.

Measuring Leaf Spectroscopy

The experiments described have used portable and laboratory spectrometers. CID Bio-Science’s  CI-710s SpectraVue Leaf Spectrometer can be used in most experiments described. It is a portable tool to estimate leaf spectroscopy, pigment changes, and stress in a non-destructive method. The results are obtained in real-time and are accurate, helping scientists by cutting analysis time, costs, and efforts and being reliable.

Sources

Barbosa, T. K. M., Fiorio, P. R., Calaboni, C., Kluge, R. A., Demattê, J. A. M., Mattiuz, C. F. M., … & Ré, N. C. (2024). Lisianthus (Eustoma grandiflorum) leaf degradation analysis in the postharvest by VIS-NIR-SWIR reflectance spectroscopy. Ciência Rural, 54(9), e20230143.

Cui, J., Sawut, M., Ailijiang, N., Manlike, A., & Hu, X. (2024). Estimation of Leaf Water Content of a Fruit Tree by In Situ Vis-NIR Spectroscopy Using Multiple Machine Learning Methods in Southern Xinjiang, China. Agronomy, 14(8), 1664.

Jia, Z., Duan, Q., Wang, Y., Wu, K., & Jiang, H. (2024). Detection Model and Spectral Disease Indices for Poplar (Populus L.) Anthracnose Based on Hyperspectral Reflectance. Forests, 15(8), 1309.

Salehi, B., Mireei, S. A., Jafari, M., Hemmat, A., & Majidi, M. M. (2024). Integrating in-field Vis-NIR leaf spectroscopy and deep learning feature extraction for growth-stage dependent and independent genotyping of wheat plants. Biosystems Engineering, 238, 188-199. https://doi.org/10.1016/j.biosystemseng.2024.01.016

Thandapani, P. U., Harahap, Z., & Nadaraj, S. (2024, February). Usage of near infrared spectrometer as an analyzing tool for nutrients in leaf, fertilizer and soil in oil palm industry. In IOP Conference Series: Earth and Environmental Science 1308 (1). IOP Publishing.