We carried out our study in a sugarcane field located near the city of Jaboticabal, São Paulo, Brazil ( Figure 1 ). The region has an Oxisol type soil with low slope (0 - 8%). The climate of the region is of type Aw with a summer dry season. Annually, rainfall reaches about 1460 mm and the average temperature is 22.6°C. We conducted our study with the cultivar RB 97-5201 in sixth ratoon. We performed 8 samplings throughout the maturity stage of the crop (beginning February 28 and ending May 8, 2022) with an interval between samplings of 15 days. In each analysis, data were collected at 30 sample points regularly distributed (9 x 9 m grid) and spaced 2 m apart ( Figure 1 ). On evaluation days we captured images with UAV and randomly collected 4 stalks at each sample point. In total, our dataset was composed of 240 samples (30 samples x 8 dates). The images were processed and the stalks were sent to the laboratory for analysis of °Brix and Purity contents. Additionally, we included growing degree days (GDD) information to establish functional relationships with crop phenology.

A multirotor UAV (DJI Phantom 4 Multispectral RTK, Shenzhen, China) was used as the remote sensing platform in this study. The UAV is equipped with a multispectral camera that has five spectral bands, namely Blue (450 nm ± 16 nm), Green (560 nm ± 16 nm), Red (650 nm ± 16 nm), RedEdge (730 nm ± 16 nm), and NIR (840 nm ± 26 nm). The UAV has a sunlight sensor on top to compensate for incident solar radiation during flight and ensure that spectral data are consistent. In addition, it is equipped with a multi-frequency GNSS receiver (DJI D-RTK2 base Station, Shenzhen, China) able to receiver signals from constellation namely GPS, GLONASS, BeiDou, and Galileo, ensuring centimeter positional accuracy, making it possible to acquire temporal data from the same point. The flight missions were performed automatically by application (DJI GS Pro, Shenzhen, China). Flight settings and parameters are described in Table 1 .

The images were stitched using Structure from Motion (SfM) software (Agisoft Metashape Professional 1.5.5, Agisoft, St. Petersburg, Russian) to generate 8 multispectral orthomosaics. To extract the spectral information and calculate the vegetation indices ( Table 2 ) we used the open-source package “FIELDimageR” (Matias et al., 2020); in the programming language R (version 4.1.0).

After the collection of stalks in the field, they were properly identified and taken to the laboratory to determine the quality parameters °Brix and Purity. Initially, the stalks from each sample point were processed individually in a hydraulic press for juice extraction. We used the juice to measure the °Brix content by digital refractometer (ABBE, Atago Pal-1, Tokyo, Japan) and recorded the value corrected to a temperature of 20°C. To measure Purity, we followed the methodology proposed by Consecana (2006). We diluted 10g of clarifying substance based on aluminum chloride in 200 mL of juice. The solution was filtered and the measured value was recorded using a polarimeter (Anton Paar, Bremen, Germany). To determine purity we used Equations 1, 2, and 3.

Where, LPb is the polarimetric reading equivalent to lead subacetate, LAI is the polarimetric reading with aluminum chloride, and %Pol is the apparent sucrose content.

We mapped the dynamic ripening on biometric data ( Figures 2 and 3 ). As the crop ripened, it accumulated SS in the stalk; hence, the °Brix ( Figure 2 ) and Purity ( Figure 3 ) of analytical juice increased temporally and spatially. For instance, °Brix initially was 12.8 ± 2.5. Such a measure of SS then rose to 14 ± 1.7 at the 2^nd evaluation. Additionally, we measured 14.4 ± 2.1°Brix from samples of the 3^rd collection, supporting a field at early maturity and still unsuitable for cost-effective harvesting. As the elongation occurred, however, the °Brix increased significantly. Therefore, its values for the 4^th and 5^th survey-level evaluations were 16.9 ± 2.1 and 17.9 ± 1.6, respectively. Summarily, sugarcane developed the highest °Brix of 19.7 ± 0.9 at the 7^th evaluation.

We identified a similar trend to Purity. The crop produced 80–85% pure juice until the 4^th evaluation. Later, in the 7^th evaluation, however, the measure for this technological feature exceeded 90%, supporting an optimal Pol/SS ratio for high-quality harvesting. More importantly, its distribution throughout the field was homogeneous, further supporting the suitability of such a phenological stage for standard operation and precision crop management. A decreasing proportion of the area with the highest °Brix and Purity at the 8^th evaluation could make the recovery of adequate raw material (sucrose) for industrial processing difficult, driving the need to determine the most reliable time to intervene in the field. Therefore, by analyzing the spatio-temporal variability of such indicators of qualitative yield, we must plan to harvest the sugarcane at the 7^th evaluation. However, we could act earlier since 50–70% of the area produced a raw material with 18–19°Brix and 85–90% purity at the 6^th evaluation.

We selected spectral features to predict °Brix and Purity by applying regression analysis to remote sensing data ( Figure 4 ). A spectral band or VI capable of predicting °Brix could not provide an accurate predictor of Purity and vice versa, supporting structural input-to-output dependencies and particularities of such indicators of qualitative yield. For instance, Blue, Red, and NIR contributed to developing an adequate ten-input predictive model for °Brix. However, they could not function as accurately as Green and PSRI in predicting Purity through a topologically less complex function consisting of seven predictors ( Table 3 ). Such a single band and VI contributed to bringing an R² of 0.85 into the biophysical modeling for Purity, while the adequacy for those above at predicting °Brix was 0.65, making them less accurate.

By analyzing MSEP, however, we could recognize a lower predictive error from such a brix-fitting model, making it more precise. In addition, the more inputs, the higher the accuracy and precision of a polynomial function ( Figure 4 ); however, its complexity can increase, potentially forcing an ML algorithm to misfit data through either underfitting or overfitting a trend. A higher number of inputs usually implies a higher degree of freedom; hence, a model becomes more robust and probable to reject a false hypothesis and produce significant output. However, further increasing the number of predictors could not increase precision ( Supplementary Table 2 and 3 ), supporting the occurrence of multicollinearity or correlation between them ( Supplementary Figure 1 ). Mutual relationships commonly reduce predictive performance in statistical modeling, driving the need to re-design or exclude part of them (Lindner et al., 2022). However, if stakeholders understand the role of independent variables, constraining them in an ML model to reduce multicollinearity is unnecessary. In such a case, it can neither determine exactness and generalization nor result in misinterpretation and misinformation (Lindner et al., 2022). Therefore, balancing adequacy and complexity is significant in addressing the biophysical modeling of °Brix and Purity upon imagery data without computational unfeasibility.

Machine-learning models effectively estimated °Brix ( Figure 5 ) and Purity ( Figure 6 ) by processing biometric and remote sensing data. They were as precise as accurate, allowing the selection of a non-linear function to best describe qualitative yield, logically, irrespective of the indicator. Random forest brought the highest R² into the biophysical modeling for °Brix; hence, it qualified as the most accurate algorithm. In addition, SVM estimated such a measure of SS as accurately and precisely as RF, outperforming both MLR and DT. These approaches developed the least accuracy and precision as ten-input regressors.

We obtained mathematical descriptions with higher precision for Purity since estimates of R², MAE, and RMSE were 0.85–0.9, 4.3–5.1%, and 5.9–6.7%, respectively; the ranges of these metrics for ML models predicting °Brix were 0.6–0.8, 1.1–1.4%, and 1.5–1.9%, respectively. Therefore, compared to Purity, such an indicator of the technological quality of juice added more systematic errors in the modeling, reducing exactness; however, it could not necessarily decrease correctness, which is another part of robustness. Random forest most accurately and precisely predicted not only °Brix but also Purity, further supporting its outperformance at learning on spatio-temporal data to map a series of spectral inputs to an agronomic output. Its predictive metrics were 0.9 R², 4.6% MAE, and 6% RMSE. Additionally, MLR and DT described the Purity at an equal level of adequacy. However, DT outperformed the MLR in supporting a seven-input model to predict Purity.

As sugarcane ripens, it accumulates sugars in organs throughout its structure, such as leaves, stalks, and roots. However, as it grows and develops vegetatively and reproductively, it significantly consumes them to sustain its physiological functions (Khan et al., 2022). In advanced phenological stages, it transports photoassimilates from older (or senescent) leaves to younger parts, such as the stalk in its parenchymal cells and vacuoles (Misra et al., 2022). While these compartments act as sucrose reservoirs in a mature plant, the flowering can manifest as a sink in an over-mature plant, decreasing its content. We could reproduce and visualize these dynamics on maps of °Brix ( Figure 2 ) and Purity ( Figure 3 ). We quantified the highest quantity of SS from samples of the 7^th evaluation; hence, they produced the purest juice. However, °Brix and Purity decreased as the field started flowering at the 8^th evaluation. In addition, stalks became more fibrous, supporting “isoporization” (Morais et al., 2015). Such a phenomenon indicates a reduction in water and sugar, making harvesting inefficient and costly (Poltroniere et al., 2021). By monitoring the field and mapping °Brix and Purity, we can offer stakeholders possibilities to optimize their on-farm management and agribusiness models.

Stakeholders usually rely on sugary substrates to recover sugarcane from the field cost-effectively. °Brix provides a reliable measure of SS in material, while Purity indicates the portion of sugar it contains. Therefore, both are significant technological features for farmers and sugar-energy plants to decide on activities and processes. Raw material with high °Brix and Purity is desirable for commercialization. However, if it excessively consists of minerals and sugars other than sucrose, such as glucose and fructose, its Purity becomes lower than usual, making harvesting and industrial processing challenging. Relevant standardization bodies of the sugar-energy sector in Brazil and abroad set °Brix and Purity to be higher than 18 and 85%, respectively, for economically sustainable mechanical harvesting. Sugarcane can develop such optimal values before or at physiological maturity. However, as it is a semi-perennial grass, it continues to vegetate during ripening; hence, these indicators of qualitative yield decrease nonlinearly over time and spatially, as evidenced by our prescriptive maps ( Figures 2 and 3 ). Prospective producers who search for precision farming support systems can ground their analytical (not empirical) decisions and actions in these digital representations of an experimental field. Perhaps, they harvest material for making food and fuel with higher accuracy and better quality while optimizing workflow.

Spectral features offer stakeholders reliable markers to monitor and map crops. They respond to modifications in nutritional composition (Shendryk et al., 2020), accumulation of biomass (Abebe et al., 2022), and physiological events of maturation (Chea et al., 2020). Single bands and their mathematical combinations into VIs allow for collecting significant imagery data on agroecosystems, whether to make decisions on operations from implementation (e.g., seeding and planting) to harvesting. Researchers often exploit them in remotely assessing the agronomic performance of sugarcane for biomass (Wang et al., 2022), quantitative yield (Sumesh et al., 2021), and standard biometric variables, such as leaf area and height of an individual (Oliveira et al., 2022). However, they still have not emphasized applying ML to UAV imagery data to predict °Brix and Purity as we focus on. Therefore, our AI-intensive approach is innovative. It can realistically monitor saccharification on canopy reflectance during ripening, as photosynthetically active leaves determine stalk sugar concentration (Khan et al., 2022). In addition, it can offer accurate and precise biophysical models to establish functional relationships between spectral features and indicators of technological quality ( Table 3 ).

We obtained evidence for Red and NIR improving the robustness of brix-predicting models. Wavelengths occurring in the electromagnetic radiation spectrum around 680 nm and above visible red light between 780 nm and 1 mm can manifest as exciters to chlorophylls, inducing them to emit either photon (reflectance) in a specific spectral band or fluorescence within a region (Jensen, 2009; Zhao et al., 2010; Stein et al., 2014; Silva Junior et al., 2018; Rodrigues et al., 2020; Barbosa Júnior et al., 2022b). Moreover, they can correlate with the concentration of nutrients (e.g., sugars and minerals) in parts of a plant, such as a stalk (Rodrigues et al., 2022), supporting the ability of our models to predict °Brix and Purity upon imagery data. However, Red and NIR could not estimate Purity as accurately and precisely as Green and RedEdge. As sugarcane grows, its photosynthetic activity intensifies, triggering chemical modifications to chloroplasts. The accumulation of sugars from leaves in the stalk further contributes to physiological reactions in these membrane-bound organelles, altering the balance of chlorophylls and the “greenness” of a plant (Chea et al., 2022). hence, we can acquire significant spectral data from a canopy to predict Purity, which provides a measure of available sucrose in SS.

Predictive data analytics can develop knowledge for advancing agriculture. However, conventional models can be statistically complex and demand considerable computational processing, making their implementation challenging. Even though fundamental approaches, such as correlational or regression analysis (Chea et al., 2020; Todd et al., 2022), can determine functional relationships between spectral and agronomic features, they could not be mathematically sufficient to address problems with a high level of abstraction. Therefore, their application in complex farming systems could not be cost-effective, driving the need to develop an alternative to explain non-linear interactions.

We can train an ML algorithm on a heterogenous and “messy” dataset to learn meaningful and non-duplicative patterns to solve a task automatically, accurately, and unbiasedly. Some applications of ML for sugarcane research and development available from earlier independent studies include predicting or forecasting chlorophyll content (Narmilan et al., 2022), standard morphophysiological variables (Oliveira et al., 2022), production of biomass (Wang et al., 2022), and classify cultivation (Nihar et al., 2022). We developed a new pathway by mapping spectral features to °Brix and Purity; hence we can fulfill a gap in analyzing qualitative yield while improving the addressability of a UAV for scalable aerial remote sensing. Our models are accurate and precise, especially RF and SVM. RF performs an independent prediction by processing data through multiple decision trees (Breiman, 2001). Support vector machine maps inputs to output as a classifier rather than as a regressor (Cristianini and Shawe-Taylor, 2000). As RF provides more parameters and higher overfitting prevention capability for ML, it can outperform SVM in predictive analysis (Yuan et al., 2022), supporting our trends.

Decision tree and MLR could be options for RF and SVM in predicting °Brix and Purity. However, they could develop a lower level of accuracy or precision, driving the need for improvement. The DT consists of an advanced problem-solving and computation-performing procedure. It splits a dataset into multiple branches to establish relationships hierarchically (Ghosh et al., 2022). However, such an algorithm has an inherent flaw, causing it to be less effective. Therefore, implementing a flawless filter could be necessary to increase its accuracy and precision in processing data with significant fluctuations. Even though MLR is topologically and operationally more basic than other ML algorithms, it can develop a highly accurate predictive model for Purity. Such a technique can work well on linear imagery data; hence, it can offer a reliable estimate of quantitative variables, such as productivity, upon VIs (Todd and Johnson, 2021; Krupavathi et al., 2022). However, it could not predict °Brix as accurately as RF and SVM, supporting a non-linear dataset. By introducing GDD into the model, however, we can optimize its predictive performance. Sugarcane’s GDD varies proportionally to its growth and development, acting as a source of constant propagation to MLR.

We demonstrated the technical viability of ML algorithms in predicting °Brix and Purity upon UAV imagery data. Our approach is still at an early stage of research and development. However, it is consistent and can offer stakeholders possibilities to address precision harvesting for cost-effective production. Such an operation is costly (Banchi et al., 2019) and determines the quantity and quality of material for industrial processing (Martins et al., 2021). Therefore, prospective stakeholders across researcher centers and industries who search for decision-making support systems can benefit from our AI-intensive biophysical models to predict the optimal time for harvesting. As sugarcane fulfills approximately 80% of global sugar production (FAOSTAT, 2020), recovering material with the highest quality possible from the field at the precise time and place can be significant to develop a thriving and responsive sugar-energy sector.

Acquiring imagery data by a multispectral sensor onboard UAV allows the development of accurate and precise biophysical modeling of qualitative yield. Our predictive frameworks can be technically comparable with those functions available in independent studies by Bégué et al. (2010), Chea et al. (2020); and Chea et al. (2022). However, they can offer farmers further information to monitor dynamic ripening and map regions of high °Brix and Purity for “smart” harvesting. In addition, our approach can work by processing only remote sensing data, not depending on a conventional ground-level survey to collect biometric measures. Therefore, such an advantage can save farmland staff time and labor, streamline workflow, and ultimately level up the cost-effectiveness of production. Furthermore, while our approach can predict qualitative yield, it can be part of a high-throughput phenotyping program to select early-maturity genotypes. Stakeholders often rely on passive sensors to monitor and assess breeding fields, opening the opportunity to investigate active devices for this purpose.