Final report for LNE23-482R
Project Information
Our project seeks to develop a novel drone-based sensing framework to help fruit growers better manage plant nutrients. Our approach relies on a number of predictive models, based on correlations found between drone imagery derived vegetative indices and individual internal plant macro- and micronutrient levels, to accurately determine the need for supplemental nutrient applications within fruit orchards. This project has three interconnected goals, together which would allow for its adoption by our local agricultural community: 1) determine if corrective action based solely on the drone models affects fruit yield and/or fruit quality, 2) determine if the inclusion of super- or suboptimal data points from less experienced growers extend the effective range of the models, and 3) if the technology is economically feasible, affording cost savings onto farmers. Over the project years, 7 farms have been participated in the research process as actively contributing members of the research team. Growers engaged with on-farm demonstrations and were responsible for applying recommended corrective actions as well as participating in the evaluation of the treatments on fruit yield and quality.
Validate and operationalize predictive models, which utilize drone images/derived vegetation indices (NDVI, NDRE) to quickly predict plant tissue nutrient levels on a whole-farm level instead of relying on traditional, limiting, and costly tissue analysis in hopes of impacting current year’s crop. Farmer’s corrective action (e.g., foliar sprays) will rely solely on model-based predictions, with control plots for comparison. Models have been developed for apple, peach, blueberry, and grape with regards to both macro- and micronutrients, comprised of three years of data across three locations in Connecticut. Once verified with growers, this technology will be available to replace existing nutrient-testing practices.
Current plant nutrient testing protocols rely on two types of tests: soil analysis and tissue analysis. Soil analysis is typically done every two years as soil nutrient levels can be slow to change. Tissue analysis is required because adequate soil nutrition does not always equate to adequate plant nutrition. Tissue analysis will always provide the best insight into the health of a plant. However, tissue analysis relies on skills and services that may not be readily available or accessible to all farmers. The cost of a single plant tissue analysis at the UConn Soil Nutrient Analysis Laboratory is $30.00 and the turn-around time for results is at least several weeks. Typically, farmers do not submit a single tissue sample for analysis and costs can accumulate quickly. This analysis also relies on proper sample collection techniques, 100 of the most recently matured leaves, to give an accurate depiction. Typically, growers receive results at the end of the growing season and may rely on interpretations and corrective action recommendations from Extension professionals. This corrective action is always taken in hopes of impacting the next year’s crop. This current process does nothing to address corrective action for the current year’s fruit crop which may suffer or fail.
Our novel approach utilizes 62 drone-image-driven predictive models developed in our previous study to accurately predict internal plant nutrient levels. The technology utilizes drones to capture multispectral data on a whole farm scale which is then run through numerous crop and nutrient specific equations to determine plant nutrient levels without the need for the traditional testing processes. This eliminates the weeks long turn-around time, providing insight into plant health the same day. Our hope is that this technology will also provide cost savings for farmers and impact their current year’s crop.
Growers have shown a high interest in plant nutrition beginning with a 2015 plant nutrition intensive short course, 5-year study with fruit growers utilizing a combination of tissue and soil analysis, crop load and environmental conditions to develop fine-tuned fertilizer programs, to the recent drone study developing a model to detect nutrient deficiencies early in the season. CT fruit growers were surveyed in June and July 2022 to gauge interest in using drones (# surveys sent out =398, # responses =41). Respondents were from across CT with farm sizes ranging from 1 to over 100 acres. 90.24% expressed interest in learning how drones can be used to detect nutritional issues; 22% presently use drones primarily for farm marketing and advertising; and 22% would like to learn how to use drones in farming. Barriers to drone use identified were ‘have not seen a need yet’, ‘do not understand how they will help my business’, ‘cost’, and ‘not tech savvy enough’. This work will show drones can be used to effectively identify plant nutrition issues. It will be beneficial to growers as it can provide actionable crop analytics in near-real time to take corrective actions. All farms in the northeast could benefit no matter the size.
Research
Hypothesis 1 [H1]: Corrective actions to improve plant nutrition based solely on our proprietary drone-imagery driven predictive models will positively influence the yield and quality of the current year’s fruit crop in participating orchards.
Hypothesis 2 [H2]: Including a broader grower group to include those of varying skill levels and expertise will provide valuable data to extend the effective range of the predictive models.
Hypothesis 3 [H3]: The economic analysis will prove that the technology is affordable and accessible to all fruit growers in CT, despite background, acreage or experience level.
Over the project years, we executed the drone remote sensing and field sampling in candidate farms based on proposed randomized block design. Each year, we tasked drone imaging in June and continued until late at bi-weekly intervals. We established permanent ground control points (GCPs) in each study block in each farm to locate reflectance targets during drone data collection missions. Using RTK GPS receiver, we collected the coordinates of the GCPs and later stages used as inputs for image orthorectification and co-registration processes. Drone images were acquired using Micasense RedEdge-MX dual camera system. We followed pre- and post-sensor calibration steps using spectral reflectance panel in each mission to make sure the radiometric consistency of acquired data. The acquired multispectral images consist of 10 spectral bands in visible and near-infrared wavelengths. We used Agisoft MetaShape software to process drone images. Main steps involved stitching, orthorectification, and radiometric calibration based on GCPs and reflectance panel data. We further analyzed bi-weekly orthomosaics using ArcGIS Pro software. Using orthomosaic we segmented out individual canopies and derived necessary vegetation indices. We are in the process of combining plant tissue analysis data with drone-based vegetation indices refine nutrient predictive models. In addition to regression models, we tested three machine learning models to predict nutrient concentration from remotely sensed spectral and site-level predictors: 1) Partial Least Square Regression, 2) Random Forest, 3) Ridge Regression.
Remote Sesning Analysis:
During project years, we conducted bi-weekly data collection (leaf sampling for nutrient analysis, leaf-level spectral data, and drone imaging) across eight weeks, from mid May (bud burst stage) to late August (harvesting stage for most fruit)
Plant tissue collection & analysis: Each sampling season, a total of 100 leaf samples were collected from apple, peach and blueberry plant, while petiole samples were taken from grapevine. These samples were thoroughly cleaned up, dried in an oven at 60 °C for 24 hours and analyzed at UConn Soil Nutrient Analysis Laboratory. The analysis focused on 14 essential nutrients including nitrogen (N), phosphorus (P), potassium (K), calcium (Ca), magnesium (Mg), iron (Fe), copper (Cu), zinc (Zn), manganese (Mn), molybdenum (Mo), sodium (Na), aluminum (Al), boron (B), and lead
Drone data collection & analysis: A Quadcopter drone, equipped with a Micasense Red edge Dual MX-sensor, was deployed for aerial flight capturing high-resolution multispectral images of orchards. These collected images were processed in Agi soft Meta shape to generate Orthomosaic, and further data extraction was performed in ArcGIS Pro. Various vegetation indices (VIs)- a total of 33 VIs- specifically related to nutrients, were derived from the extracted spectral band values. These indices were correlated with actual lab-analyzed nutrient value. In addition to correlation analysis, simple linear regression, multiple linear regression, and machine learning models (random forest, partial least square regressions) were performed to predict nutrient status. While machine learning models were overfitted by sample size, significant relationships between nutrient levels and spectral indices were observed in simple and multiple linear regressions. As a result, nutrient deficiencies were detected in some of the fruit blocks, and growers were informed to apply targeted foliar nutrient applications of (micro and macro nutrients) as corrective actions.
Spectrometer data collection and Analysis: Leaf-level hyperspectral data was collected using a handheld-spectrometer, which records reflectance data from 300nm to 1200nm-providing a broader spectral range. Total of 30 spectral readings were taken from each block. These spectral readings were taken from the same leaves that underwent laboratory nutrient analysis. From the analysis results of spectrometer, the most sensitive wavelength of each nutrient was determined.
Economic Analysis
The economic analysis was based on representative extension enterprise budgets for major fruit crops, including apples, peaches, blueberries, table grapes, and wine grapes, with all costs standardized to constant 2024 dollars. Three implementation scenarios were evaluated: (1) individual grower ownership of the drone system, (2) cooperative/shared ownership among multiple growers, and (3) commercial service-provider delivery. A cost model was developed to estimate annualized monitoring costs by incorporating equipment, software, labor, maintenance, travel, and operational expenses. Partial-budget and break-even analyses were then used to determine the level of yield and/or fruit quality improvement required for the technology to offset its costs under each implementation model. Finally, a Monte Carlo simulation was performed to evaluate the effects of uncertainty in crop response, market conditions, and operational costs on the probability of achieving positive economic returns. This framework provided a practical assessment of the economic viability and potential adoption pathways for drone-based nutrient monitoring in Northeastern fruit production systems.
Study Findings
Remote Sensing
Wavelength Sensitivity Analysis
Leaf-level hyperspectral data analysis revealed possible fruit-specific sensitive wavelengths for nutrient deficiencies.
Apple: N-528nm, P & K-728nm, Ca-971nm, Mg-415nm, Zn- 936 nm, B-464 nm, Fe-974 nm, Mn-660 nm, Cu-405 nm, Na- 490 nm, and Al- 917 nm.
Peach: N-523nm, P-539nm, K-1001nm, Ca-537nm, Mg-533nm, Al-695nm, B-698nm, Cu-512nm, Fe-669nm, Mn-594nm, Na-508nm and Zn-708nm.
Blueberry: N & P-537nm, K-584nm, Ca-532nm, Mg-521nm, Al-548nm, B-543nm, Cu-777nm, Fe-532nm, Mn-525nm, Na-986nm and Zn- 552nm
Grape: N-583nm, P-902nm, K-503nm, Ca-528nm, Mg-717nm, Al-747nm, B-761nm, Cu-678nm, Fe-695nm, Mn-736nm, Na-725nm, Pb-499nm and Zn-975nm
In year 3, we further evaluated the operational feasibility of predicting macronutrient concentrations (N, P, K, Ca, Mg) using multispectral drone imaging and leaf-level hyperspectral spectroscopy. We tested three machine learning algorithms (Random Forest (RF), Partial Leastsquaer Regression (PLSR), Ridge Regression (RR)). Our results suggested a notable sensor-algorithm dependency: the RF algorithm maximized the predictive accuracy of broadband multispectral data, whereas the PLSR algorithm effectively managed the high-dimensional collinearity of hyperspectral narrow bands. Across both remote sensing platforms, K, Ca, and Mg exhibited the strongest spectral associations (R2≥ 0.50, while highly mobile nutrients (N and P) demonstrated weaker predictability. The feature importance analysis (SHAP and VIP) revealed that phenological progression (cGDD) was the most important predictor for nutrient variability. While drone multispectral imaging offers spatial advantages, leaf-level hyperspectral models were best at determining spectral signatures from complex orchard conditions.
Multispectral Image Analysis
For apple, RF consistently outperformed both PLSR and RR across most macronutrients. The highest predictive accuracy was achieved for Mg (R2 = 0.55, RMSE=0.04 and MAE= 0.03), followed closely by K (R2 = 0.54, RMSE=0.21 and MAE= 0.16). Ca was predicted with moderate accuracy (R2= 0.44), as shown by both RF and PLSR. N showed lower predictability (R2 = 0.34) compared to K, Ca, and Mg. P demonstrated the weakest performance among the macronutrients, with RF explaining 26% of the variance (R² = 0.26), while PLSR and RR had even lower R² values (0.05 and 0.07, respectively). Overall, in apple, RF explained approximately 20-55% of the variability across macronutrients, whereas PLSR and RR showed reduced predictive capability.
In contrast to apple, model performance for peach was less consistent and generally weaker. Although RF again outperformed PLSR and RR, predictive accuracy remained limited for most macronutrients. K was the only nutrient predicted with relatively stronger performance (R² = 0.61, RMSE = 0.28, MAE = 0.21). For N, Mg, Ca and P, RF explained less than 30% of the variance (R² ≤ 0.30). PLSR and RR frequently produced negative R² values for several macronutrients except N by PLSR. Overall result of multispectral data shows that RF was the best predicting model for prediction macronutrients in apple and peach.
Hyperspectral Analysis
In apple, PLSR demonstrated the strongest and most consistent predictive performance across most macronutrients (N, K, Ca and Mg), whereas RF provided the best prediction for P. Ca exhibited the highest overall predictability, with PLSR explaining 59% of the variance (R2= 0.59, RMSE = 0.18, MAE= 0.14). K models also accounted for 50% of the variance (R2= 0.50, RMSE = 0.22, MAE= 0.17). Mg and N showed comparable performance, with PLSR explaining 41% and 40% of the variability respectively. In contrast, P remained the least predictable macronutrient (R2= 0.26) which was by RF. PLSR and RR performed worst for P with very low R2 (0.09 and 0.089 respectively).
In contrast to apple, predictive performance in peach was generally lower and more variable across models. RF consistently outperformed PLSR and RR for all macronutrients but with very low predictive power. Mg showed highest predictability using RF (R2= 0.45, RMSE = 0.06, MAE= 0.05). N, K, Ca, and P all exhibited weaker model performance, with R2 values below 0.40 across models. As observed in apple, P remained the least predicted macronutrients in peach, with limited variance explained regardless of modeling approach. As for fruit types, model performance was nutrient dependent. In apple, PLSR was most effective, whereas in peach RF provided more robust predictions. P consistently showed the weakest estimation in both fruit types.
Economic Analysis
Results showed that the economic viability of drone monitoring is driven primarily by operational scale and the delivery model. Estimated net monitoring costs were approximately $256.10 per acre for grower-owned systems, $79.08 per acre under cooperative ownership, and $53.50 per acre when provided as a commercial service. High-value crops such as apples, peaches, blueberries, and table grapes required only modest improvements in yield or fruit quality to recover monitoring costs under cooperative and service-provider models, whereas lower-value crops, particularly native wine grapes, required substantially larger gains. Monte Carlo simulations further demonstrated that cooperative and service-provider approaches consistently provided a high probability of positive economic returns while reducing investment risk. Overall, the analysis indicates that shared-service and cooperative delivery models are the most economically feasible pathways for expanding access to drone-based nutrient monitoring among small and medium-sized fruit producers in the Northeast, supporting broader adoption of precision nutrient management technologies.
Overall, this project demonstrates the promising potential of combining drone-based remote sensing and machine learning to support nutrient management in perennial fruit production. The predictive modeling component showed that multispectral and hyperspectral imagery can be used to estimate nutrient concentrations in apples, peaches, blueberries, and grapes with good accuracy, particularly for potassium, calcium, and magnesium. Among the algorithms evaluated, Random Forest consistently provided the best predictive performance, suggesting that machine learning approaches are well suited for capturing the complex relationships between canopy spectral responses and plant nutrient status. The study also identified the pre-fruit to mid-season growth period as the most suitable period for nutrient monitoring, when spectral responses were most sensitive to nutrient variation. The economic analysis indicated that the feasibility of adopting drone-based nutrient monitoring depends largely on the implementation model. While individual ownership of drone equipment may not be economically practical for many small and medium-sized fruit growers because of relatively high capital and operational costs, cooperative ownership and commercial service-provider models substantially reduced per-acre monitoring costs and improved the likelihood of achieving positive economic returns for high-value fruit crops. These findings suggest that shared-service approaches may provide the most practical pathway for expanding access to precision nutrient monitoring technologies across the Northeastern United States. For growers, this research suggests that drone-based nutrient monitoring can serve as a valuable decision-support tool that complements, rather than replaces, conventional soil and tissue testing. By providing spatially explicit information across entire orchards and vineyards, drone imagery can help identify nutrient deficiencies earlier, guide targeted tissue sampling, support more timely corrective fertilizer applications, and potentially improve nutrient-use efficiency while reducing unnecessary inputs. These capabilities have the potential to improve fruit quality, maintain productivity, and reduce environmental impacts associated with fertilizer management.
Additional research across more growing seasons, cultivars, and production regions will help further validate and refine the predictive models. Continued advances in remote sensing technologies, machine learning algorithms, and service-delivery models are expected to improve the practicality and accessibility of these tools. Collectively, the technical and economic findings indicate that drone-based nutrient monitoring is a promising approach for supporting precision nutrient management in perennial fruit production and provides a foundation for broader adoption as the technology continues to mature.
Education & outreach activities and participation summary
Educational activities:
Participation summary:
During the project year, there were a total of 3 individual educational activities as part of the outreach efforts of this project. The first was an annual update at the 2025 CT Pomological Society's Summer Twilight meeting. This grower event had around 115 attendees, most of whom were farmers. During this meeting, a final summative update was provided to growers on the project. In addition, the CT Wine Council invited us to speak at their annual meeting about the technology. We were able to share some of the latest data and updates from the project with grape growers. Since grapes were one of the best candidate crops, it was great to be able to speak specifically with grape producers in the state. As always, this technology and its implications were presented in the undergraduate class "Small Fruit Production - PLSC 3440" as part of the novel nutrient management tools section.
Learning Outcomes
- Understanding the potential use of drone-based multispectral imaging for monitoring nutrient status in perennial fruit crops.
- Learning how drone-derived information and machine-learning models can support nutrient-management decisions.
- Gaining hands-on experience with model-based nutrient recommendations through on-farm demonstrations and corrective nutrient applications.
- Increased exposure to evaluating the effects of nutrient-management decisions on fruit yield and quality.
- Awareness of the strengths and limitations of drone-based nutrient monitoring, including differences in model performance among nutrients and crop growth stages.
Project Outcomes
The Master’s-level graduate student supported by this grant successfully completed the research and defended the thesis in Fall 2025.