A Satellite-UAV System for Localized Forest Structural Analysis for Risk Monitoring
Problem: Current forest monitoring systems lack accurate, localized measurements of canopy structure needed for reliable environmental monitoring. Satellite observations provide broad coverage but limited structural detail at local scales, while field surveys and airborne LiDAR are expensive and spatially limited. As a result, many forests lack reliable structural data required to accurately characterize local forest conditions.
Name: Xizhi Wang & Xuanzhe Yao
School: Troy High School
Teacher: Thomas Mosig
Abstract
Forests are increasingly threatened by ecological stresses such as wildfires and drought. Between 2001 and 2024, California lost 2.2 million hectares of trees due to wildfire, highlighting the need for an accurate, localized monitoring system. Structural information, including canopy density and spatial heterogeneity, is a strong predictor of ecological risks and wildfire dynamics. Public satellites for environmental monitoring, such as Sentinel and Landsat, provide extensive multispectral data. However, they operate at spatial resolutions inadequate for detailed analysis of canopy structures.
More expensive systems, such as commercial high-resolution satellites, help bridge this spatial resolution gap, but atmospheric effects such as clouds and illumination variability introduce significant noise and temporal inconsistency. Field measurements and airborne LiDAR also provide accurate structural information, but are spatially limited. These systems, due to high operating costs, often have practical observation intervals of several weeks or months. Consequently, current satellite models often rely on indirect proxies, which struggle to generalize across forests without ground-truth data for calibration.
This project introduces a low-cost, scalable satellite–unmanned aerial vehicle system that uses UAV-collected ground truth to calibrate satellite models for localized monitoring. We used GEDI LiDAR data to train a baseline structural model with Sentinel-1 SAR and Sentinel-2 multispectral features. We derived local canopy and textural metrics based on orthomosaics generated from imagery collected with a custom drone. After calibration, canopy cover prediction improved to R² = 0.58 on held-out testing regions. Vegetation indices and texture metrics achieve R² values of 0.32–0.47, demonstrating improved risk assessment.
Research
Criteria:
Constraints:
Forests must be monitored at large scales to track canopy loss, degradation, wildfire risk, and broader ecosystem change. However, large-scale monitoring is difficult because no single method is simultaneously accurate, affordable, and scalable.
Current approaches have solved various parts of the problem together:
To reduce these tradeoffs, we propose a multi-source monitoring framework that combines satellite imagery for scalable wall-to-wall coverage, spaceborne LiDAR for broad structural supervision, and UAV-derived high-resolution mapping to locally calibrate and adapt models to a specific forest. This creates a pathway toward monitoring that is both scalable and locally accurate, enabling more reliable forest-state and risk assessment across diverse regions.
Engineering Prototype or Model
Figure 1. Pixhawk UAV with IMX477 Camera & Raspberry Pi onboard computing
The initial data satellite prototype consisted of building 30-day data composites with Sentinel-1, Sentinel-2 composites every month over the period of 2019-06-01 to 2022-07-31 using an NCLD land mask across the entire regions of San Bernardino, Los Padres, and Cleveland forests. We then evaluated training with a 80/20 split on available GEDI points. We used spatiotemporal tensors where certain features would have their 10/50/90 percentiles (Dubayah et al.) for a number of indices, resulting in ~150 bands. This original prototype was going in the right direction, however it failed to consider the importance of exposure to lots of different forest types. Also, because we were only taking composites for a singular month the use of spatiotemporal tensors was redundant; we would be trying to show temporal relationships for only about ~5 images per month.
The initial drone prototype was a custom quadcopter built on a Hexsoon EDU-450 frame. It is equipped with a Cube Orange+ flight controller and Here4 GNSS module for autonomous waypoint navigation. A Raspberry Pi for onboard processing and an Arducam IMX477 camera automated image capture during flight missions. The system initially performed basic autonomous mapping flights in testing areas to collect imagery for orthomosaic construction. This prototype focused on validating flight stability, GPS navigation accuracy, and the feasibility of acquiring drone datasets.
Initially, we planned to do a simple fine tuning: linear regression between between the predicted values and the drone orthomosaic.
Figure 2. Initial Data Exports
Materials
Some equipments were graciously borrowed from friends and teachers
Procedures
Initial Testing & Results - Data/Observations
Figure 4. Model performance for initial prototype for 80/20 split and held-out forests. Tested on four forest structure targets: canopy height (rh95), canopy cover, plant area index (PAI), and foliage height diversity (FHD).
Figure 3. UAV-derived orthomosaic spatially aligned with satellite analysis grids for data integration. The highlighted region shows an example grid cell used for localized feature extraction and calibration of satellite-based models.
The initial drone flight system had several limitations, including motion blur from vibration, inconsistent lighting, and uneven overlap due to unoptimized flight paths. These issues reduced orthomosaic quality. Figure 3 shows one of the earliest orthomosaics generated, which appears pixelated and spatially distorted. However, it served as a proof of concept, demonstrating that UAV imagery could be successfully processed into orthomosaics and used as an initial prototype for deriving structural and textural indices.
The initial results showed that the current prototype was problematic. The random 80/20 was alright with R^2 ~ 0.42, however it dropped off rapidly when Los Padres and San Bernardino were the forests being held out.
The Cleveland R^2 made sense because the Cleveland AOI was significantly smaller than the other two. ��The rapid drop-off with holdout made us realize the importance of having more varied data, because currently we cannot say that it generalized.
Prototype or Model Revisions
Figure 5. UAV field deployment and ground control setup during autonomous flight testing and data collection.
Figure 6. UAV in flight over the study area during aerial image acquisition for orthomosaic generation
We changed from 3 regions into 8 different regions after realizing that the original model did not generalize too well with the limited amount of information that we gave it. ��We also changed to a GEDI-first exporting design. We started exporting GEDI shots first, and then building our composite on the exported GEDI shots. This drastically sped up total export times; speeding up from ~12 hours to ~2 hours.��We also changed from a 30-day composite per point into a 60-day composite. The reason for that is that we believed that a longer time frame captured greater stability and because some months had only 1-2 Sentinel images. We also cut down a lot of our bands by removing the spatiotemporal and choosing instead to use mean and standard deviation. A 2 month time frame did not feel long enough to justify having 5 different percentiles per value. We also changed evaluation from a simple 80/20 split on testing with GEDI data to also test with various forest holdouts to test how well it is capable of generalizing.
�We changed our fine-tuning strategy to be to retrain the model with the drone data inside. We assigned the drone data a higher weight (10x) so that the model can maintain its global structure. We did a spatial holdout split test: left side of the site for training, right side for testing to avoid spatial leakage In real world scenarios we would only have part of a forest mapped out and it must be able to generalize.
Figure 4. System Workflow
We modified flight pathing for UAV to ensure consistent altitude and greater image overlap. We added checkpoints between flight pathing so drone can continue mission if it needs to swap battery. We also included additional features for battery optimization including segmented flight missions and optimized waypoint spacing; as we only have a battery life of around 20 minutes. We also added anti-vibration mounts and adjusted the camera lens. Additional preprocessing steps were added before the indices calculation.
Retesting & Results - Data/Observations
Figure 5. Model performance tests on GEDI on new prototype. 80/20 split and held-out forests split. Tested on four forest structure targets: canopy height (rh95), canopy cover, plant area index (PAI), and foliage height diversity (FHD).
See Figure 5. Better results than in the initial prototype. The random 80/20 split achieved all R^2 values greater than 0.6. Both RMSE and MAE dropped or stayed the same with the new design. Performance decreases under forest-level spatial holdout; however, 80/20 split represents reality. one would be able to train the GEDI-based model on the history of a said forest to monitor the forest.
Figure 6. Side-by-Side Comparison of Orthomosaic Processing Pipeline
See Figure 6. The left panel shows the original orthomosaic generated from UAV imagery, which contains visible noise, lighting variability, and fine-scale artifacts due to vibrations and illumination differences. On top of anti-vibration mounts, we applied Gaussian blur to reduce motion noises. We also applied CLAHE enhances local contrast while normalizing brightness across the mosaic, producing a more uniform and analytically stable image. The right panel demonstrates the conversion of the orthomosaic into a 10m × 10m spatial grid, matching the resolution of satellite datas.
Figure 7. Fine-tuning improves site-level performance significantly.
See Figure 6. Our tests reveal that the GEDI canopy cover and the RGB canopy cover had very little correlation, but can be fine tuned to get R^2 ~ 0.5. A few possibilities: 1) GEDI and Sentinel-2’s cover do not measure the same thing. Plausible, since they are inherently different tools. 2) GEDI may not generalize well since only sparse data points are available, making that highly susceptible to noise and variability. However, we were able to get good predictive results with fine tuning using the ground truth.
Conclusion/Results Discussion
This project found that it is possible to map forest structure by combining Sentinel satellite images, GEDI LiDAR data, and local UAV surveys with machine learning. Our results partly supported the idea that using several sensors is better than using just one. Models trained with all the datasets gave more consistent predictions of canopy structure than those using only satellite data. This suggests that each sensor captures different ecological details. Satellite data provided wide coverage, GEDI added height information, and UAV images gave detailed local observations for calibration.
One key finding was that models trained across large regions did not work well in new forest environments unless they were adapted locally. Forest ecosystems differ a lot in species, canopy shape, terrain, and lighting, which can lower prediction accuracy. Recalibrating with local UAV data greatly improved performance at each site, showing that local ground truth is needed for reliable monitoring. This supports the idea that large-scale monitoring needs global data for coverage and local data for accuracy. The improvement after UAV calibration suggests that combining global and local approaches may be needed for practical forest monitoring.
Unexpectedly, canopy cover relationships differed between GEDI structural measurements and RGB-derived UAV metrics, preventing completion of the full UAV feedback loop originally planned. RWe found that canopy cover measured by GEDI and by UAV RGB images did not match as expected, so we could not complete the planned UAV feedback loop. This was not a failure, but showed that each sensor captures different ecological information. GEDI worked well in tall forests but had limited coverage, while RGB images showed detailed surface changes but did not measure height directly. These results point out challenges like spatial mismatch, mixed pixels, limited LiDAR samples, and differences in image mosaics. They also suggest future work on better calibration and combining different signals. Overall, the study shows that using multiple data sources with local recalibration is a practical and scalable way to monitor forest structure.
Our background research showed that there was a gap between large-scale environmental monitoring and localized measurements of forest structure. This understanding guided the design of our prototype, which aims to integrate satellite observations with UAV-derived ground truth. Our project can support wildfire and other risk assessments by improving estimates of canopy structure and vegetation distribution. Additionally, the system may assist forest managers and conservation agencies in monitoring ecosystem health, detecting environmental changes over time, and guiding restoration or land management decisions. Because the method relies on publicly available satellite data combined with low-cost UAV observations, it provides a scalable solution for continuous environmental monitoring across large regions.
Through this project, we learned that inaccuracies in satellite models are often caused by ecological differences between regions rather than insufficient data. Massive datasets for forest monitoring already exist; however, local forests exhibit unique structural relationships that global models cannot fully capture without calibration. Therefore, we are exploring a key trade-off between leveraging the massive scalability of satellites and the precision accuracy of drones. We gained experience combining hardware engineering, remote sensing, and machine learning into a unified system. We learned how to design, calibrate, and operate the drone from the ground up, as well as how to refine models to account for noise sources such as lighting variation, motion blur, and flight stability.
Going forward, we will focus on expanding testing across multiple forest ecosystems to evaluate model generalization under different environmental conditions. We plan to integrate additional UAV sensors, such as multispectral cameras and LiDAR, to further improve structural and vegetation measurements. Future development will include automating the calibration pipeline and optimizing flight planning to enable repeated monitoring over time. Our long-term goal is to develop a real-time monitoring system in which periodic UAV observations continuously refine satellite predictions, enabling adaptive, scalable forest monitoring.
Application/Future Research
References Cited
Dubayah, R., Blair, J. B., Goetz, S., Fatoyinbo, L., Hansen, M., Healey, S., Hofton, M., Hurtt, G., Kellner, J., Luthcke, S., Armston, J., Tang, H., & Silva, C. (2020). The Global Ecosystem Dynamics Investigation: High-resolution laser ranging of the Earth’s forests and topography. Science of Remote Sensing, 1, 100002. https://doi.org/10.1016/j.srs.2020.100002
Haralick, R. M., Shanmugam, K., & Dinstein, I. (1973). Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3(6), 610–621. https://doi.org/10.1109/TSMC.1973.4309314
Mullissa, A. G., Vollrath, A., Odongo-Braun, C., Slagter, B., Balling, J. E., Gou, Y., Gorelick, N., & Reiche, J. (2021). Sentinel-1 SAR backscatter analysis ready data preparation in Google Earth Engine. Remote Sensing, 13(10), 1954. https://doi.org/10.3390/rs13101954
Vollrath, A., Mullissa, A., & Reiche, J. (2020). Angular-based radiometric slope correction for Sentinel-1 on Google Earth Engine. Remote Sensing, 12(11), 1867. https://doi.org/10.3390/rs12111867
Gitelson, A. A., Kaufman, Y. J., Stark, R., & Rundquist, D. (2002). Novel algorithms for remote estimation of vegetation fraction. Remote Sensing of Environment, 80(1), 76–87. https://doi.org/10.1016/S0034-4257(01)00289-9
Woebbecke, D. M., Meyer, G. E., Von Bargen, K., & Mortensen, D. A. (1995). Color indices for weed identification under various soil, residue, and lighting conditions. Transactions of the ASAE, 38(1), 259–269.
Dubayah, R., et al. (2021). GEDI L3 gridded land surface metrics and biomass density, 1 km resolution. NASA LP DAAC. https://lpdaac.usgs.gov/products/gedi03_bv002/
Fassnacht, F. E., Latifi, H., Stereńczak, K., Modzelewska, A., Lefsky, M., Waser, L. T., Straub, C., & Ghosh, A. (2016). Review of studies on tree species classification from remotely sensed data. Remote Sensing of Environment, 186, 64–87. https://doi.org/10.1016/j.rse.2016.08.013
Lausch, A., Erasmi, S., King, D. J., Magdon, P., & Heurich, M. (2017). Understanding forest health with remote sensing—Part I: A review of spectral traits, processes and remote-sensing characteristics. Remote Sensing, 9(10), 1029. https://doi.org/10.3390/rs9101029