River Ridge Ranch

Oak Tree Mortality Detection Through Imagery Analysis

Steven Swain, Kevin O'Sullivan, Mitch Alexander, Nicholas Tarpinian

INTRODUCTION:

Over much of the last decade, the Western United States has been subject to extreme drought conditions which has had a significant impact on forested lands. California in particular has been dealing with drought for many decades causing havoc on its agricultural system and resulting in yearly record-breaking fires. As drought conditions persist, oak tree stress and mortality seems to also have become a more prominent issue. This issue leads us to our study site a 722-acre land called River Ridge Ranch located in Tulare County, California with a range of oak woodland species. The ranch has become an ecological preserve near the foothills of the Sierra Nevada Mountain range and Sequoia National Park with the Blue Oaks becoming the most prominent trees on the ranch. This ranch land contains a variety of trees and shrubs with the four main tree species being Blue Oak, Live Oak, Valley Oak, and Buckeye.

OBJECTIVE:

The purpose of this project is to locate and map the dead, distressed, and dying oak tree species at the study site of River Ridge Ranch in Springville, California. Utilizing imagery from unmanned aerial systems (UAS) and aerial photography to assist in the analysis. This imagery will help in the identification of oak trees through the use of classification systems, GIS software’s, deep learning algorithms, and the understanding of spectral indices. Visiting the ranch to conduct fieldwork will help in the validation of dead or distressed trees that will complement the imagery and classification analysis. Furthermore, a suitability analysis will take place to help identify vulnerabilities on the ranch that is causing or leading towards the death of oak trees. We are hoping that this research will provide a good understanding of what is causing the death of oak trees and the impact drought has on the landscape.

Background:

River Ridge Ranch was once a cattle ranch that had been constantly clear-cut for grazing purposes since the 1800s. Within the last six years or so, the new ranch owner Dr. Gary Adest, removed the cattle from the ranch and started working with the California State University of Long Beach (CSULB) Geography department to map out and conduct research on the ranch. This research is being conducted to see how the environment bounces back after being used as grazing land for such a long period of time. The research is also being done to analyze the carbon sequestration capabilities of the land to better inform policy on how old ranch land should be used after a ranch is "retired". Our goal of mapping dead, dying, and distressed trees is key to know where and why trees are dying to help better understand how we prevent oak tree die back and maximize carbon sequestration in ranch oak woodlands located in this region. Exploring past research, we wanted to get a better understanding of how drought affects the landscape. The result of extreme droughts leading to invasive pathogens called Phyophthora ramorum also known as the sudden oak death is leading to the death of oak trees and providing fuel to wildfires across California and the Pacific Northwest. Studies on drought, pathogens, and diseases leading to extensive tree mortality can help us provide a better picture of how this affects the landscape. Also understanding past analysis techniques on how researchers approached land use classification of forests and exploring deep learning algorithms that assist in the imagery classification.

Data and Sources

Our data was pulled from and stored in multiple locations. These locations included personal hard drives, Microsoft One Drive, and on our organizational account on ArcGIS Online.

Methods:

Methodology Flowchart

To more effectively analyze our data, we split up and ran multiple types of analysis to see what method would be the most efficient in locating and mapping dead, dying, and distressed trees within River Ridge Ranch. When we finalize the comparison of our analyses, we will select the best method that we believe could most effectively create an accurate vulnerability map of the study area. These different analysis consist of multiple different types of supervised and unsupervised classifications and multiple different types of vegetation analysis.

01 / 05

Data Aquisition:

We acquired our data in three separate ways. We had data from previous field research conducted by previous cohorts that had been stored on CSULB computers and we had recorded data collected from personal field research. Our field research was conducted by creating a survey utilizing ESRI’s Survey123 application. The survey included GPS and photographic data of dead trees gathered by us utilizing our cellular devices. The GPS data had a +/- 4-meter accuracy. This field data was collected in the late morning and early afternoon hours for two days in late May. We walked through the study area collecting point locations of dead and dying trees to be used as training samples for our supervised classification and the Deep Learning model. Overall, we collected approximately 155 tree samples. Other data such as boundary data and location data were acquired through the University’s ArcGIS Online Organization profile.

UAV Data

Each UAV flight of the study area was flown at 400 feet AGL in overlapping swaths utilizing the eBee fixed wing UAV. The sensors used for this study are the S.O.D.A, the MicaSense RedEdge-MX, the MicaSense RedEdge-MX blue, and the Parrot Sequoia+. This data was then uploaded and rendered by Professor Scott Winslow at CSULB and then given to us for analysis.

NAIP Unsupervised Classification:

We started this process by obtaining NAIP imagery from the years 2005 to 2020 from the USDA NAIP website. Each image was then uploaded into ArcGIS Pro v. 2.9.3 as TIFFs and were left as true-color images so they would be as similar as possible to ensure accuracy. Each NAIP image was then individually ran utilizing the ISO Cluster classification tool and creating seven classifications and leaving the rest of the settings as their default values. We then created a symbology and label file within ArcGIS Pro so that each classified image would have the same symbology and labels in order to ensure an accurate depiction of the data from image to image. The seven classes are Dirt, <25% Green Canopy, 25-75% Green Canopy, 75-90% Green Canopy, 0-100% Green Canopy, Dead/Dry Grass, and No Reflectance.

UAV Supervised Classification:

The supervised classification of the UAV imagery was conducted using Google Earth Engine. Our UAV iamgery had a spatial resolution of 0.12m and the training data was polygons of dead trees from our in field research and that we created in ArcGIS Pro v.2.9.3. The training samples were then split into training and validation groups at a 70%/30% ratio. A Random Forest classifier was used with 140 trees to generate a classification with the same labeling scheme as the training data. This process was iterated, changing the training data, band combinations, sample resolution, and tree size until accuracy was maximized with the resulting details listed above. We then created six classes, Dead Trees, Live Trees, Rocks, Bare Earth, Grass, and structures.

Deep Learning Object Detection:

First we prepared the data for object detection by placing each object (tree) within a polygon slightly larger than the object. We then converted the 5-band images to 3-band RGB images as it is easier to detect the dead trees its natural color state. The each image was then made into an image chip to allow for the proper overlap so that no pixels that made up the dead trees were missed. We then input the image chips into the training modle based on the parameters set by the Silivi-Net team's research we found during our literature review. Once we trained the model, we were then able to use this model to detect objects (dead trees) in all of our UAV imagery.

Suitability Analysis:

A suitability analysis helps in discovering the best-known location based on the inputted criteria and research scenario in question. Based on a multitude of environmental factors on the ranch, we want to know if the death of oak trees on the ranch is being caused by the ongoing drought in California. The criteria used to conduct a suitability map includes a digital terrain model (DTM) of the ranch, slope layer, aspect layer, hillshade layer, NAIP imagery land use classification, and NAIP imagery modified soil adjusted vegetation index (MSAVI).

Every individual criteria required for the suitability analysis.

With this suitability analysis we wanted to determine which areas are more likely to stun growth of canopy or cause death. With each criteria receiving a weight equaling to a 100% to determine the model. The DTM, Slope, Aspect and Hillshade each received 10%; while the NAIP dead tree classification and NAIP MSAVI carried the most weight at 30% each since they are both the main criteria in determining the effects of death on the ranch. The results showed , most areas on the ranch seem to be not ideal spots for growth especially during times of drought. The low moisture content from the MSAVI when comparing both the suitability map and MSAVI determines areas where there would be no growth.

To validate the suitability analysis, an overlay of the dead trees surveyed on the ranch was compared. Each dead tree point collected in the field matches to areas that are not suitable for growth. Overall, inputting different criteria into a suitability model can help determine areas of research that are helpful for the environment or causing harm.

Results:

01 / 06

NAIP Imagery Classification

This map series shows vegetation health of the River Ridge Ranch study area from the years 2005-2020. The classification process used to complete this analysis was the ISO Cluster Unsupervised classification. We learned that NAIP imagery is not the best for species identification and analysis but is more than adequate for vegetation health analysis.

NAIP imagery taken on August 8, 2020 and classified into 7 classes using the ISO Cluster classification method showing vegetation health on River Ridge Ranch.

Google Earth Engine/R UAV Imagery Supervised Classification:

The supervised classification using Google Earth Engine and R produced a map with an 88% producer's accuracy, a 79% consumer's accuracy, and 12.26% OOB accross the entire UAV image.

These tables show the final accuracy results of our supervised classification.

Deep Learning Object Detection:

The deep learning program was successfully able to detect and differentiate dead trees from the rest of the environment in the UAV imagery.

Suitability Analysis:

The suitability analysis, with the vegetation indicies carrying the most weight, show large portions of the ranch become unsuitable to oak trees during times of extended drought.

Discussion

The NAIP Imagery came out with fairly reliable results in determining the overall vegetation health of tree clusters and large individual trees but not the best results for individual trees in the area. This particular NAIP image also shows why you have to go through each image carefully to choose the best quality image as this particular image is extremely noisy due to shadows being casted. This is caused because of the time of day that the image was taken.

This NAIP image from 2005 is not good for determining vegetation health analysis to it being 3-band imagery and does not allow for accurate vegetation health analysis.

Using Google Earth Engine for the supervised classification allows anyone with access to recreate our project with similar results, with the biggest limitation being the access to accurate aerial imagery.

The methodology we utilized for our deep learning analysis in the UAV imagery did not translate as weel to the NAIP Imagery.

Both the deep learning analysis and the suitability analysis require a powerful computer to run effectively.

Project Management:

We were able to stick to our timeline fairly well. The main issue we ran into for meeting our timeline was being able to go out to River Ridge Ranch as a group. We had to change the deadline for that step multiple times due to schedule conflicts.

We would probably due a couple more visits to the ranch to get even more training points and go during different times of they ear to get better data to use with the NAIP imagery.

Ethics:

The most relevant ethical considerations of this study were primarily focused on privacy and providing accurate and appropriate data that benefits society and could hopefully influence policy changes for the better in the future.

Conclusion:

Overall, we were able to accomplish our initial research goal of being able to detect dead oak trees on River Ridge Ranch. We were able to show vegetation health on the ranch on a spatiotemporal scale, identify individual dead trees, and create a suitability analysis that shows how how the droughts have effected oak tree mortality on the ranch.

References:

-       Abdollahnejad, A., and D. Panagiotidis. 2020. Tree Species Classification and Health Status Assessment for a Mixed Broadleaf-Conifer Forest with UAS Multispectral Imaging. Remote Sensing 12 (22):3722.  https://www.mdpi.com/2072-4292/12/22/3722 .  

-       Briechle, S., P. Krzystek, and G. Vosselman. 2021. Silvi-Net – A dual-CNN approach for combined classification of tree species and standing dead trees from remote sensing data. International J ournal of Applied Earth Observation and Geoinformation 98:102292.    https://linkinghub.elsevier.com/retrieve/pii/S0303243420309351 .  

-      Brown, B. J., B. C. McLaughlin, R. V. Blakey, and N. Morueta‐Holme. 2018. Future vulnerability mapping based on response to extreme climate events: Dieback thresholds in an endemic California oak ed. I. Ibáñez. Diversity and Distributions 24 (9):1186–1198.   https://onlinelibrary.wiley.com/doi/10.1111/ddi.12770 .  

-      Byer, S., and Y. Jin. 2017. Detecting Drought-Induced Tree Mortality in Sierra Nevada Forests with Time Series of Satellite Data. Remote Sensing 9 (9):929.  https://www.mdpi.com/2072-4292/9/9/929 .  

-      .Cobb, R. C., S. E. Haas, N. Kruskamp, W. W. Dillon, T. J. Swiecki, D. M. Rizzo, S. J. Frankel, and R. K. Meentemeyer. 2020. The Magnitude of Regional‐Scale Tree Mortality Caused by the Invasive Pathogen Phytophthora ramorum. Earth’s Future 8 (7).  https://onlinelibrary.wiley.com/doi/10.1029/2020EF001500 .  

- Guo, Q., M. Kelly, P. Gong, and D. Liu. 2007. An Object-Based Classification Approach in Mapping Tree Mortality Using High Spatial Resolution Imagery. GIScience & Remote Sensing 44 (1):24–47.  https://www.tandfonline.com/doi/full/10.2747/1548-1603.44.1.24 .  

-Huesca, M., S. L. Ustin, K. D. Shapiro, R. Boynton, and J. H. Thorne. 2021. Detection of drought‐ induced blue oak mortality in the Sierra Nevada Mountains, California. Ecosphere 12 (6).  https://onlinelibrary.wiley.com/doi/10.1002/ecs2.3558 .  

- Johnson, B., R. Tateishi, and N. Hoan. 2012. Satellite Image Pansharpening Using a Hybrid Approach for Object-Based Image Analysis. ISPRS International Journal of Geo-Information 1 (3):228–241.  http://www.mdpi.com/2220-9964/1/3/228 .  

  • Kakooei, M., and Y. Baleghi. 2018. Leaf-Less-Tree feature for semantic labeling applications on Google Earth Engine. In 2018 9th International Symposium on Telecommunications (IST), 609–613. Tehran, Iran: IEEE  https://ieeexplore.ieee.org/document/8661123/ .

-Koppen climate classification | Definition, System, & Map | Britannica.  https://www.britannica.com/science/Koppen-climate-classification .  

-Lisein, J., A. Michez, H. Claessens, and P. Lejeune. 2015. Discrimination of Deciduous Tree Species from Time Series of Unmanned Aerial System Imagery ed. M. Cristani. PLOS ONE 10 (11):e0141006.  https://dx.plos.org/10.1371/journal.pone.0141006 .

-Liu, D., M. Kelly, and P. Gong. 2006. A spatial–temporal approach to monitoring forest disease spread using multi-temporal high spatial resolution imagery. Remote Sensing of Environment 101 (2):167–180.  https://linkinghub.elsevier.com/retrieve/pii/S0034425705004232 .  

-Mayr, W., H. Mayer, U. Bacher, and H. Ebner. 1999. Automatic extraction of trees from aerial imagery. Semantic Modelling for the Acquisition of Topographic Information from Images and Maps :155–166.  

-Meddens, A. J. H., J. A. Hicke, and L. A. Vierling. 2011. Evaluating the potential of multispectral imagery to map multiple stages of tree mortality. Remote Sensing of Environment 115 (7):1632–1642  https://linkinghub.elsevier.com/retrieve/pii/S003442571100068X .  

-Polewski, P., W. Yao, M. Heurich, P. Krzystek, and U. Stilla. 2015. Active learning approach to detecting standing dead trees from ALS point clouds combined with aerial infrared imagery. In 2015 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 10–18. Boston, MA, USA: IEEE  http://ieeexplore.ieee.org/document/7301378/ .  

-Prasai, R., T. W. Schwertner, K. Mainali, H. Mathewson, H. Kafley, S. Thapa, D. Adhikari, P. Medley, and J. Drake. 2021. Application of Google earth engine python API and NAIP imagery for land use and land cover classification: A case study in Florida, USA. Ecological Informatics 66:101474.  https://linkinghub.elsevier.com/retrieve/pii/S157495412100265X .  

-Weinstein, B. G., S. Marconi, M. Aubry‐Kientz, G. Vincent, H. Senyondo, and E. P. White. 2020. DeepForest: A Python package for RGB deep learning tree crown delineation ed. S. Record. Methods in Ecology and Evolution 11 (12):1743–1751.  https://onlinelibrary.wiley.com/doi/10.1111/2041-210X.13472 .  

-What are Convolutional Neural Networks? IBM.   https://www.ibm.com/cloud/learn/convolutional-neural-networks .  

-Wood, J. D., B. O. Knapp, R.-M. Muzika, M. C. Stambaugh, and L. Gu. 2018. The importance of drought–pathogen interactions in driving oak mortality events in the Ozark Border Region. Environmental Research Letters 13 (1):015004.   https://iopscience.iop.org/article/10.1088/1748-9326/aa94fa .  

-Xiao, C., R. Qin, X. Huang, and J. Li. 2018. A STUDY OF USING FULLY CONVOLUTIONAL NETWORK FOR TREETOP DETECTION ON REMOTE SENSING DATA. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences IV–1:163–169.  https://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/IV-1/163/2018/ .  

-Xiao, C., R. Qin, X. Xie, and X. Huang. 2019. Individual Tree Detection and Crown Delineation with 3D Information from Multi-view Satellite Images. Photogrammetric Engineering & Remote Sensing 85 (1):51–59.  https://www.ingentaconnect.com/content/10.14358/PERS.85.1.55 . 

- USDA, National Agricultre Imagery Program

- California State University, Long Beach, Geography Department

Methodology Flowchart

Every individual criteria required for the suitability analysis.