Recipient Organization
UNIV OF CONNECTICUT
438 WHITNEY RD EXTENSION UNIT 1133
STORRS,CT 06269
Performing Department
Natural Resources & the Environment
Non Technical Summary
Understory vegetation plays essential roles in forest ecosystems. Understory vegetation exhibits strong temporal dynamics, which have been attributed to different drivers, such as forest management, nitrogen deposition and climate change. However, the spatial distribution and temporal change of understory is still unavailable at state scale because the fieldwork is labor-intensive and sometimes logistically unfeasible. Further, remote sensing of understory is always shaded by the above dense canopy, which are mostly deciduous forests in Connecticut. This proposal uses medium resolution satellite imagery (i.e., Sentinel-2 and Landsat) acquired during the "peek window", which is defined as the short period in early spring, when understory vegetation produce new leaves while the above deciduous forests are still in the leaf-off condition. We will develop time-series-based algorithms to map annual understory and to detect the understory change from 2017 and 2022 based on satellite time series observations from Landsat and Sentinenl-2. The accuracy of these annual maps will be assessed by fieldwork. The outputs of this project will provide the first map to track the spatio-temporal dynamics of understory vegetation statewide over time for Connecticut. This new information will provide supports for sustainable forest ecosystem management and wildlife and biodiversity conservation, thus meeting the need of NIFA Priority Area "Bioenergy, Natural Resources, and the Environment", sub-priority "Forestry".
Animal Health Component
50%
Research Effort Categories
Basic
(N/A)
Applied
50%
Developmental
50%
Goals / Objectives
The project goal to understand the spatio-temporal dynamics of understory vegetation will provide support for sustainable forest ecosystem management and wildlife and biodiversity conservation, thus meeting the need of NIFA Priority Area "Bioenergy, Natural Resources, and the Environment", sub-priority "Forestry". By addressing this NIFA sub-priority, the potential long-term impacts to forests include monitoring understory change in response to forest disturbance and evaluating its impacts on regional biodiversity and potential risks under different future climatic scenarios.
Project Methods
We propose to detect the change of understory vegetation at 10-meter resolution from 2017 to 2023 using Sentinel-2 A/B and Landsats 7-8 (Sentinel-2 C and Landsat 9 will be used if successfully launched in 2021). To create a clear-sky dense time series of satellite observations at 10-meter resolution, we will use Fmask 4.0 approach (Shi Qiu et al., 2019; Zhe Zhu and Woodcock, 2012) to detect clouds, cloud shadows, and snow, and then apply an Extended Super-Resolution Convolutional Neural Network (Shao et al., 2019) to convert 30-meter resolution Landsat pixel into 10-meter resolution pixels, and if combined with 10-meter resolution Sentinel-2 time series, we can create extremely dense time series data (~3 days) at 10-meter resolution. There are mainly three components in this approach (c1-c3).c1. Reference sample collectionWe will collect high-quality reference samples for "understory" and "deciduous forest without understory" for training and accuracy assessment. As it is hard to interpret understory vegetation from remote sensing images, we propose to collect the sample data based on field visits. For training data selection, we will create large polygons of reference data based on limited field visits. For accuracy assessment, we will use a stratified random sampling approach. The stratum will be determined by annual understory and understory change maps (understory vs. non-understory and change vs. no-change). The sample size and the allocation of samples in each stratum will follow the good practice for accuracy assessment (Olofsson et al., 2014).c2. Mapping understory vegetation in deciduous forest in 2017We aim to monitor the changes of understory vegetation in deciduous forest since 2017. The 2017 understory map will be the base map for change detection. The first step is to map deciduous forest at 10-meter, but as far as we know, there is no public available 10-meter land cover/land use map for deciduous forest. Therefore, we will apply the classification component of the widely used Continuous Change Detection and Classification (CCDC) algorithm (Zhu and Woodcock, 2014) to map the deciduous forest in 2017. The basic idea of the classification component is that for each individual pixel, a harmonic modelis estimated based on the time series of clear-sky Sentinel-2/Landsat observations in 2017 (assuming no change within the year), and the land cover/land use class will be labeled based on the model-derived variables, such as coefficients of time series models, Root Mean Square Error (RMSE) from model fit, using the Random Forest (RF) classifier. The RF classifier can be trained based on 2016 National Land Cover Database (NLCD).The second step is to map the understory vegetation from the deciduous forest map. A new random forest classifier will be trained based on the random survey samples regarding to understory and deciduous forest, and the inputs will consist of the variables derived from satellite time series (spectral and temporal dimensions) and metrics from 3D LiDAR data (structural dimension). We propose to extract spectral and temporal features from the dense time series data. A preliminary study, in which all coefficients of time series modelswere used, demonstrated the time series-based approach has high potential to distinguish understory vegetation from deciduous forest. To reduce the redundant information, in this proposal, we will calculate and select new variables that are sensitive to understory vegetation for improved classification and compare with the results by using all the coefficients. We also will calculate metrics (i.e., height in multiple intervals) from LiDAR 3D points, which is capable of representing the understory's and forest's structure (Singh et al., 2015). This method has been recently applied to map understory vegetation for the entire Connecticut land area and achieved an overall accuracy above 90%.c.3. Continuous change detection of understory vegetation in deciduous forestsWe will develop a new algorithm for detecting understory vegetation change based on the 2017 map generated by the above section. We will not compare the understory classification maps at different years to identify the change, because the error from the classification is much larger than the amount of change, and the classification errors will accumulate (Friedl et al., 2010; Fuller et al., 2003; Zhu and Woodcock, 2014). Instead, we will design new variables that are sensitive to understory vegetation from the dense satellite time series during the early spring peek window. For instance, understory vegetation usually becomes greener earlier than deciduous forest without understory. Thus, we can extract the day-of-year (DOY) of start greening at the pixel level (DOY of greener), as well as the maximum slope of the time series model during the DOY peek window (DOY of maximum slope) to separate deciduous forest pixels with and without understory vegetation. By differing these time series variables at two consecutive years, we will be able to identify all kinds of understory vegetation change for each year. It is worth noting that young regenerating trees may behave similarly as understory vegetation, which could increase map commission errors.