3. GLAD Tools for Landsat ARD Applications
The primary purpose of the GLAD ARD is to simplify time-series analysis and to create multi-temporal metrics for land cover and land cover change mapping. To facilitate classification model extrapolation, we have implemented globally consistent reflectance normalization. The ARD product does not represent actual surface reflectance and is not suitable for precise measurement of photosynthesis, water quality, and other variables that require actual surface reflectance.
In this section, we provide tools for application of the ARD time-series for land cover and land cover change mapping and analysis of classification results. To use the ARD tools, you need to install ActivePerl (https://www.activestate.com/products/activeperl/), OSGeo4W (usually installed as QGIS-OSGeo4W package from here https://qgis.org/downloads/) and the GLAD_1.0 package (https://glad.umd.edu/gladtools/Complete_package/GLAD_1.0_master.zip). The installation instructions are provided in the 3.1. GLAD Tools Setup section. Additional GLAD tools may be added from the 3.2. GLAD Tools Depository.
The ARD product is used to create multi-temporal metrics that serve as a consistent input for annual land cover mapping and change detection models. Metrics represent a set of statistics extracted from the normalized surface reflectance time-series within a calendar year. Two independent types of metrics may be created from 16-day time-series data using GLAD Tools: annual phenological metrics and annual change detection metrics.
The annual land cover mapping is based on annual gap-filled phenological metrics which simplify the analysis of surface reflectance and land surface phenology. Phenological metrics may be calculated using a complete or incomplete set of 16-day ARD for at least one year. Several years of data are preferable to implement gap-filling. The GLAD system is designed to provide different sets of phenological metrics for various applications. Please, read the metrics methodology and select the metric types that better suit your needs here: 3.3. Phenological Metrics.
The annual change detection metrics were designed to highlight inter-annual changes of spectral reflectance and are used as source data for change mapping. To generate this metric type, at least two complete years of 16-day data is required. Several types of change detection metrics are designed for different applications. Please, read the metrics methodology and select the metric types that better suit your needs here: 3.4. Change Detection Metrics.
After creating a set of metrics, the data from different tiles must be mosaicked for visualization. Section 3.5. Using Image Mosaics provides tools for stitching tiled data into multi-band image mosaics.
Land Cover and Land Cover Change Classification
The GLAD Tools provide software for the application of a simple but powerful decision tree model for land cover mapping and change detection. The decision tree is a type of supervised classification which requires a set of training data for parameterization of the model. The following principles are currently implemented by GLAD Tools:
a. The classification model is created for two classes only. A “target” class usually represents the class of interest (i.e., forest or forest loss), while the “background” class represents all other land cover types. Training for both classes is required. The model output represents the likelihood of each pixel to belong to a “target” class.
b. The classification model uses training (dependent variable) in the form of polygonal ESRI shapefiles. Shapefiles may be created in QGIS (as shown in sections 3.6 Land Cover Classification and 3.7. Change Detection) or in any other GIS software.
c. Independent variables are represented by multi-temporal metrics. Currently, the GLAD classification tool supports several types of metrics. Please, check the metric types and instructions for software application.
d. To reduce model overfitting and the adverse effect of errors in training data, the decision tree model implements bootstrap aggregation (bagging) and pruning techniques. Parameters for the bagging (number of trees in an ensemble) and pruning (deviance decrease limit) can be adjusted for each classification.
e. The accuracy of the classification model is improved through the iterative process of adding training data, known as “active learning”. The active learning method consists of iterations of model construction, application, evaluation of results, and adding new training data until the desired map quality is achieved.