Abstract

Due to a rapid increase in accessible Earth observation data coupled with high computing and storage capabilities, multiple efforts over the past few years have aimed to map land use/land cover using image time series with promising outcomes. Here, we evaluate the comparative performance of alternative land cover classifications generated by using only (1) phenological metrics derived from either of two land surface phenology models, or (2) a suite of spectral band percentiles and normalized ratios (spectral variables), or (3) a combination of phenological metrics and spectral variables. First, several annual time series of remotely sensed data were assembled: Accumulated growing degree-days (AGDD) from the MODerate resolution Imaging Spectroradiometer (MODIS) 8-day land surface temperature products, 2-band Enhanced Vegetation Index (EVI2), and the spectral variables from the Harmonized Landsat Sentinel-2, as well as from the U.S. Landsat Analysis Ready Data surface reflectance products. Then, at each pixel, EVI2 time series were fitted using two different land surface phenology models: The Convex Quadratic model (CxQ), in which EVI2 = f(AGDD) and the Hybrid Piecewise Logistic Model (HPLM), in which EVI2 = f(day of year). Phenometrics and spectral variables were submitted separately and together to Random Forest Classifiers (RFC) to depict land use/land cover in Roberts County, South Dakota. HPLM RFC models showed slightly better accuracy than CxQ RFC models (about 1% relative higher in overall accuracy). Compared to phenometrically-based RFC models, spectrally-based RFC models yielded more accurate land cover maps, especially for non-crop cover types. However, the RFC models built from spectral variables could not accurately classify the wheat class, which contained mostly spring wheat with some fields in durum or winter varieties. The most accurate RFC models were obtained when using both phenometrics and spectral variables as inputs. The combined-variable RFC models overcame weaknesses of both phenometrically-based classification (low accuracy for non-vegetated covers) and spectrally-based classification (low accuracy for wheat). The analysis of important variables indicated that land cover classification for this study area was strongly driven by variables related to the initial green-up phase of seasonal growth and maximum fitted EVI2. For a deeper evaluation of RFC performance, RFC classifications were also executed with several alternative sampling scenarios, including different spatiotemporal filters to improve accuracy of sample pools and different sample sizes. Results indicated that a sample pool with less filtering yielded the most accurate predicted land cover map and a stratified random sample dataset covering approximately 0.25% or more of the study area were required to achieve an accurate land cover map. In case of data scarcity, a smaller dataset might be acceptable, but should not smaller than 0.05% of the study area.

Highlights

  • Knowledge about land use/land cover (LULC) is fundamental for natural resource management, agricultural policy making, and regional and urban planning

  • This study focused on evaluating the performance of alternative land cover classifications using either (1) only phenological metrics derived from either of different land surface phenology (LSP) models: The Convex Quadratic Model, in which EVI2 = f (AGDD) [11,21] and the Hybrid Piecewise Logistic Model, in which EVI2 = f [22], or (2) a suite of spectral band percentiles and normalized ratios, or (3) both phenological metrics and spectral variables

  • There was no obvious choice between the 2016 Random Forest Classifiers (RFC) models using Convex Quadratic model (CxQ) versus Hybrid Piecewise Logistic Model (HPLM); the HPLM RFC models performed better on Analysis Ready Data (ARD) data and the CxQ RFC models better on Harmonized Landsat and Sentinel-2 (HLS) data

Read more

Summary

Introduction

Knowledge about land use/land cover (LULC) is fundamental for natural resource management, agricultural policy making, and regional and urban planning. Most reliable data sources for LULC information are periodic surveys from governmental agencies, e.g., the National Resource Inventory and the National Agricultural Statistics Service (NASS), both in the United States Department of Agriculture (USDA) [1,2] Those datasets often lack spatial and temporal details, which prevents a comprehensive analysis of land change. Remote sensing technology can complement field observations and surveys Conventional classification approaches, such as those applied in the National Land Cover Dataset (NLCD) [3,4,5] or the Cropland Data Layer (CDL) [6], were developed in an era of data scarcity and limited computational power and data storage. Compared to the traditional approach, the use of image time series often improves classification accuracy by incorporating both spectral and temporal profiles [8,9,10]

Methods
Results
Discussion
Conclusion
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call