Crop-Type Classification for Long-Term Modeling: An Integrated Remote Sensing and Machine Learning Approach

Henrique G Momm,Racha Elkadiri,Wesley Porter

doi:10.3390/rs12030449

Abstract

Long-term temporal and spatial information of crop type supports a wide range of applications including hydrological and climatological studies. In the U.S., yearly crop data layers (CDLs) are available starting in the early 2000s and have been developed using combined field information and sets of temporal imagery from multiple sensors. Development of long-term crop-type layers similar to CDLs is restricted by reduced accessibility to imagery and the necessary auxiliary datasets. In this study, a procedure to generate a historical crop type was developed and evaluated. Time series of Normalized Difference Vegetation Index (NDVI) datasets from Landsat 5 TM sensor for the Lower Bear Creek watershed were collected and processed. Object-based pseudo phenology curves, represented by the NDVI time series, were generated using noise filtering and dimensionality standardization procedures for the years 1985, 1990, 1995, 2000, and 2005. Classifiers were developed and evaluated using random-forest machine learning algorithms and CDL datasets as the reference. Increased generalization performance was obtained when the model was developed using multi-year datasets. This can be attributed to improved crop type representation during the training phase coupled with characterization of yearly variations due to natural (weather) and anthropogenic factors (farming management). Source of uncertainties were the presence of multiple crops within objects, phenological similarities between soybean and corn/maize, and the accuracy of CDL itself. The proposed procedure supports the development of historic crop types for long-term studies at the field scale in agricultural watersheds.

Highlights

Datasets on crop type classification and its distribution in space and time support a wide range of direct applications and can be incorporated into a variety of environmental and hydrologic models
In the U.S, the cropland data layer (CDL) product generated by the U.S Department of Agriculture (USDA) National Agricultural Statistics Service (NASS) provides a valuable database of crop classification and distribution for the conterminous U.S [11]
Imagery is combined with auxiliary datasets such as digital elevation models (DEMs), the U.S Geological Survey-National Land Cover Dataset (NLCD), and the USDA-Farm Service Agency (FSA)-Common Land Unit (CLU)

Summary

Introduction

Datasets on crop type classification and its distribution in space and time support a wide range of direct applications and can be incorporated into a variety of environmental and hydrologic models. Applications include, but are not limited to, crop yield estimation and prediction [1], natural hazard (e.g., floods and droughts) impact assessment on agricultural commodities [4,5,6], and nonpoint source pollution quantification and mitigation [5,6] For the latter, hydrological watershed models, in particular the annualized agricultural non-point source (AnnAGNPS) [7,8] and the Soil and Water Assessment Tool (SWAT) [9,10], rely heavily on spatiotemporal crop type information to characterize agricultural watershed and best represent existing conditions of land use and farming practices. Imagery is combined with auxiliary datasets such as digital elevation models (DEMs), the U.S Geological Survey-National Land Cover Dataset (NLCD), and the USDA-Farm Service Agency (FSA)-Common Land Unit (CLU)

Methods

Results

Discussion

Conclusion