Evaluation and Comparison of Random Forest and A-LSTM Networks for Large-scale Winter Wheat Identification

Tianle He,Chuanjie Xie,Gaohuan Liu,Qingsheng Liu,Shiying Guan

doi:10.3390/rs11141665

Abstract

Machine learning comprises a group of powerful state-of-the-art techniques for land cover classification and cropland identification. In this paper, we proposed and evaluated two models based on random forest (RF) and attention-based long short-term memory (A-LSTM) networks that can learn directly from the raw surface reflectance of remote sensing (RS) images for large-scale winter wheat identification in Huanghuaihai Region (North-Central China). We used a time series of Moderate Resolution Imaging Spectroradiometer (MODIS) images over one growing season and the corresponding winter wheat distribution map for the experiments. Each training sample was derived from the raw surface reflectance of MODIS time-series images. Both models achieved state-of-the-art performance in identifying winter wheat, and the F1 scores of RF and A-LSTM were 0.72 and 0.71, respectively. We also analyzed the impact of the pixel-mixing effect. Training with pure-mixed-pixel samples (the training set consists of pure and mixed cells and thus retains the original distribution of data) was more precise than training with only pure-pixel samples (the entire pixel area belongs to one class). We also analyzed the variable importance along the temporal series, and the data acquired in March or April contributed more than the data acquired at other times. Both models could predict winter wheat coverage in past years or in other regions with similar winter wheat growing seasons. The experiments in this paper showed the effectiveness and significance of our methods.

Highlights

For many years, remote sensing (RS) systems have been widely applied for agricultural monitoring and crop identification [1,2,3,4], and these systems provide many surface reflectance images that can be utilized to derive hidden patterns of vegetation coverage
We proposed two models, random forest (RF) and attention-based long short-term memory (A-long short-term memory (LSTM)), that can be efficiently used for large-scale winter wheat identification throughout the Huanghuaihai Region, by building an automatic data preprocessing pipeline that transforms time-series Moderate Resolution Imaging Spectroradiometer (MODIS) tiles into training samples that can be directly fed into the models
(1) Both the RF and A-LSTM models were efficient for identifying winter wheat areas, with F1 scores of 0.72 and 0.71, respectively

Summary

Introduction

Remote sensing (RS) systems have been widely applied for agricultural monitoring and crop identification [1,2,3,4], and these systems provide many surface reflectance images that can be utilized to derive hidden patterns of vegetation coverage. According to phenology and simple statistics, several key phenology metrics, such as the base level, maximum level, amplitude, start date of the season, end date of the season, and length of the season, extracted from time-series RS images are used as classification features that are sufficient for accurate crop identification. Cornelius Senf et al [5] mapped rubber plantations and natural forests in Xishuangbanna (Southwest China) using multispectral phenological metrics from MODIS time series, which achieved an overall accuracy of 0.735. Yang Shao et al [15] compared the Savitzky-Golay, asymmetric Gaussian, double-logistic, Whittaker, and discrete Fourier transformation smoothing algorithms (noise reduction) and applied them to MODIS NDVI time-series data to provide continuous phenology data for land cover classifications across the Laurentian Great Lakes Basin, proving that the application of a smoothing algorithm significantly reduced image noise compared to the raw data

Methods

Results

Discussion

Conclusion