A Comparison of Estimating Crop Residue Cover from Sentinel-2 Data Using Empirical Regressions and Machine Learning Methods

Ding, Yanling; Zhang, Hongyan; Wang, Zhongqiang; Xie, Qiaoyun; Wang, Yeqiao; Liu, Lin; Hall, Christopher

doi:10.3390/rs12091470

Cited by 28 publications

(16 citation statements)

References 60 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In the past decade, the focus was mainly on the quantification of crop residue or NPV coverage (in %) from remote sensing data (e.g., [ 33 ]). Hereby, applied methods ranged from empirical algorithms (e.g., crop residue indices), over classification [ 34 ], to spectral unmixing [ 11 ], spectral angle methods [ 35 ] and spectral mixture analysis (SMA) [ 36 , 37 ].…”

Section: Introductionmentioning

confidence: 99%

Assessing Non-Photosynthetic Cropland Biomass from Spaceborne Hyperspectral Imagery

et al. 2021

View full text Add to dashboard Cite

Non-photosynthetic vegetation (NPV) biomass has been identified as a priority variable for upcoming spaceborne imaging spectroscopy missions, calling for a quantitative estimation of lignocellulosic plant material as opposed to the sole indication of surface coverage. Therefore, we propose a hybrid model for the retrieval of non-photosynthetic cropland biomass. The workflow included coupling the leaf optical model PROSPECT-PRO with the canopy reflectance model 4SAIL, which allowed us to simulate NPV biomass from carbon-based constituents (CBC) and leaf area index (LAI). PROSAIL-PRO provided a training database for a Gaussian process regression (GPR) algorithm, simulating a wide range of non-photosynthetic vegetation states. Active learning was employed to reduce and optimize the training data set. In addition, we applied spectral dimensionality reduction to condense essential information of non-photosynthetic signals. The resulting NPV-GPR model was successfully validated against soybean field data with normalized root mean square error (nRMSE) of 13.4% and a coefficient of determination (R2) of 0.85. To demonstrate mapping capability, the NPV-GPR model was tested on a PRISMA hyperspectral image acquired over agricultural areas in the North of Munich, Germany. Reliable estimates were mainly achieved over senescent vegetation areas as suggested by model uncertainties. The proposed workflow is the first step towards the quantification of non-photosynthetic cropland biomass as a next-generation product from near-term operational missions, such as CHIME.

show abstract

Section: Introductionmentioning

confidence: 99%

Assessing Non-Photosynthetic Cropland Biomass from Spaceborne Hyperspectral Imagery

et al. 2021

View full text Add to dashboard Cite

show abstract

“…In fact, Li et al (2014) showed that the larger the dataset, the better the performance of the random forest model in their study on improving the linkage affinity prediction of scoring functions on the substitution of random forest for linear regression. This conclusion was also reached by Ding et al (2020) in their study on the comparison of empirical regressions and machine learning methods for crop residue cover estimation using Sentinel-2 data, as well as by Wang et al (2016) in their study on wheat biomass estimation using the random forest regression algorithm and remote sensing data. They all reported that the accuracy of the machine learning approaches was improved when increasing the training sample size relative to the total sample population.…”

Section: Selection Of Modeling Approachmentioning

confidence: 78%

Dry season forage assessment across senegalese rangelands using earth observation data

Diouf

Diédhiou

et al. 2022

Front. Environ. Sci.

View full text Add to dashboard Cite

Strengthening of feed security in the Sahel is urgently needed given the climate change and growing human population. A prerequisite to this is sustainable use of rangeland forage resources for livestock. Many studies have focused on the assessment of rangeland resources during the rainy season, while only a few have focused on the dry season which is the longest and most demanding period for livestock in Sahelian rangelands. The objective of this study is to develop remote sensing-based models for estimating dry season forage vegetation mass. To that end, 29 vegetation indices calculated from each of the MODIS-MCD43A4 (500 m), Landsat-8 (30 m), and Sentinel-2 (10 m) satellite products were used and tested against in situ data collected during three field-measurement campaigns in 2021 at eleven monitoring sites across Senegalese rangelands. Four statistical models were tested, namely, random forest, gradient boosting machines, and simple linear and multiple linear regressions. The two main vegetation mass variables modeled from remote sensing imagery were the standing herbaceous and litter dry mass (BH) and total forage dry mass (BT) with a dry mass of woody plant leaves added to BH. Overall, Sentinel-2 data provided the best performance for the assessment of BH with multiple linear regression (R2 = 0.74; RMSE = 378 kg DM/ha) using NDI5 (Normalized Difference Index5), GRCI (Green Residue Cover Index), SRI (Simple Ratio Index), TCARI (Transformed Chlorophyll Absorption in Reflectance Index), and DFI (Dead Fuel Index) indices. For BT, the best model was also obtained from Sentinel-2 data, including RVI3 (Ratio Vegetation Index3) (R2 = 0.78; RMSE = 496 kg DM/ha). Results showed the suitability of combining the red, green, blue, NIR, SWIR1, and SWIR2 bands in monitoring forage availability during the dry season. Our study revealed that the spectral richness of the optical sensor systems Sentinel-2, Landsat-8, and MODIS-MCD43A4 allowed for accurate assessments of dry-season forage mass of semi-arid rangelands. Adding to this, the high spatial and temporal resolution of Sentinel-2 satellite imagery makes this a promising data source for timely monitoring. These findings can support the monitoring of the animal feed balance in Sahelian countries and contribute to enhancing the resilience of pastoralism toward feed shortage through early warning systems.

show abstract

“…Recently, Beeson et al, (2020) used Landsat-derived NDTI to map conservation tillage in three Midwest states over a 10-year period, with overall accuracies ranging from 64% to 78%, which is a typical accuracy range for studies employing this index [4]. In addition to use of spectral indices for NPV cover estimation, researchers using broadband multispectral imagery have employed various innovative methods such as spectral unmixing (Laamrani et al, 2020; R 2 = 0.70 [29]) and machine learning (Ding et al, 2020; R 2 = 0.69 [30]), but seldom do such studies produce accuracies exceeding 80%.…”

Section: Crop Residue Measurement Using Broadband Multispectral Indicesmentioning

confidence: 99%

Evaluation of SWIR Crop Residue Bands for the Landsat Next Mission

et al. 2021

View full text Add to dashboard Cite

This research reports the findings of a Landsat Next expert review panel that evaluated the use of narrow shortwave infrared (SWIR) reflectance bands to measure ligno-cellulose absorption features centered near 2100 and 2300 nm, with the objective of measuring and mapping non-photosynthetic vegetation (NPV), crop residue cover, and the adoption of conservation tillage practices within agricultural landscapes. Results could also apply to detection of NPV in pasture, grazing lands, and non-agricultural settings. Currently, there are no satellite data sources that provide narrowband or hyperspectral SWIR imagery at sufficient volume to map NPV at a regional scale. The Landsat Next mission, currently under design and expected to launch in the late 2020’s, provides the opportunity for achieving increased SWIR sampling and spectral resolution with the adoption of new sensor technology. This study employed hyperspectral data collected from 916 agricultural field locations with varying fractional NPV, fractional green vegetation, and surface moisture contents. These spectra were processed to generate narrow bands with centers at 2040, 2100, 2210, 2260, and 2230 nm, at various bandwidths, that were subsequently used to derive 13 NPV spectral indices from each spectrum. For crop residues with minimal green vegetation cover, two-band indices derived from 2210 and 2260 nm bands were top performers for measuring NPV (R2 = 0.81, RMSE = 0.13) using bandwidths of 30 to 50 nm, and the addition of a third band at 2100 nm increased resistance to atmospheric correction residuals and improved mission continuity with Landsat 8 Operational Land Imager Band 7. For prediction of NPV over a full range of green vegetation cover, the Cellulose Absorption Index, derived from 2040, 2100, and 2210 nm bands, was top performer (R2 = 0.77, RMSE = 0.17), but required a narrow (≤20 nm) bandwidth at 2040 nm to avoid interference from atmospheric carbon dioxide absorption. In comparison, broadband NPV indices utilizing Landsat 8 bands centered at 1610 and 2200 nm performed poorly in measuring fractional NPV (R2 = 0.44), with significantly increased interference from green vegetation.

show abstract

A Comparison of Estimating Crop Residue Cover from Sentinel-2 Data Using Empirical Regressions and Machine Learning Methods

Cited by 28 publications

References 60 publications

Assessing Non-Photosynthetic Cropland Biomass from Spaceborne Hyperspectral Imagery

Assessing Non-Photosynthetic Cropland Biomass from Spaceborne Hyperspectral Imagery

Dry season forage assessment across senegalese rangelands using earth observation data

Evaluation of SWIR Crop Residue Bands for the Landsat Next Mission

Contact Info

Product

Resources

About