GEOG 6120 Advanced Optical Remote Sensing

Exploratory Analysis of Crop Classification through Remote Sensing Techniques

Introduction

Remote Sensing and precision agriculture have a new and unfolding relationship. Remote sensing can be more cost effective for decision makers, especially when dealing with larger study areas. Current technology allows for analysis on a finer scale than previously performed. Application of certain plant based indices, such as NDVI and LAI, has proven effective at identifying geophysical conditions of the underlying imagery.

There are some known issues with using remote sensing for land classification. In order to determine the accuracy of the classification, ground data must be collected to verify the analysis. This action requires field collection and can be costly both in resources and time.

The objective of this project is to attempt several methods of crop classification in order to determine their effectiveness. Due to the limitations of this project, no ground data has been collected for verification. In place, the cropland layer from the US Department of Agriculture will be used to perform a remote verification. An example of the cropland data layer can be seen in Figure 2.

  • The following imagery sources will be used from the California agriculture belt

    • Aviris-NG - Northern California (see Figure 1a)

    • EO-Hyperion - Bakersfield (see Figure 2)

  • Spectral profiles gathered from the below sources will be used in the analysis

    • Lab/Greenhouse

    • Field data

    • Image based

Anticipated Outcomes:

There are some expectations for the the outcome of this analysis. The first is to determine through qualitative or quantitative assessment, the most effect means of applying remote sensing techniques to crop land classification. Another expectation is that many crops will likely be classified similarly owing to similarities in their spectral profiles. Below are listed the crops expected to create confusion in the classification:

  • Leafy ground crops (high LAI):  Zucchini, tomato, strawberries

  • Grasses: Alfalfa, barley, wheat

  • Orchards and Trellised crops: Trees, vines

Data

Aviris NG Data Overview

The study are is located in northern California near Sacramento, see Figure 1. The imagery is from June, 2018. This time frame coincides with the cropland data available through the USDA and an optimal crop growth time frame before harvesting. The imagery has 240 bands, and 10nm bandwidth. Below is a list of the crops that can be found in the study area for which there is a corresponding spectral profile:

  • Tomato

  • Soybeans

  • Sunflower

  • Cucumbers

  • Zucchini

 
 

Earth Observing (EO)- Hyperion Data Overview

This study area is in southern California focused around the agricultural fields outside of Bakersfield The imagery is from June, 2016. For this study are USDA data from 2016 was utilized. Hyperion imagery has 240 bands, and 10nm bandwidth. Imagery is provided as a 30m resolution. Below is a list of the crops that can be found in the study area where the cropland data correlated to obviously discrete crop fields:

  • Corn

  • Pistachios

  • Alfalfa

  • Almonds

  • Cotton

  • Barley

Spectral Profiles - Lab

Lab based spectral profiles used in analysis. The Wisconsin data contains many different lab and field collected spectra for corn. In the Serbin dataset are various types of crop spectra, of which five were utilized in this analysis for their overlap

 
 

Spectral Profiles - Imagery

Image based spectral profiles, see Figures 5 & 6. Collection of these spectra were performed using visual inspection of pixels and the cropland data layer.

 
 

Methods

There a several approaches to classifying hyper-spectral imagery into land use types that involve spectral matching algorithms. The methods used in this analysis are Spectral Angle Mapping (SAM), Iterative End-member Selection (IES) and Multi-Range Spectral Feature Fitting (MSFF). These were chosen with consideration of the availability of spectral libraries and the characteristics of the spectrum in these libraries. All three classification were performed using ENVI software, specifically ENVI Classic. The IES classfication also utilized the ViperTools extension in ENVI.

Spectral profiles collected in labs were used on both images. The Serbin library was used in the spectral matching algorithms for the northern California study area, and the West Madison library was used with the Bakersfield imagery. Spectral libraries were also collected using the image pixels in both study areas, see Figures 6 & 7 . Pixels were compared against the cropland layer to determine what crop was cultivated in specific pixels to acquire plant profiles and then deployed in the classification process.

SAM can be very effective in classification as pixels within an image are compared to every other pixel in the image to determine their similarity. The pixel values are treated like vectors and plotted on an axis. Angles between pixels are compared, the smaller the angle the more similar the pixels are. Pixels are then grouped into categories that represent a collection that are more similar to each other than other pixels. This process is done until relatively discrete categories are established, and an output image with the categories reflected as pixel values is produced.

Many of the spectra used in the classification had similar spectral profiles so a semi-automated process, IES, was utilized for both study areas to determine the best end-members for classification. IES uses the input spectral library to classify imagery by identifying which end-member provides the highest accuracy in classifying the whole image. Then the process is run again adding additional end-members until the accuracy no longer improves with the addition of end-members. While IES was performed on the northern California study area, the process was not successful in producing any viable output and so has not been included in the Results section below.

The third classification method used was MSFF, which uses spectral features from multiple wavelengths in a spectral profile to compare to the image. Users can identify these features through a user interface with the ENVI software tools. An output classification is created based on each pixels similarity to the spectral features identified. Spectral features used in this analysis can be seen in Figures 11 - 15 for the AVIRIS- NG imagery and Figures 30 & 32 for the EO-Hyperion imagery.

Results:

AVIRIS-NG Spectral Angle Mapping

Output from SAM performed on AVIRIS-NG Imagery, Figures 8 - 10. Figure 8 is a subset of the original AVIRIS-NG image, while Figures 9 & 10 are the same area after the SAM analysis was performed. Figure 9 reflects the SAM results utilizing the Serbin spectral library, which were collected in a lab. The next image in Figure 10 is the result of SAM utilizing the spectral library generated from the image, using the cropland layer to generate regions of interest (ROIs). It is clear from the images that SAM was not effective in classifying the agricultural plots, the only features clearly are the water features on either edge of the image.

AVIRIS- NG - MSFF - Serbin Dataset Spectra

Below in Figures 11 - 15 are the spectral features from the Serbin spectral library identified for use in MSFF on the AVIRIS-NG imagery. There is an obvious similarity in the spectral profiles between several of the crops and this is reflected in the resulting classification imagery in Figures 16 - 20. For much of the imagery, there is a distinction between agricultural plots and surrounding features, however the accuracy of the classification ends there. Many of the plots are classified with varying degrees of confidence as the same crop.

Hyperion Spectral Angle Mapping

They Bakersfield data produced better results from the SAM analysis. As shown in Figures 21 - 23, there are distinct classes identified in the imagery which aggregate on plots. There is little increase in accuracy after setting the similarity threshold at 0.2 radians, the difference between Figures 22 & 23 is minimal. While there is still much confusion in the classified images, this is an improvement to the SAM output from AVIRIS -NG imagery. Individual plots are clearly recognizable and are classified with regularity as the same crop type.

 
 

Hyperion - IES Fraction Results

The IES analysis performed similarly to SAM for they Hyperion imagery, producing results that indicate clearly the extent of individual plots and distinguishes, with some degree of accuracy, between crops. Output from IES Fraction analysis can be seen below in Figures 24 - 29. There is some confusion between Alfalfa and Barley and between Corn and Cotton, however there was an expectation that plants with similar spectral profiles would be classified similarly. This method distinguishes orchard like crops, such as pistachios and almonds, very well. This may be a result of the difference in number of pixels with crops in them, plots with orchard like crops often have more mixing between bare soil and the crops.

 
 

Hyperion - MSFF - West Madison Corn Spectra

The MSFF analysis was carried out using several distinct spectral features from the Wisconsin spectral library. Below, figures 30 and 32 illustrate a selection of the spectral features. Figures 31 and 33 are the output from each MSFF analysis. It is interesting to note the similarity in the classification of corn as a result of IES and MSFF. It is clear MSFF is able to distinguish individual plots, and classifies distinct areas within the Hyperion imagery as definitely corn and other areas as definitely not corn.

 
ARS_10.2.jpg
 

Figure 30 Spectral feature - absorption in visible spectrum ~ 500-600nm

 
 

Figure 31 MSFF Output for spectral feature at 500-600nm

Hyperion - MSFF - West Madison Corn Spectra

 
ARS_11.2.jpg
 

Figure 32 Spectral feature: Absorption in visible spectrum ~ 500-600nm & ~ 1100 - 1200nm

 
 

Figure 33 MSFF Output for spectral feature at 500-600nm & 1100 - 1200nm

Conclusion

It is clear that the Hyperion data performed significantly better than AVIRIS-NG imagery in this classification project. Outputs from IES, MSFF and SAM produced meaningful results that can be pursued further in other analysis. The amount of confusion in the classification of AVIRIS-NG data might be related to the types of crops being classified, and the similarity in spectral profiles. The imagery collection was larger, which may have also impacted the classification process.

There was an expectation that the similarity of spectral profiles of crops would make it difficult to differentiate certain crops. For example orchard and grasses, and grasses and fallow land have similarity in their spectral profiles. This causes confusion in the classification output. The data used for ground verification, the cropland layer from the USDA, has certain issues for this scale of analysis. The accuracy and quality of the data is difficult to determine, and visual inspections shows there to be discrepancies between the data and actual crops being cultivated.

Performing the analysis using lab spectra and field collected spectra illustrates some of the issues with this type of analysis. The locality of the spectra had an impact on the output, specifically using the West Madison crop spectra to classify imagery from California. Also using lab and image based spectra allowed for an overview of the effectiveness of these methods. Overall the image based spectra were more effective at classifying the related imagery, which was expected.

This analysis would benefit in an emphasis on crop spectra and building libraries for crop data. Many of the spectral libraries have spectra for invasive and riparian plants, very few crop spectra have been cataloged. Among the crop spectra that is collected, most are not staple crops or reflect the breadth of the agricultural industry.

References

Data Sources:

AVIRIS-NG: Scene - f180608t01p00r10

EO - Hyperion: File - EO1H0420352016203110K

CropScape Data Layer, USDA.  https://nassgeodata.gmu.edu/CropScape/

Couture, John.  University of Wisconsin Environmental Spectroscopy Laboratory, 

http://labs.russell.wisc.edu/townsend/

ly K.S. Serbin S.P. Lieberman-Cribbin W. Rogers A.. 2018. Leaf spectra, structural and biochemical leaf traits of eight crop species. Data set. Available on-line [http://ecosis.org] from the Ecological Spectral Information System (EcoSIS). doi:10.21232/C2GM2Z

Literature:

Delegido, Jesús & Verrelst, Jochem & Meza, C.M. & Rivera Caicedo, Juan & Alonso, Luis & Moreno, Jose. (2013). A red-edge spectral index for remote sensing estimation of green LAI over agroecosystems. European Journal of Agronomy. 46. 42-52. 10.1016/j.eja.2012.12.001. 

L. Garcıa Torres, J.M. Pena-Barrag, F. Lopez-Granados, M. Jurado-Exposito, R. Fernandez-Escobar, Automatic assessment of agro-environmental indicators from remotely sensed images of tree orchards and its evaluation using olive plantations. Computers and electronics in agriculture 61 (2008) 179–191

L.F. Johnson, D.E. Roczen, S.K. Youkhana, R.R. Nemani, D.F. Bosch, Mapping vineyard leaf area with multispectral satellite imagery.  Computers and Electronics in Agriculture 38 (2003) 33/44

Mulla, David J., Twenty five years of remote sensing in precision agriculture: Key advances and remaining knowledge gaps, Biosystems Engineering 114 (2013) 358 e371

Raffaele Casa, Fabio Castaldi, Simone Pascucci, Stefano Pignatti.  Potential of hyperspectral remote sensing for field scale soil mapping and precision agriculture applications.  Italian Journal of Agronomy 2012; volume 7:e43