Research Article: Modified shape index for object-based random forest image classification of agricultural systems using airborne hyperspectral datasets

Date Published: March 7, 2019

Publisher: Public Library of Science

Author(s): Eric Ariel L. Salas, Sakthi Kumaran Subburayalu, Changshan Wu.


This paper highlights the importance of optimized shape index for agricultural management system analysis that utilizes the contiguous bands of hyperspectral data to define the gradient of the spectral curve and improve image classification accuracy. Currently, a number of machine learning methods would resort to using averaged spectral information over wide bandwidths resulting in loss of crucial information available in those contiguous bands. The loss of information could mean a drop in the discriminative power when it comes to land cover classes with comparable spectral responses, as in the case of cultivated fields versus fallow lands. In this study, we proposed and tested three new optimized novel algorithms based on Moment Distance Index (MDI) that characterizes the whole shape of the spectral curve. The image classification tests conducted on two publicly available hyperspectral data sets (AVIRIS 1992 Indian Pine and HYDICE Washington DC Mall images) showed the robustness of the optimized algorithms in terms of classification accuracy. We achieved an overall accuracy of 98% and 99% for AVIRIS and HYDICE, respectively. The optimized indices were also time efficient as it avoided the process of band dimension reduction, such as those implemented by several well-known classifiers. Our results showed the potential of optimized shape indices, specifically the Moment Distance Ratio Right/Left (MDRRL), to discriminate between types of tillage (corn-min and corn-notill) and between grass/pasture and grass/trees, tree and grass under object-based random forest approach.

Partial Text

Broadband vegetation indices (VIs) reduce spectral data dimension by limiting the number of bands at different ranges of the electromagnetic spectrum to extract vegetation information from remotely sensed images. Mostly, the bands are selected from the visible and near/mid infrared regions in order to measure the photosynthetic activity of the plant [1] [2], vegetation dynamics [3], biomass abundance [4], predict crop yield [5], and biotic stresses [6]. This reduction of spectral information could pose some drawbacks such as index saturation beyond certain level when estimating high vegetation biomass [7] [8]. Another constraint in the use of existing broadband VIs is the challenge of choosing relevant band centers and widths [9] for agricultural management system mapping, particularly if it involves hyperspectral data where there is increased number of near-continuous bands. Under such circumstances, broadband VIs resort to using only average spectral information over wide widths resulting in loss of crucial information, such as little absorption features caused by the differences of spectral responses from agricultural fields that may be available in those specific narrow bands [10]. These hardly noticeable spectral absorption features could be the key for differentiation of landcover classes with similar spectra, as in the case of crop residue and soil.

The methodology was based upon the steps displayed in Fig 1, which included (1) processing and segmenting the images, (2) applying the random forest classifier, and (3) evaluating and assessing the results.

The modified MDIs added another breadth of possibilities in the analysis of hyperspectral images. The results demonstrated the potential significant challenges in mapping and classifying landcover, specifically vegetation/crops and their management practices, using traditional approaches.

We developed and proposed a new and optimized moment distance index to improve the spatial-spectral classification of hyperspectral data for agricultural management systems. We conclude, based on our goal to obtain better classification accuracies not only for vegetation classes but for other landcover types, that it is worth integrating optimized MDIs for object-oriented classification of hyperspectral images. However, it is still unknown how optimized MDIs would perform when hyperspectral bands are limited, say for instance limiting the distance between LP and RP near the 2100 nm cellulose absorption region, which other existing indices had utilized in mapping tillage systems. This is something worth looking into in the future. One thing that is certain, however, that with proper selection of variables—spectral indices, textural variables, and optimized MDIs—we could obtain relatively high classification accuracies for individual landcover classes.




Leave a Reply

Your email address will not be published.