IEEE Geoscience and Remote Sensing Magazine - December 2020 - 80

to the joint use of spectral and spatial information. This
indicates that, besides the rich spectral information, spatial information is also important for HSI classification.
◗◗ PCNN and CNN versus CAE and CRNN: Similar to SAE,
CAE focuses on image reconstruction rather than classification. In contrast, PCNN and CNN are exclusively
designed for the classification task, so they are able to
learn more discriminative features than CAE, leading
to better classification performance, especially on the
Houston 2018 data set. Although CRNN also focuses on
the classification task, it has more parameters to train.
Using the same number of training samples and epochs,
PCNN and CNN can achieve better results than CRNN
in terms of AA, OA, and kappa.
◗◗ PCNN versus CNN: PCNN outperforms CNN in terms of
classification accuracies for all three data sets. We should
note that the improvements are substantial in the case of
the Houston 2013 and 2018 data sets. Due to the use
of PCA, most of the redundant spectral information is
reduced. Therefore, the number of trainable parameters
in PCNN is smaller than that of CNN, making it easier
to learn under the same conditivon.

is time consuming due to the distribution matching
between input samples and labels. The requirement to
estimate the statistical distribution also limits LSDR's
stability, especially when the training set is available on
a small scale (e.g., for the Indian Pines 2010 and Houston 2013 data sets).
◗◗ LDA versus JPlay: Unlike conventional regression techniques, JPlay is capable of extracting semantically meaningful and robust features, due to the multilayered structure
and self-reconstruction constraint (32). Quantitatively
speaking, JPlay outperforms the other SFE methods. The
CV provides a feasible solution to automatically determine the parameter combination in JPlay. Despite the
ADMM solver designed for speeding up the optimization
process, such a multilayered parameter update inevitably
suffers from high computational cost.
DEEP FE
◗◗ Spectral versus spectral-spatial models: Most of the spectral-spatial models (i.e., CNN, PCNN, CAE, and CRNN)
achieve superior performance compared to spectral models
(i.e., SAE and RNN) in terms of AA, OA, and kappa due

TABLE 7. THE CLASSIFICATION ACCURACIES OBTAINED ON FEATURES EXTRACTED FROM
THE HOUSTON UNIVERSITY 2018 DATA SET USING DIFFERENT SHALLOW AND DEEP FE TECHNIQUES.
HOUSTON 2018
SHALLOW FE

DEEP FE

UFE
SPECTRAL PCA
1

0.3088

0.8781

SFE

MSTV

OTVCA

LPP

LDA

CGDA

0.0536

0.6842

0.6618 0.6256 0.7575

SFE

LSDR

JPLAY

SAE

RNN

CNN

0.7969

0.5991

0.794

0.5702 0.7516

CAE

CRNN

0.4428 0.6338

PCNN
0.6638

2

0.7603

0.8396

0.7046

0.6376

0.8122 0.8474 0.8076

0.7747

0.8347

0.7893 0.6975 0.8173 0.8849 0.8707

0.8376

3

1

1

1

0.9972

1

1

1

1

0.9972 0.7739 0.8482 0.9924

0.8045
0.9595

1

1

4

0.9134

0.9494

0.6238

0.6775

0.9453 0.9059 0.9265

0.9276

0.9298

0.9221 0.8613 0.9444 0.9362 0.9439

5

0.4119

0.4668

0.2676

0.1679

0.4728 0.5258 0.4661

0.4289

0.3971

0.4982 0.404

0.433

0.5396 0.5404 0.48

6

0.257

0.299

0.3835

0.3164

0.3008 0.291

0.2776

0.2726

0.278

0.2585 0.2537 0.305

0.2902

0.3377

7

0.3025

0.3025

0.3025

0.3025

0.3025 0.3025 0.3109

0.3025

0.2857

0.3025 0.3025 0.2908 0.2723 0.2997

0.3176

0.308

0.7356 0.8538 0.8583 0.8092 0.8677

8

0.7657

0.7675

0.7599

0.7417

0.7785 0.7849 0.7544

0.7518

0.7771

0.7216

9

0.3849

0.3877

0.5767

0.599

0.4887 0.3917 0.3672

0.5255

0.5877

0.6302 0.4186 0.7970 0.7371 0.3717

10

0.3603

0.436

0.3747

0.4491

0.423

0.3497

0.401

0.3819 0.3465 0.5902 0.4957 0.5484 0.5778

0.4086 0.379

0.8659

11

0.4162

0.4792

0.7862

0.7596

0.5085 0.4667 0.4266

0.4422

0.5359

0.4143 0.4699 0.5456 0.5781 0.6048 0.5948

12

0.0132

0.0046

0.0093

0.0077

0.007

0

0.0302

0.0152 0.0697 0.0511

0.0477 0.0927

0.0579

13

0.4525

0.5556

0.4238

0.409

0.5442 0.5164 0.5707

0.5603

0.5324

0.4523 0.4789 0.5148 0.5619 0.4246

0.5811

0.0023 0.017

14

0.3019

0.2629

0.546

0.4665

0.3651 0.4152

0.2073

0.3212

0.3789 0.3309 0.5289 0.6763 0.3375

0.5705

15

0.6303

0.4721

0.4457

0.4887

0.4602 0.5549 0.418

0.5234

0.5944

0.5197 0.5289 0.6277 0.6476 0.6447

0.6591
0.8572

0.2294

16

0.6412

0.7611

0.622

0.7648

0.7559 0.5888 0.7374

0.6403

0.6688

0.7457 0.7086 0.8498 0.7594 0.7173

17

1

1

1

1

1

1

1

1

1

18

0.4983

0.6885

0.6576

0.5197

0.714

0.6581 0.6625

0.5686

0.7366

0.5346 0.608

0.9545

1

0.9545 0.8909 1

1

0.6365 0.5981 0.7692

0.702

0.8006 0.9476

19

0.5265

0.6363

0.906

0.8777

0.7323 0.6989 0.61

0.6266

0.6771

0.6569 0.6545 0.9102 0.896

20

0.4444

0.8904

0.9706

0.5253

0.8479 0.9189 0.6797

0.5955

0.5393

0.5388 0.4801 0.6246 0.5566 0.4879

0.6519

AA

0.5195

0.6039

0.5707

0.5696

0.606

0.5952 0.5676

0.5647

0.5863

0.5777 0.5458 0.64

0.6667

OA

0.4634

0.5101

0.575

0.5899

0.5552 0.5027 0.4825

0.5492

0.5944

0.5938 0.4851 0.7278 0.6969 0.5116

0.7728

l

0.3732

0.4317

0.4833

0.4974

0.4714 0.4231 0.4018

0.456

0.5037

0.4948 0.3936 0.6474 0.6124 0.4372

0.7011

0.6268 0.609

The highest accuracy in each row is shown in bold.

80

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2020
IEEE Geoscience and Remote Sensing Magazine - December 2020

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2020
