Coarse-Fine classification of landsat image using Neural Network

Coarse-Fine classification of landsat image using Neural Network

Kozo Okazaki*, Yutaka Fukui*, Yoshihiro Nobuoka*,
Hiroshi Mitsumoto**, Shinichi Tamura***, Guo-Fang Shi***
*Faculty of Engineering, Tottori University
**Faculty of Engineering Science, Osaka University
***Osaka University Madical School

Takashi Hoshi+, Kiyoshi Torii++,
+Institute of Information Sciences and
Electronics University of Tsukuba
++ Faculty of Agriculture, Kyoto University

Abstract
We have already presented a neural approach to the landsat window image classification by personnal computer based system. Here, we used a 32 bit personal computer (NEC PC-9801 RA) , Hyper frame board memory (HFBM,3 planes) and Image Pipeline Processing (Im PP) board. In this case, however, the classification is often rough in some areas. If we make the size of window smaller in order to resolve this, we need much more time for processing. Considering the classification, a larger window is reasonable for a region such as sea, lake, etc. and a smaller window is suitable for a compound area.

In this paper, we propose a method of coarse-fine classification of landsat image using error back propagation algorithm (BP), where we used several sizes of windows. Furthermore, we made up the software of using two HFBMs and applied them to BP for processing the 4ch. image data. .

Introduction
We have propose a neural network approach to the remote sensing image data [1]. Here, multi-channel image data composed of neighboring pixels are used as input to the back propagation network. The training is done by error back propagation (BP) algorithm. The classification of multi-spectral remote sensing image is, usually, based on multi-variable analysis. This method examines the statistical characteristics for every pixels. Generally, the image shows a marked trend of being classified excessively. However, classifications of he ample flat area including sea, lake area, etc. need to use more large regions for processing. This paper deals with a method of coarse-fine classification of landsat image using BP, where we use several sizes of windows. Furthmore, we made up the software of using two HFBMs and applied them to BP for processing the 5ch . image data. .

System configuration
We used a personal computer based system; computer (NEC PC9801 RA), ImPP neural board (NEC), Neuro07 software (NEC Information Technology Co. Ltd.), Hyper Frame (Digital Arts Co. Ltd.) and 80387 co-processor (Fig. 1). The classification of image is carried out at high speed by ImPP and co-processor. Usually, one HFBM (64x400pixelx3plane) is used. We made up the software of using two (three or four boards are possible as well) HFBMs. We can process the six landsat channels data as the maximum. .
Fig. 1 System Configuration by personal computer

Neural network model
Due to simpleness of the calculation algorithm and excellent ability of learning , BP is widely and actively used for many fields [eg. [2]]. Fig. 2 shows a schematic diagram of neural networks for multi-input, multi-output. .

Fig. 2 Error Back Propagation

The number of the layers of BP is three. The values of the weight of the BP are initialized randomly. The data we use is Ishigaki Island (Japan) got from landsat 3 (4ch.). The images are classified by the window size. We call this method as window size method (WSM) for short. Windows for teaching are specified by a mouse on the Hyper frame plane with its category (sea, cloud, coast, plain). . input layer : (10x10pixel) x3ch.
hidden layer :10.
output layer : 5.
Display : sea-blue, cloud- red, coast-violet, plain-green.
Fig. 3 Original Image

Fig. 4 An example of classification with window
size 10 X 10 (3 Ch. are used. Number of categonies is 4.
Gray level of the color is propotional to the activation.)

Images of composite RGB image are shown in Fig. 3. Fig 4. shows the results of classification. The gray level of the color is propotional to the activation. Most parts of area are recognized correctly, but in some part, fault. Table 1 is some of the detailed result of BP output layer activation levels (AL).

Table 1. Ssome detailed results of BP output layer

	Sea	Cloud	Coast	Plain
1	0.8937	0.0154	0.2000	0.0196
2	0.0058	0.0476	0.0682	0.9444
3	0.0603	0.1972	0.0780	0.1783
4	0.1006	0.1957	0.0745	0.1381
5	0.4451	0.0479	0.4551	0.0262
6	0.0117	0.4306	0.4554	0.0887
7	0.0045	0.2810	0.2440	0.4779
8	0.0385	0.4108	0.1694	0.0604

	Sea	cloud	coast	pain
(ex.) No. 1	0.8037	0.0154	0.200	0.0196

This area (No. 1) is decided as "sea", because its active value=0.8937 is max. AL of (No. 1) and No.2 are high and the difference between the maximum and second one (DBMS) is large. The classification is carried out correctly. AL of No. 3 and No. 4 are low and the DBMS is small. AL of No. 5 and No. 6 are relatively high, but the DBMS is small. AL of No. 7 and No. 8 are relatively high and the DBMS is relatively high. Many mis-classifications occur in cases of No.3-No.6. For these cases, we need to classify once more using smaller window. .

Coarse-Fine classification
The resolution corresponding to one pixel of landsat image data used in the experiments is 80 m. For the purpose of increasing the classification accuracy, the experiment of reducing the window size of learning/ recognition to a pixel was done. However, in this case the learning has not converged. In BP processing by pixel unit, the method using pseudo-variance in neighboring 3x3 window can be considered, and it's now under consideration. Alternatively, it can be considered that the weight of recognizing window of 10x10, 5x5, 3x3, 2x2 size etc. which have clearly obtained by the learning, in used to classify the center pixel of the window by moving one by one. We call this method as window center method (WCM) for short. .

Although there is a problem of processing time (about 3.5 hours), it is done and compared with the original image. Then it agreeded well with inspecting consideration. Using the BP classifying the data based on the pattern shape, the experiment for the size of 5x5 (about one hour) become better in the boundary and compound region (WCM). The obtained results are good one the whole. The classification (3 ch. and 4categories) results are shown in Fig. 5.

Fig. 5 Result of WCM ( Window 5 X 5, 3 Ch. Number of categories is 4. )

Therefore, the learning window size is not simply reduced, but it is taken largely in ample uniform regions. The window size of 5x5 is applied in complicated boundary or compound regions, hereupon, and the category of the center point of window is determined one by one by moving the window. In this way, the coarse0fine classification has been done by the multiple window size. Fig. 6 and Fig. 7 show the results of classification using 4 channel image data. Shadow of cloud category is added and displayed by black. Window size is 5x5 and 3x3 respectively. Classification is done by WSM and WCM. In Fig. 8, it is shown the gray leveled results: first; it is classified initially by 10x10s9ze (WSM). second; the WCM is applied to re-classify by 5x5 sized for the regions with lower activation than 0.6 As the classification results the activation in the boundary and compound regions gets higher value than that of obtained from the experiment with one size, and the classification accuracy becomes higher and reduction of the processing time ( about 35 minutes) is possible. The more the size of window area, the more we easily recognize the features of the area, and spend less time for learning, but the classification becomes coarse instead. .

Fig. 6 Classification by 4 Ch. 5 categories

Fig. 7 Classification by 4 Ch. 5 categories (the same as Fig. 6)

Fig. 8 Coase-fine classification (3 ch. and 4 categories )
1st step: window 10 X 10; WSM 2nd step: window 5 X 5; WCM

Conclusion
In this paper, using a neural network the coarse-fine classification to the land-sat image has been done. Although the unnatureness of classification remains near the boundary region in the processing with large rectangle window, speed up of processing by coarse-fine method is realized. But the category" shadows of cloud" has not been dealt with in the coarse-fine method. Including this, the comparison and estimation with the conventional method are being considered. .

In case of classifying the data including the noise, diversive or unstable component such as the landsat image with some categories, when the window size is small, especially 1x1, the classification by BP becomes difficult. To cope with this, the energy minimizing approach can be applied [3]. As the energy; the term of the difference between the neighboring pixel, the term of constraint such as one pixel necessarily belongs to one category, etc. are considered. Also, the relization of highly adaptabe classification method by neural network which includes the method of detection and normalization of pattern location [4] are the subjects to be applied. .

We programmed the BP on PC-9801 RA and Excel image processing unit (A vionix Co. Ltd.) using MS-C Ver. 5.1. Next , we made up the system using one HFBM instead of the Excel using MS-C for cost down. As the third stage we made up the present system with two HFBMS. We need not necessarily use the ImPP board, but can process the image more than 10 times speedy using it. .

Reference

Okazaki K. et. al. Neural Approach to Remote Sensing Image Classification, 10th ACRS, Malaysia, pp. H2.1-6, 1989.
Jian-Tong W. et. al.: Neuro-Voice- Recognition jointly Using Mouth Shape Image and Voice Features, Trans. IEICE (kn Japanese), J73,-D-II, 1990.
Mitsumoto H. et. al : Lip Contour Extraction, Complement, and Tracing by Using Energy Function and Optical Flow, Trans. IPSJ (in Japanese), Vol. 31, No 3, pp. 444-453, 1990.
Okazaki K. et. al. : Detection and Normalization of Pattern Location by Neural Network, Pacific Rim Int. Conf. on Artificial Intelligence ' 90, Japan, 1990.