Automatic Detecting Rice Fields by Using Multitemporal Satellite Images,Land-parcel Data and Domain Knowledge
The Use of Multitemporal Images
Using Multitemporal images is helpful inn many application of image interpretation. In this study, the use of Multitemporal images is to increase the potential of differentiating rice field from other land-use types. It should be applied based on the domain knowledge mentioned above. Due to the land-cover variation of rice fields within a planting season, the difficulties can be expected in recognizing rice fields by using a single image epoch. If one uses images taken during the period of transplanting, one would not digtinguish rice fields from wet lands. Similarly, using images of the growing period we may confuse with other vegetation area, and using images taken after harvestinating we will be able to differntiate them from regions of bare soil. However, by investigation the variation of spectral responses of Multitemporal images, we can minimize the possibility of misinterpreting .
Checking temporal variation of sprectral responses has proven effective in agricultural land-cover and forest classification [Lo et al.1995]. Lo et al. analyzed temporal profiles of ratio or green vegetation index to identify crop types. Based on the similar idea, Wolter et.al. utilized temporal NDVI ( Normalized Difference of vegetation Index) differences to distinguish forest types. Therefore the temporal NDVI profile in a rice season investigated. According to the report of Huang et,al. [1985] and some sample data, the temporal NDVI profile of a rice field can be described as in figure 1. Some useful information could be extracted from the temporal profile for distinguishing rice fields, for example the curve shape of the profile or the NDVI differences among epochs.
Figure 1. The expected temporal profile of vegetation index in a rice season (the star symbols represent the value of sample data)
The second use of Multitemporal image is decreasing occlusion of cloud. It is in common that images are partially clouded in Taiwan, so that we frequently need to combine more than one image to get a complete visibility. In order to reduce the effect of land-cover changes, the images to be combined should be taken closely in data.
Automatic Detection of Ratio Fields
Three methods were developed to recognized rice fields from SPOT images. They are named profile matching, peak detection, and differences classification.
Profile Matching (PM) An expected temporal profile similar to the curve in the figure1 could be obtained, if an agricultural fields is planted rice. By comparing the expected profile with each temporal profile of the agricultural fields,one could distinguish rice fields from the other agricultural fields. Automatic comparison can be completed by matching the curves using the cross-correlation method (figure2). The threshold value of the cross-correlation function to classify fields into rice and non-rice should determine in advance. Evenly distributive image epochs over the rice season are required to obtain a good result. In addition to detecting rice fields, this method could also determine the planting time of a rice field.
Peak Detection (PD) The temporal profile of the vegetation index in a rice season is a curve of a mountain shape. The peak represents a growing of vegetation in the field. If a peak that locates within a rice season and whose bottom spans about a time period of a rice-growing cycle is detected in the temporal profile of a filed, this field has the evidence of being planted rice. To detect the peak, one can set a horizontal line to intersect with the temporal profile, so that a pair of intersections defines a peak which width and height are estimated as in figure3. There is no rigorous rule to set the NDVI level of the horizontal line. In general, it should be a little bit larger than the NDVI of bare-soil pixel. At least three image epochs those distribute in the beginning, middle, and end of the rice season are required. This method could also roughly determine the planting time of a rice field.
Figure.2 The temporal profile matching using the cross-correlation method.
| 
Figure 3.The method of peak detection |
Difference Classification (DC) Instead of checking the shape of the temporal profile, one could also take the NDVI differences between some specific epochs as the features to classify the field. If rice season is equally divided into 3 period. Therefore, the NDVI obtained from an image epoch of the second period minus that of the first or the third periods will provided appropriate NDVI differences fro the classification. The features of such NDVI differences will increase the reparability between the polygon of rice an d non-rice (figure4), so that it improves the classification accuracy. Positive NDVI differences will be obtained if a field is planted rice. Unsupervised classifying the field into two groups, the group has the larger NDVI differences should belong to the class of rice field. At least two images that provide an appropriate NDVI difference are required.
Figure 4. NDVI differences the separability between rice and non-rice fields.
The Study Site
The study site located in Tainan County, Taiwan. It is atypical agricultural area in Taiwan. The area of the study site is about 7.7 km
2 . The average size of the agricultural fields is about 3000m
2. Rice crops are the major agricultural products in this area. Usually there are more than 70% of the fields being planted rice crops.
Twelve Multitemporal SPOT XS images were studied. The images were selected referring to the local calendar of the second rice season in 1993.The images are shown identification number and the collecting dates in figure 5. The land-cover changes are obviously presented in the time series of images.
The images were selected from the product list of the collected images by the Center for space and Remote Sensing Research, National Central University, Republic of China. All of the images were registered to the land-parcel map of the study area, and were resampled to the resolution of 12.5m by 12.5m pixel size.
The use Multitemporal Landsat Tm images was also considered. However, although Landsat TM images provide better spectral resolution than SPOT image repeatability frequency is generally too low to fit this application
The land-ownership data were digitized from a land-parcel map provided by the crop Bureau of Taiwan. The data contain 2874 polygons of land parcel as figure 6. These data provide us the divide the images into regions with respect to the agricultural fields.
Figure.5 The Multitemporal SPOT-XS images used for the study
The Crop Bureau of Taiwan also provides us data of the ground truth, which will be used to evaluate the results of our experiments. The data were generated by means of manual interpretation of aerial photographs. The data contents are attributes of land parcels that indicate whether a land parcel is a rice field or not. In the second rice season of 1993, there were 2043 rice fields,737 non-rice fields, and 94 mixed fields.
