Automatic Detecting Rice Fields by Using Multitemporal Satellite Images,Land-parcel Data and Domain Knowledge
Yi-Hsing Tseng,pai-hui hsu and Huang Chen
Department of Surveying Engineering
National Cheng Kung University
Taiwan, Republic od China
E-mail: tseng@mail.ncku.edu.tw
Keywords: Rice, Inventory, Image Interpretation, Multitemporal Images, and Temporal Profile.
Abstract
A remote- sesing technology for automatically detection rice field is proposed. The principal of this technology is applying a region based classification by means of integrating geographical data and domain knowledge with multitermporal Image. Based on the principal, there methods of investigating the temporal NDVI profile to detect rice fields were implemented. They are profile Matching (PM), peak Detection (PD), and difference classification(DC). All of the methods were tested on a set of Multitemporal SPOT XS images (12 epochs) collected during the second rice season of 1993. comparing to the traditional supervised classification using a single image epoch, all the methods can easily improve the accuracy about 20 %. The PM and PD method. When the number of image epochs is small the PM and PD methods may not work, but the DC method works well even if there are only 2 or 3 epochs. All of the methods do not require any training data for classification. We expect that this approach will dramatically reduce the needs human work and increase the efficiency of the rice inventory work.
Introduction
Inventorying the rice corp. overall the country for each rice season has been an important task of the Taiwan government fro decades. The purposes of carrying on the task are to monitor and control the overall rice product and to provide evidences for the government to compensate farmers when planted crops are damaged by disaters such as typhoons or floods. The current inventory system is founded on a region -based photogrammetric procedure. Aerial photos will first be visually registered onto the corresponding maps of land ownership. Manual photo interpretation is then implemented to determine each land-ownership polygon is a rice field or not. Polygons of rice field will be marked and counted and their areas will be summed to estimate rice product.. this system is developed based on the fact that most boundaries of rice field coincide with the boundaries of land parcel
In order to reduce the cost of the inventory work, using satellite images is suggested to avoid taking aerial photos [Huang,1984]. In this paper, to put into practice, we develop an automatic image interpretation to reduce the need of manual work. The proposed method is a region-based classification by means of integrating land-ownership data domain knowledge with Multitemporal SPOT imagery. The region-based classification is used based on the concept of integrating geographical data and remotely sensed images proposed by Johnsson (1994) and Derenyi & Tuerker (1996)
The consideration of domain knowledge is critical when applying Multitemporal images to image interpretation [Argialas and Harlow.1990]. The knowledge of rice crop including the local crop calender, time table of rice growing, as well as the variations of spectral reflectance within a rice season should be taken into account. Based on the knowledge, the spectral variation derived from a time series of Multitemporal images becomes meaningful and could be used to detect rice field automatically. This idea is similar to the use of temporal profile or temporal differences of spectral properties in image classification proposed by Brisco &Brown [1995],Lo et al.[1986] and Wolter et al. [1995].
The Use of Land-ownership Data
The basic idea of a region-based classification is to determine the region class by means of regional staistic measures of spectral data [Derenyi and Tueker,1996]. The applied images should be segmented into regions, and the spectral means of a region can be calculate for classification [Johansson,1996]. Because the unit of inventory is land parcel, land-ownership data are approprite to defining regions. The classification requires that a single spectral class should dominate each region. In most cases, this requirement can be satisfied, because the boundaries of land parcel mostly coincide with field boundaries and each field usually planted only one kind of crop. The exceptions are usually less then 1% of the total number of field in an agricultural area.
There are two steps to segment a satellite image into region based on the corresponding land-ownership data. First, the image should be registered to the vector data of land-ownership. Second, extracting and storing the spectral data of pixel within each polygon as a data set. In order to determine that in which polygon a pixel locates, one can use the center point of the pixel to perform a point-in-polygon check. This procedure can also avoid that a pixel is assigned to me\ore than one polygon.
The data arrangement mentioned above transforms raster data structure to object-based data structure. All of the data in a data asset are associated with a single region. It is then easier for us to calculate statistic measures of spectral data for each region and to attach the statistic measures to the data set. It is also convenient to add other attribute data further application. Furthermore, such data arrangement is suitable to the integration of remote sensing and geographical information system.
The Use of Domain Knowledge
Discipline specific knowledge including information about spectral, temporal, and structure properties of objects is critical to many inventorying and monitoring works using remote sensing. For rice inventory using remote sensing imagery, the following item are the domain knowledge should be taken into account: the local crop calendar, the time table of rice growing, and the variations of spectral reflectance of a rice field within a rice season
Most rice fields in Taiwan have two planting seasons a year. Each period of rice season is about four mouths. In general, the first season starts in spring and harvests in summer. The second season usually begins right after the end of the first season and harvests before winter is coming. Usually the fields in an agricultural area have similar calendar of planting. However, different agricultural areas may have time differences of rice requirement of using Multitemporal images to detect rice fields.
Basically the whole season of the rice growing can the divided into 5 periods:
-
Transplanting: fields are covered by water with little vegetation;
- Growing :fields are getting vegetation entropy;
- Reproducing : vegetation entropy reaches the maximum and starts to decrease slowly;
- Mellowing :vegetation entropy continuously a little;
- Harvesting :fields become bare soil with a little crop residue.
On should keep in mind that the land cover of a rice field is changing during a rice season. On the one hand, knowing this phenomenon is the requirement of choosing appropriate image epochs. On the other hand, this knowledge given us the clue to distinguish rice fields from other land-use types by using Multitemporal images.
Knowing the timetable of rice growing, a remote-sensing expert would not have difficulties to imagining the variation of spectral reflectance of a rice field within a rice season. The pattern of the temporal spectral variation of a rice field provides solid evidence to identify rice fields. This is the most important reason that we suggest the use of Multitemporal images.
