Home Articles Monitoring small-scale farming using RS data

Monitoring small-scale farming using RS data

Bambang H. Trisasongko, Dyah R. Panuju, Laode S. Iman
Bogor Agricultural University, Indonesia
Email: [email protected]

Rice is an essential food in Indonesia along with cassava, corn and sago. In spite of the fact that its production and distribution have always remained one of the major concerns in government policies, its production is continuously reducing in some areas, mainly in the Northern Coastal Region of Java. In that area, uncontrolled urban sprawl has eaten up most of the agricultural lands. In past, using Landsat data, monitoring of agricultural land field proved effective for successfully identification of various rice growth stages in two locations in Java (Panuju et al., 2007). Another test using ALOS AVNIR-2 (Tjahjono et al. 2009) and a next generation Indonesian micro-satellite (Trisasongko et al. 2010) showed similar results.

Small-scale farming systems as seen in Indonesia require special consideration to the data. At the moment, there is no updated database on rice field at the high resolution available in Indonesia. However, it is crucial to accommodate timely rice monitoring. Previous attempt to classify all agricultural commodities in Indonesia has failed due to the complexity of land uses. Based on previous experiences, provision of high resolution rice distribution map is critical to this task. The role of low altitude aerial photograph has been replaced by high resolution satellite data such as IKONOS or QuickBird. In future, this two-step approach will assist in monitoring rice fields in a cost-effective way.

This paper discusses applicability of WorldView-2 high resolution data to provide high resolution rice field distribution map, especially to identify segmented rice fields. To extend the application, an outlook of routine rice monitoring is also examined.


Study site
The site was situated in Subang Regency, Indonesia (Figure 1). Subang is one of the main rice producers in the country, situated in northern region of Java Island. Rice planting can be easily found throughout the region. However, most productive region in Subang has been the Northern Coastal Region (NCR). Alluvial plain which is the most preferable land for rice planting is found throughout the NCR. The area is also supported by a large scale irrigation network from Jatiluhur multipurpose dam, making the area suitable for intensive planting scheme.

Figure 1. Administrative boundary of Subang. Site location is indicated by a black box.
Datasets and field observation
The main dataset for the experiment includes WorldView-2 (panchromatic and multi-spectral) data, which was corrected according to the Indonesian Base Map (Rupa Bumi Indonesia) at 1:25.000 scale. In order to assist the interpretation and analysis, additional data were exploited. Specific area selection was assisted by a rice coverage map derived previously from high-resolution multi-sensor observation, including ALOS PRISM, AVNIR-2, and SPOT-5 in 2009. The map was validated by a thorough field observation which collected 55 site identifiers.

Parcel edge (Galengan) identification
One of primary interesting investigations on a new sensor has been capability to identify agricultural parcel edge, locally called the galengan. The role of galengan identification is to identify segmented land parcel. Higher number of galengan indicates that the agricultural region is highly segmented, which in turn reduce optimal acreage and the yield. Segmentation in smaller land parcel could signify a diversity of parcel management; consecutively create complexity in yield data acquisition.

Irregular size of galengan mostly found in Indonesia generates complication in extraction from high resolution remotely-sensed images. To date, only wider galengans could be indicated from visual inspection or through digital methods. However, this is not the case of Subang, on in wider term, in Indonesia. Visual observation by means of ALOS PRISM data was failed due to insufficient spatial resolution. Higher resolution data such as pan-sharpened IKONOS are capable to identify a limited number of galengan. Most of the identification scheme has been pseudo-recognition which observes galengan from different land cover (or growth stages in this case) of the adjacent parcels.

In this research, two types of galengan viz major and minor galengans were of our primary interest. For this purpose, we employed a pan-sharpened WorldView-2 data using visual inspection. To assist the identification, a collection of field dataset was utilized.

Agricultural land cover mapping
Monitoring agriculture should be taken in a frequent way using various Earth observation data available. In order to support the task, spectral signature should be employed based on physical dataset. In this research, the raw digital number format was converted into radiance using specification delivered in the header.

To accommodate agricultural mapping through classification scheme, two decision trees were employed, namely QUEST (Loh and Shih 1997) and CRUISE (Kim and Loh 2003). Previously, QUEST was tested for remote sensing problems by Pal and Mather (2003). However, both algorithms have insufficient exploitation in an application perspective. In this research, classification accuracy was assessed through computing kappa coefficient from testing dataset.

Results and discussion

Identification of Galengans
Availability of very high resolution imagery such as WorldView-2 delivers a broader application in agricultural data capture. In Indonesia, a special case was made due to fragmented land parcels denoted by galengans. During field visit, there are at least two types of galengan, minor and major galengans. Comparison between their appearance in WorldView-2 image and field photograph is presented in following figures.

Figure 2. Minor galengan

Figure 3. Major galengan
As shown, WorldView-2 data has capability to capture major galengans with the width ca. 50-60 cm. Minor galengans, however, require special attentions. This is quite similar to those high-resolution imageries such as QuickBird or IKONOS. Minor galengans could be potentially identified if the land cover permits. Based on the field information, higher capability on capturing minor galengans could be made on dry fallow if the planting date between fields is quite similar. At the beginning of a new planting season where the fields are waterlogged, visual identification of minor galengans is fairly difficult.

Spectral signatures
During the field visit, data collection was taken place to acquire rice growth status. Whole data were separated into primary and testing data. Primary data were employed on spectral signature study and classification. Based on primary field data, spectral diversity of various rice growth conditions could be observed (Figure 4).

Figure 4. Extracted radiance data
Using extracted radiance, it is shown that dried fallow stage could be easily discriminated on all bands, except on yellow and NIR-2 bands. On the preparation of a new planting season, waterlogged rice field is benefited from a new WorldView-2 band; that is the Coastal band. Similar condition applies to transplanting. In addition, the new yellow band is an advantage as an aid to discrimination of transplanting stage.

Classification performance
In order to assist mapping task, this research employed two variants of decision tree algorithm, namely QUEST and CRUISE. Using primary dataset for training purpose, we found that decision trees derived from CRUISE and QUEST methods are slightly complex than our previous study using Landsat or ALOS AVNIR-2 data. The decision trees are presented in Figure 5 and 6.

Figure 5. Class discrimination by CRUISE algorithm

Figure 6. Class separation by QUEST algorithm
As previously indicated, the role of the Yellow Band is obviously important in this case. This is reflected by primary separation node on both algorithms which use Yellow Band as the first discriminator. Separation is further made by inclusion of Red Edge Band mainly to assist differentiation of dry conditions. It is shown that addition of Coastal, Yellow and Red Edge Bands in WorldView-2 data delivers enhancement to previous understanding on the problem and therefore opens a better perspective in mapping purpose.

Quantitative assessment on mapping was made through Kappa coefficient. Tables 1 and 2 respectively present accuracies based on decision tree classifiers. The Kappa coefficients indicate that QUEST was slightly robust than CRUISE. This is fairly consistent with our previous finding using ALOS AVNIR-2 and Landsat TM/ETM (Panuju et al. 2007; Trisasongko et al. 2010).

Probable cause of slightly moderate accuracies was textural information which is prominent in high resolution imageries. Despite slight inaccuracies, we found that addition of Yellow, Red Edge and Coastal Bands conveys a positive role in rice monitoring.

World-View 2 delivers a new insight on rice monitoring problems in Indonesia. The analysis indicated that the inclusion of Yellow, Coastal and Red Edge Bands is substantial to discrimination of diverse growth stages. In particular, Coastal band aided separation of waterlogged stage, which indicates the beginning of a new planting season. Red Edge is particularly important for dry fallow discrimination.

Using CRUISE and QUEST decision tree algorithms, we found that sufficient thematic map could be produced. QUEST delivered a slightly better accuracy than of CRUISE, which is consistent to our previous experiments. Accuracy could be improved by insertion of textural information. Implementation solely based on tonal data is therefore discouraged.

The WorldView-2 dataset was obtained from Digital Globe through the 8-band Research Challenge. We would like to express our gratitude to Mr. Ian Gilbert of Digital Globe for his assistance and helpful initiatives. Field work was partially supported by KKP3T and ALOS PP-2 Projects.