3.2 Sujetos y fuentes de información
4.1.1 Análisis descriptivo de los datos
From the multitemporal SPOT 1998 and 2001 satellite imagery spectral measurements for band 1 (green), band 2 (red) and band 3 (near infrared) were retrieved for crop type barley, canola, chickpeas, lentils and wheat. It was observed in the typical spectral properties for each crop type that as the amount of green biomass on the ground increased, so did the absorption in band 2 (red wavelength) and the reflectance in band 3 (near infrared wavelength). During senescence the crops had reduced amounts of green biomass and hence the reflectance in band 3 decreased coupled with an increase in band 2. These observations agreed with reports in the literature (Badhwar and Henderson, 1981; Tucker and Sellers, 1986; Russ, 1993; Russ et al. 1993; for more details refer to Chapter 2). Chickpeas and lentils were found to be distinctively different to barley, wheat and canola. Together with their respective NDVI values all crop types could be visually distinguished.
Multiple factors influenced the crop development during a season. 1998 was a la Niña year (Wright, 2001); la Niña years are associated with higher spring rainfall and cooler daytime temperatures in the south east Australian region (Jones and Trewin, 2000) and hence an earlier season “break” (Liu et al,2004). Sufficient rain fall was a prerogative for successful crop emergence (Hammer, 1983 considered a rainfall of 20 mm in winter over 1 or 2 days to be the criterion for planting on cracking clay soil). The 1998 crop season started distinctly earlier than in 2001. When analysing the sowing dates of canola (supplied by the farmers) it was observed that the average sowing date in 1998 was early May, while in 2001 canola was not sown until end of June. This obviously had a significant impact on the satellite signals obtained in i.e. July, when in 1998 some crops had emerged while there was no vegetation signal in 2001 yet. Vegetation signatures were compared in Figure 5.46; the NDVI of canola fields was plotted in both years 1998 and 2001, respectively.
It was observed that the seasonal development was not linear; the speed of crop development was influenced by climatic events (such as temperature and rainfall) during the season.
Canola (NDVI) 0 20 40 60 80 100 180 210 240 270 300 330 DOY ND V I [* 1 0 0 ] 2001 Data 1998 Data
Figure 5.46: Temporal shift in the NDVI development of canola in 1998 and 2001 (see text for full explanation)
Seasonal effects are a significant challenge when transferring crop discrimination models without prior knowledge from one year to another. It is therefore necessary to build a database that includes many crop seasons to determine a “typical” crop season in south east Australia as a base line to adjust seasons to each other. Aigner (1999), and Aigner et al. (1999) reported building such a crop database with NOAA-AVHRR data (1995-1998) for the Gooroc test site when relating the satellite data to grain crop yields. In his study Aigner observed that the temporal behaviour of the NDVI varied with respect to season onset date and plateau duration. Li and Kafatos (2000) found the biosphere vegetation patterns in AHHRR data in the USA to be related to the El Niño/ La Niña effect. Reed et al (1994) related vegetation phenology to quantified AVHRR NDVI curve properties in the USA and Hill and Donald (2003) used such NDVI metrics in Western Australia to derive information about seasonal agricultural productivity. A regional multi-seasonal database utilizing NDVI metrics of remote sensing data with high temporal resolution (AVHRR or MODIS) together with climatic records needs to be built in future research to use seasonal information for the crop monitoring system in south east Australia.
Classification accuracies were obtained for each acquisition date with a discriminant function analysis. Using datasets from multiple dates, the classification accuracies could be improved significantly; in several models all fields were classified correctly. The data were compared to a similar dataset from 2001 and equivalent spectral
properties and classification results were found. However, the models are currently limited to the five investigated crop types. Applying the derived discriminant functions to fields with other crop use will result in misclassifications. Hence further research is required to derive spectral properties of other crop types and to include those into the discriminant functions.
Wilkinson (2005) analysed over 500 classification results reported in the literature from 138 separate papers over a time frame of 15 years (1989-2003). He found that the mean classification accuracy (overall per cent correct) was 76.19% with a standard deviation of 15.59%, and that reported classification results did not improve over the 15 year time frame. He noted that the number of features used in classification experiments (mean 7.85) was relatively low, given the potential value of multitemporal and multi-sensor mapping approaches and the apparent sophistication of classification approaches. The work presented in this thesis took advantage of a multitemporal dataset in the progressive date models and achieved results (up to 100%) which were superior to the average results found by Wilkinson (2005).
Information on crop type and status, together with area statements can also give valuable information to other service providers of the farming community, such as logistical planning in receiving docks, insurance companies, etc.
5.5 Conclusions
Typical crop signatures were derived from SPOT satellite data for barley, canola, chickpeas, lentils and wheat. The NDVI gave a good understanding of crop status during the season. Two years, 1998 and 2001 were observed. The crop spectral reflectance values showed similar behaviour in both years, however the temporal pattern was not consistent when comparing both years. The temporal crop development was compressed and stretched, subject to climatic conditions. Thus seasonal shifts complicate classification model transfer from one year to the next. When using the discriminant models derived in this study, climatic information and approximate sowing dates need to be integrated to address the seasonal shift aspects.
The accuracies of crop discrimination models were greatly enhanced by multitemporal satellite data as there was much information about crops in the temporal domain. Classification accuracies greater than 80% were obtained as early as the end of August for both investigated years (1998 and 2001). Knowledge of the crop spectral properties derived from this study coupled with in situ data of only a few selected “typical” crop fields should result in very good classification accuracies for the five investigated crop types in the future. It is anticipated that crop signatures in other south east Australian regions under similar cropping systems and soil types are comparable to the ones observed in the Gooroc area. However this will need to be confirmed.
The derived spectral properties of crops grown in south east Australian conditions comprise a valuable baseline data set for typical crop fields, allowing discrimination of atypical field; this information can be used to address atypical fields with precision farming management.