International Archives of Photograminetry and Remote Sensing. Vol. XXXII Part 7C2, UNISPACE III, Vienna, 1999 
46 
/•SKn 
I5FR5 
UNISPACE III - ISPRS Workshop on 
“Resource Mapping from Space” 
9:00 am -12:00 pm, 22 July 1999, VIC Room B 
Vienna, Austria 
ISPRS 
The performance of this approach in crop area assessment 
proved to meet the strict requirements (Figs, l.a.b.) both for the 
validation period (1991-96) and in the operational one. These 
two comprises an 7 years, 33 cases (county/year) sample. The 
strong relationship in the Landsat TM derived (FOMI RSC) and 
Central Statistical Office, Hungary (CSOH) data for the major 
crops proved, that this method was independent from the given 
year or the area, the different terrain and complexity of the 
counties. 
2.2. Crop monitoring and yield estimation methods 
The novel result of HARSP are the purely remote sensing crop 
monitoring and yield forecast models. The models were 
developed by FOMI RSC. They integrated NOAA AVHRR and 
high-resolution satellite data (e.g. Landsat, IRS-1C/D, SPOT). 
The models combine the benefits of both data sources: the 
frequency of NOAA AVHRR data and spatial resolution of 
high resolution images. This approach requires fairly accurate 
crop maps. With the adaptation and improvement of a linear 
unmixing model (Puyou Lascassies et. al., 1994) to NOAA 
AVHRR series a crop development assessment and quantitative 
yield forecast model was developed. The model was calibrated 
at the spatial units level of 400-500 ha. That is the guarantee for 
its good performance at the counties level (approx. 0,5 million 
hectare each, in Hungary) and further. That is also why it can 
produce a crop yield distribution map. The county wheat and 
maize yields predicted by the model compared favourably to the 
official data (Figs. 2.a.b.) both in the pre-validation period 
(1991-96) and in the operational one (1997-99) as well. The 
structure of the model is similar for different crops and it does 
not depend on the area and the given year’s whether. It was also 
found that the timeliness requirement can be met by the yield 
forecast model. 
3. OPERATIONAL CROP AREA ASSESSMENT 
AND YIELD FORECAST FROM 1997-99 
The substantial R+D and validation created a firm basis to move 
forward to an operational program: Crop Monitoring and 
Production Forecast Program (CROPMON 1997-1999). The 
crop data-reporting calendar was set by the customer, the 
Ministry of Agriculture and Regional Development. 
It consists of five dates from June 15 to October 1 in harmony 
with the existing traditional production forecast system of 
MoARD. The area covered directly have been a characteristic 
subsample (6-9) of all the counties (19), so that 40-57 % of the 
total cropland in Hungary have directly been monitored. 
Beyond the crop area assessment and yield prediction for the 
counties, these data are expanded to the entire area of Hungary. 
This expansion uses a subregional temporal correlation analysis 
plus a direct robust method (see 4.). The eight main crops 
monitored are winter wheat, winter and spring barley, maize, 
sugar beet, sunflower, alfalfa and maize to ensilage. These crops 
together represent the 78-82 % of the entire Hungarian 
cropland. 
The crops area assessment is based on the quantitative analysis 
of multitemporal high resolution images (Landsat TM and IRS- 
1C/1D LISS HI.) from early May through August, to 
compensate for the cloudiness. The comparison of the remote 
sensing results with CSOH data is obviously an indication only. 
The differences cannot be interpreted, by any means, as errors 
of the remote sensing technology. A thorough study is under 
way that will produce confidence values attached to the area 
estimates. The difference of crop areas estimates of FOMI RSC 
and the Central Statistical Office, Hungary (CSOH) is in the 
range of 0.8-3.7 % (Fig.5.a.) for the entire cropland in Hungary. 
The county crop area differences occurred in the interval of 1.5- 
21 % depending on the crop and county. However the area 
weighted average difference was 4.08 %. 
This partially can be explained by the main differences in 
definitions, that is the ownership based sampling of CSOH and 
the administrative, topographic boundary based total coverage 
of cropland by the satellite images (FOMI RSC). The actual 
standard crop maps derived were also provided to MoARD 
(Fig.3.). 
The crop yield forecast was accomplished by the application of 
FOMI RSC developed model which combines high-resolution 
satellite (Landsat TM and IRS-1C/1D LISS III. or SPOT) data 
and NOAA AVHRR time series. The reporting dates 
corresponded to those of the operative Production Forecast 
System of the Ministry of Agriculture and Regional 
Development. Both appeared prior to the beginning of harvest. 
The final official data are available after the harvest: by the end 
of August for wheat and barley and in December (January) for 
the rest. FOMI RSC provided yield estimates for the counties 
and expanded them to Hungary using a regional-historical 
correlation scheme. The country average yield data compare 
favourably with CSOH preliminary values, that appear six 
weeks later (Fig.5.b.). The differences are less than 1 % for 
wheat and 4.5 % for maize average yields in Hungary. The 
differences at county level averages are somewhat bigger. 
Because of the method applied, yield spatial distribution maps 
could also be reported (Fig.4.) for the major crops. 
4. ROBUST YIELD PREDICTION BY 
NOAA AVHRR SERIES 
The primary yield forecast model (see 3.) performed well. There 
were two reasons to develop a robust yield forecast model: 
the need for a parallel, independent technique to control the 
primary yield forecast model extrapolation (see 3.) from 
the average yields of the directly monitored counties to the 
entire cropland in Hungary and 
the need for a stand-alone method that uses only very basic 
land use information: the boundary of cropland in the 
country. This was readily available from the CORINE 
Land Cover data base in Hungary (Biittner et. al., 1995). 
Beyond the land use NOAA AVHRR data series make up 
the basis of the model. The model is directly applied to all 
the individual counties and also for the national crop 
production forecast. 
The pre-processed and normalised NOAA AVHRR data set is 
temporally filtered. The average reflectance profile and the 
NDVI can be decomposed in time by a thorough spectral- 
temporal correlation analysis. This substantial analysis shows 
an extremely strong relationship between the predicted county 
yields by this decomposition method and the CSOH data (Figs.