AGRICULTURE LAND USE MAPPING USING MULTI-SENSOR AND MULTI- TEMPORAL EARTH OBSERVATION DATA INTRODUCTION

Transcription

1 AGRICULTURE LAND USE MAPPING USING MULTI-SENSOR AND MULTI- TEMPORAL EARTH OBSERVATION DATA Jiali Shang Catherine Champagne Heather McNairn Agriculture and Agri-Food Canada 960 Carling Avenue, Ottawa, ON, Canada K1A 0C6 ABSTRACT Large area mapping of crop information is a crucial source of information on agricultural land use, and can be done most efficiently using Earth observation (EO) technology. Agriculture and Agri-Food Canada (AAFC) is developing a crop inventory system based on EO data. Detailed crop type identification relies on image data acquired during key crop phenological stages in order to capture the unique temporal signature of individual crop types. Optical data such as Landsat and SPOT can provide valuable information for crop classification, however, cloud cover is a reoccurring obstacle to the application of optical data, hindering mapping and monitoring at regional and national scales. Consequently, the integration of radar with optical EO data is essential for operational monitoring over many areas. Crop information mapping from EO data was conducted by AAFC for two consecutive years (2004 and 2005) over an Eastern Ontario pilot site. Landsat, SPOT, RADARSAT (standard mode), and Envisat ASAR (VV, VH) data were acquired during the growing seasons for both years. This site consists primarily of corn, soybean, cereal, production, and animal pasture. Three classification techniques (Maximum- Likelihood Classifier, Decision-Tree Classifier, and Neural Network Classifier) were applied to various image combinations. Overall, these results show that data acquired later in the growing season provide a better classification accuracy for both optical and radar data. The integration of radar and optical data resulted in a synergistic effect, producing an increase in classification accuracies (individual class accuracy, overall accuracy, and Kappa coefficient). INTRODUCTION The collection of spatial information on land use is one of many applications that can benefit from Earth observation (EO) technology. Agriculture and Agri-Food Canada (AAFC) is developing a crop inventory system based on EO data that will characterize the location, extent and characteristics of agricultural land use across Canada. Due to the magnitude of conducting an annual land use inventory at a national scale, an operational approach using EO data is needed. Detailed crop type identification relies on image data acquired during key crop phenological stages in order to capture the unique signature of individual crop types over time. The main objectives of this project are to identify the ideal timing of EO data acquisition, the optimal EO data inputs, and the most robust and operationally suitable classification methods for crop-type mapping. To date, five pilot sites have been established by AAFC over which methods for crop classification will be developed and tested. These sites cover a wide range of climatic zones with different agriculture practices across Canada, including Eastern Ontario, Swift Current (Saskatchewan), Lethbridge (Alberta), Prince Edward Island, and the Red River Valley (Manitoba). Extensive testing is needed over multiple sites and for multiple years prior to drawing final conclusions with respect to the methodology and classification accuracies. This study reports on the results derived from the Eastern Ontario site over two growing seasons (2004 and 2005).

2 STUDY SITE The study site is centred around the town of Casselman, about 50 km by 50 km in area. It belongs to the South Nation sub-watershed, about 50 km east of Ottawa, Ontario, Canada. This site is consists primarily of agricultural land with corn, soybean, cereal, and production. Ontario Study Site Figure 1. Map of Canada showing location of the Casselman study site. DATA COLLECTION & PRE-PROCESSING Earth Observation Data Identifying the timing of satellite acquisitions, as well as the frequency of acquisitions during the growing season, are important components of this project. Data were acquired from two optical sensors (Landsat-5 and SPOT-4) as well as two SAR sensors (RADARSAT-1 and Envisat-ASAR) throughout the growing season to examine the ideal timing and data inputs required for crop-type mapping. Details on the optical data used in this study are listed in Tables 1 & 2. Details on the radar data used in this study are shown in Tables 3 & 4. Table 1. Optical data acquisitions over the Casselman study area for the 2004 growing season. Date Sensor Spatial Resolution 10 June SPOT-4 ~ 20 m 17 July SPOT-4 ~ 20 m 17 July Landsat-5 ~ 30 m 22 August SPOT-4 ~ 20 m Table 2. Optical data acquisitions over the Casselman study area for the 2005 growing season. Date Sensor Spatial Resolution 2 June Landsat-5 ~ 30 m 15 July SPOT-4 ~ 20 m 21 August Landsat-5 ~ 30 m 6 September Landsat-5 ~ 30 m 10 September SPOT-4 ~ 20 m

3 In 2004, four Envisat ASAR images and four RADARSAT images were acquired over the study site. For 2005, six Envisat ASAR images were acquired whereas sixteen RADARSAT images were programmed. However due to planning conflicts with other data users, 12 RADARSAT images were successfully acquired. Details on the radar data used in this study are shown in Tables 3 & 4. Table 3. Radar acquisitions over the Casselman study area for the 2004 growing season. Sensor Date Beam Mode Polarization Spatial Resolution RADARSAT-1 17 June S7 HH ~ 30 m RADARSAT-1 1 July S5 HH ~ 30 m RADARSAT-1 25 July S5 HH ~ 30 m RADARSAT-1 26 August S6 HH ~ 30 m ASAR 13 June IS6 VV & VH ~ 30 m ASAR 2 July IS7 VV & VH ~ 30 m ASAR 18 July IS6 VV & VH ~ 30 m ASAR 22 August IS6 VV & VH ~ 30 m Table 4. Radar acquisitions over the Casselman study area for the 2005 growing season. Sensor Date Beam Mode Polarization Spatial Resolution RADARSAT-1 2 June F1F HH ~ 10 m RADARSAT-1 19 June F3 HH ~ 10 m RADARSAT-1 6 July F5F HH ~ 10 m RADARSAT-1 13 July F3 HH ~ 10 m RADARSAT-1 20 July F1F HH ~ 10 m RADARSAT-1 30 July F5F HH ~ 10 m RADARSAT-1 6 August F3 HH ~ 10 m RADARSAT-1 13 August F1F HH ~ 10 m RADARSAT-1 23 August F5F HH ~ 10 m RADARSAT-1 30 August F3 HH ~ 10 m RADARSAT-1 6 September F1F HH ~ 10 m RADARSAT-1 16 September F5F HH ~ 10 m ASAR 27 June IS3 VV & VH ~ 30 m ASAR 13 July IS2 VV & VH ~ 30 m ASAR 1 August IS3 VV & VH ~ 30 m ASAR 17 August IS2 VV & VH ~ 30 m ASAR 5 September IS3 VV & VH ~ 30 m ASAR 21 September IS2 VV & VH ~ 30 m

4 Ground Data Field observations were gathered during the summers of 2004 and Fields were visited 3 times in mid June, late July, and late August. The crop type and crop growth stage were noted for each field. All field data were collected using ArcPad and were collated in an ArcGIS geo-database. A total of 459 fields were surveyed in 2004, and 397 fields were surveyed for Data Preprocessing Otho-correction: For the purposes of integrating the various data sources, and in order to facilitate comparisons with ground data, all of the Landsat, SPOT, RADARSAT-1 and ASAR data were ortho-rectified. The orthorectification process used a 1:10,000 DEM and GCPs collected from the National Road Network vectors. A nearest neighbor re-sampling method was applied on the optical and radar data with a 20 m output resolution. Atmospheric correction: Atmospheric correction was applied to all optical data to retrieve the at-surface reflectance using the Atcor algorithm in Geomatica of PCI Geomatics (Richter, 2004). The Atcor algorithm uses the MODTRAN 4.2 radiative-transfer code for the radiance to reflectance conversion (Champagne et al., 2005). Filtering: The RADARSAT and ASAR data were received absolutely calibrated. Prior to image analysis, a Gamma filter was applied to suppress the speckle inherent in all SAR data. Considering some of the fields under study are narrow or small, a small filter window of 3 by 3 was adopted to minimize contamination from field boundaries and adjacent fields. After applying the Gamma filter, the speckle effects were still visible. Consequently, the Gamma filter was applied a second time, using the same window size (3 by 3). With two passes of the filter, noise in the SAR imagery was significantly reduced. Data Fusion Due to differences in the swaths, the images collected from the various satellite platforms or from different beam modes from the same satellite were not entirely geo-spatially coincident. A common window was selected to maximize the spatial coverage of the overlapping image acquisitions. If a particular pixel had no value in one image, then that pixel was masked out of all images and was excluded from further analysis. Only the pixels within the overlapping area were classified. Compared with the 2004 data, the data used in 2005 classification contained more useable pixels due to reduced cloud cover in the input imagery, and improved spatial targeting of planned imagery. METHODOLOGY Classification Methods When adequate ground truth data are available, supervised classification approaches generally produce better results relative to unsupervised classifications (Schowengerdt, 1997). For both 2004 and 2005, three supervised classification techniques were applied to the same dataset in order to identify an optimal classifier for land use mapping. These approaches included a Maximum-Likelihood Classifier (MLC), a Decision-Tree (DT) approach, and a Neural Network Classifier (NNC). All the classifications were performed on a per-pixel basis without a null class. Details on these models can be found in McNairn et al., Training and Testing Site Selection The same set of training and testing samples were used for all three classifiers. To reduce bias, the training pixels and testing pixels were selected from different fields. For each crop type, half of the fields were selected for training the classifier and the second half were reserved for testing. Consequently, there was no overlap between the training and testing pixels. In the 2004 analysis, the training and testing fields were selected manually. For each crop type, the training sites were selected in such a way that every spatially alternate field was chosen (Figure 2). The rest of the fields were used as testing sites.

5 Field boundary for corn Training site for corn Testing site for corn Figure 2. An example showing the spatial arrangement of training and testing sites used for 2004 classification. For the 2005 classification, the training and testing sites were selected randomly in ArcGIS (Figure 3). First, a 30 m buffer was applied to each field boundary. The pixels within this boundary buffer zone were excluded from training or testing in order to reduce contamination from headlands and mixed pixels. Given all the fields surveyed, a 50% breakdown was provided to the system which resulted in about half of the fields being the training sites and half of the fields being the testing sites. Again, there was no overlap between training and testing fields. Field boundary for corn Training site for corn Testing site for corn Figure 3. An example showing the spatial arrangement of training and testing sites used for 2005 classification. For the 2004 growing season, seven crop types were identified from the field survey. These crops included barley, oats, wheat, corn, soybean, potato, and pasture. Barley represented 1.8% of the total fields visited and oats for 1.6%; therefore barley and oats were not included in the classification. For the 2005 growing season, eight crop types were identified from the field survey. These crops included barley, buckwheat, oats, wheat, corn, soybean, potato, and pasture. Corn, soybean and pasture/ are the major types, and represent 333 of the 397 fields surveyed, covering approximately 87% of the surveyed agricultural area. There were two buckwheat fields, and only one was within the data image area, therefore buckwheat was not included in the classification. Classification There were two steps in the classification process. Initially all data combinations were classified into seven classes, with each class representing a single crop type. Due to the amount of confusion in classification among the cereal crops, barley, oats and wheat were grouped together to make a cereal class after running the classification. The grouping of the cereal classes was also tested prior to the classification; however, the accuracy levels were not as high as merging the cereals after the classification (likely as a result of differences in the spectral characteristics

6 of each crop). Comparisons among the classification results were accomplished using producer s accuracy (including individual crop classification and overall classification) and the Kappa coefficient. Individual crop classification accuracies as well as overall accuracies in the tables are presented as percentages of correctly classified pixels. RESULTS AND DISCUSSION The classification results for the 2005 data set are discussed in detail, followed by a comparison of the two data sets. More detailed results on the 2004 classification can be found elsewhere (McNairn et al., 2005; Champagne et al., 2005). Timing of Acquisitions The classification accuracies for single-date imagery were compared for optical and radar data. Classifications were run on individual Landsat, SPOT (Table 5) and ASAR (Table 6) images to identify which acquisitions during the growing season resulted in the highest accuracies. Since each RADARSAT-1 scene contains only one channel (HH polarization), the likelihood of achieving meaningful classifications with a single channel is low. To circumvent this limitation, RADARSAT-1 data were classified in pairs by combining the two adjacent dates (Table 7). Table 5. MLC results using single optical images, summer ID Data Used Cereal Corn Pasture/ Potato Soybean Overall Kappa 1 Landsat 2 June SPOT 15 July Landsat 21 Aug Landsat 6 Sept SPOT 10 Sept For the optical data, single-date classifications using data acquired later in the growing season provided the best overall classification accuracy and Kappa coefficient. Of the three Landsat images, the August 21 st image gave the highest overall accuracy and Kappa coefficient. Classification using the early June Landsat acquisition provided the poorest results. This confirms the conclusion drawn from the 2004 classification, and is consistent with observations made elsewhere (Allen et al., 2005). Table 6. MLC results using individual ASAR images. ID Data Used Cereal Corn Pasture/ Potato Soybean Overall Kappa 1 27 June July August August September September As with the single date optical imagery, one ASAR image or two Radarsat images did not provide enough information to adequately classify crop types. Individual crop classifications, as well as overall classification accuracies for the single date radar classification fell well below those reported using the optical data. The only exception is that the September 5 th ASAR produced slightly higher overall accuracy (58.5% vs 56.3%) and Kappa (0.43 vs 0.40) compared to the June 2 nd Landsat. The late season radar imagery also proved to produce higher classification accuracies. Among the six ASAR images, the early September image produced the highest overall accuracy and Kappa coefficient. Similarly, the

7 September 6 th and 16 th RADARSAT image pair gave the highest overall accuracy and Kappa coefficient. From the beginning of June to early September, there is a general increase in overall accuracy and Kappa coefficient. This is mainly driven by the increased classification accuracies of the three main land use classes, corn, pasture/, and soybean. However data acquired too late in the season will cause the overall classification accuracy to drop. As shown in Table 6, the September 5 th data gave an overall accuracy of 58.5% and a Kappa coefficient of 0.43 whereas the September 21 st data produced an overall accuracy of 43.2% and a Kappa coefficient of Table 7. MLC results using RADARSAT image pairs from two adjacent dates. ID Data Used Cereal Corn Pasture/ Potato Soybean Overall Kappa 1 2 June, 19 June July, 13 July July, 30 July August, 13 August August, 30 August September, 16 September The importance of late season SAR acquisitions for crop classification in western Canada was also reported by McNairn et al. (2002). With relatively low vegetative cover early in the growing season, a significant contribution to radar backscatter will originate from the soil. Between late July and late August, crops have completed their vegetative growth period and are in the reproductive and seed development stages. During this period, canopy moisture content and canopy structures change dramatically. These changes can be exploited to enhance crop separability. Comparison of Optical and Radar Data Inputs Multi-temporal Landsat and SPOT provided better classification accuracies (Table 8) than single-date images. The addition of the July 15 th SPOT to the two Landsat images (August 21 st and September 6 th ) improved the classification accuracy compared with using the two Landsat images alone (#4 vs #5). However by adding an additional June 2 nd Landsat, the classification accuracies stay almost exactly the same as using one SPOT and two Landsat images, indicating that this image does not add any critical information to the classifier (#5 vs #6). The results also suggest that there is a saturation point when more optical images are incorporated the classification and overall accuracy no longer improves. For example, when all optical data were used (3 Landsat plus 2 SPOT), the overall accuracy and Kappa coefficient decreased compared to the results using just the 3-date Landsat (#3 vs #7). This is mainly caused by the soybean class: its accuracy decreased from 92.1% to 79.4%. The potato class, however, experienced a large increase in accuracy, from 56.0% to 84.6%. This could be due to the contribution of the mid July SPOT. The potatoes in this site were harvested around late July and early August. The cereal, corn and pasture class accuracies stay relatively constant. Table 8. MLC results using multi-temporal optical images. ID Data Used Cereal Corn Pasture/ Potato Soybean Overall Kappa 1 SPOT 15 July, Landsat 21 Aug All SPOT (2 dates) All Landsat (3 dates) Landsat 21 Aug, 6 Sept SPOT 15 July, Landsat 21 Aug, 6 Sept Landsat, 15 July SPOT All opticals (5 dates)

8 When more dates of radar datasets were used, classification accuracies increased steadily compared with when fewer dates of radar data were used (Table 9: #3 to #8). With 12 RADARSAT images, an overall accuracy of 61.4% was achieved. The relatively low accuracy from using temporally dense RADARSAT coverage indicates that using HH single-polarization radar alone will not likely be a viable method for classifying crop types. A multidate ASAR dataset outperformed a multi-date RADARSAT dataset (Table 9: #2 vs #8). With six ASAR images, overall classification accuracy reached 72.4% compared to an overall accuracy of 61.4% achieved with 12 RADARSAT images. These results reflect the richer information that accompanies multi-polarization data. In summary, satisfactory classification results (greater than 85% overall accuracy) were not achieved using radar data alone, even when multi-temporal data were introduced to the classifier. The best classification results were achieved with 18 dates of radar data (75.4% overall accuracy and 0.66 Kappa coefficient). When the number of images used reaches a certain level, the contribution of additional images will become smaller. For example, when using six ASARs, an overall accuracy of 72.4% and a Kappa coefficient of 0.62 were achieved (Table 9: #2). When adding 12 more RADARSAT images, the overall accuracy increased from 72.4% to only 75.4%, and the Kappa coefficient increased from 0.62 to 0.66 (#9). Table 9. MLC results using multi-temporal radar data. ID Data Used Cereal Corn Pasture/ Potato Soybean Overall Kappa 1 4 ASAR (27 June, 13 July, 1 & 17 Aug) All ASAR (6 dates) RADARSAT (6 & 16 Sept) RADARSAT (19 June, 6 & 30 July, 23 Aug) RADARSAT (19 June, 6 & 20 July, 13 & 23 Aug, 6 Sept) RADARSAT (19 June, 6, 13 & 20 July, 13, 23 & 30 Aug, 6 Sept) RADARSAT (19 June, 6, 13, 20 & 30 July, 6, 13, 23 & 30 Aug, 6 Sept) All RADARSAT (12 dates) All ASAR, all RADARSAT (18 dates) The synergy of using optical and SAR data in combination was also tested (Table 10). SAR sensors respond to large-scale crop structure (size, shape and orientation of leaves, stalks, and fruits) and the dielectric properties of the crop canopy. Crop structure and plant water content vary as a function of crop type, growth stage and crop condition. Optical sensors respond primarily to plant biochemical properties, as well as structural characteristics such as leaf area index. For this dataset, the best combinations of single-date ASAR and single-date optical happened between mid August and early September pair (#4 and #8). The images in these pairs also provided the highest accuracy when used individually. When the August 21 st Landsat and the August 17 th ASAR images were used, the classification achieved an overall accuracy of 82.7% and a Kappa of The second best pair included the August 21 st Landsat and September 5 th ASAR, resulting in an 81.6% overall accuracy and a 0.75 Kappa. When the August 17 th ASAR image was replaced with the August 30 th RADARSAT image, the overall classification accuracy was only slightly reduced (from 82.7% to 80.8%). This result was achieved despite the limitation of RADARSAT data having only one polarization.

9 Table 10. MLC results using combinations of optical and radar images, summer ID Data Used Cereal Corn Pasture/ Potato Soybean Overall Kappa 1 SPOT 15 July, ASAR 13 July SPOT 15 July, RADARSAT 13 July SPOT 15 July, ASAR 13 July, RADARSAT 13 July Spot 15 July, 6 ASAR Landsat 21 Aug, ASAR 17 Aug Landsat 21 Aug, RADARSAT 23 Aug Landsat 21 Aug, RADARSAT 30 Aug Landsat 21 Aug, ASAR 17 Aug, RADARSAT 30 Aug Aug. Landsat, 6 ASAR Landsat 21 August, ASAR 5 Sept SPOT 10 September, ASAR 5 Sept SPOT 10 September, ASAR 5 Sept, RADARSAT 6 Sept All Landsat (3 images) and all ASAR (6 images) All optical and radar (23 images) For all of the data combinations tested, the highest overall accuracy was produced by using 3 dates of Landsat, with an accuracy of 86.2%. Accuracies comparable to this, however, were achieved when only a single optical image (August 21 st Landsat) is used in combination with a single ASAR image (August 17 th ), with an overall accuracy of 82.7%. In addition, this Landsat/ASAR combination also showed a higher individual accuracy for potato crops (from 56.8% for the 3 Landsat to 82.9% for the single date Landsat/ASAR pair). This is consistent with the results from the 2004 classification. This offers some flexibility in the acquisition of optical data, where cloud conditions limit the availability of such data. When multi-date radar data are combined with single-date optical data, the synergistic effects are even more pronounced. For example, when only the July 15 th SPOT was used, an overall accuracy of 70.8% and a Kappa of 0.61 were achieved (Table 5). By adding one ASAR (July 13 th ), the overall accuracy increased to 77.5% (a relative increase of 9.5%) and the Kappa increased to 0.70 (a relative increase of 14.8%). Compared with using July 15 th SPOT alone, when 6 ASAR were incorporated, the overall accuracy increased from 70.8% to 83.8% (a relative increase of 18.4%), the Kappa increased from 0.61 to 0.78 (a relative increase of 27.9%). A similar trend was found using Landsat. When only the August 21 st Landsat was used, an overall accuracy of 74.9% and a Kappa of 0.65 were achieved. By adding one ASAR (August 17 th ), the overall accuracy increased from 74.9% to 82.7% (a relative increase of 10.4%), and the Kappa increased from 0.65 to 0.76 (a relative increase of 16.9%). Compared with using August 21 st Landsat alone, when 6 ASAR were incorporated, the overall accuracy increased from 74.9% to 85.4% (a relative increase of 14.1%), and the Kappa increased from 0.65 to 0.80 (a relative increase of 23.1%). When multi-date optical data are available, the addition of radar data does not seem to further improve the classification accuracies. For example, using all three Landsat images, an overall accuracy of 86.2% and a Kappa of 0.81 were achieved. By adding 6 ASAR images, there was little improvement in the classification accuracy. When combining all radar data (18 images) with all available optical data (5 images), the classification accuracies stayed constant. This suggests that while multi-date optical data are available, they provide the most information for crop inventory mapping.

10 Comparisons among Different Classification Methods Classifications were run on selected combinations of optical and radar data using all three classifiers (MLC, DT and NNC). Results are shown in Table 11. Among the 8 combinations tested, NNC constantly produced lowest classification accuracies, except for when a large number of images were used (#3, #6, #8). MLC and DT generally gave comparable results. This is comparable to results obtained over the Swift Current test site that showed that the MLC and DT models gave consistently better results over the NNC model (Champagne et al., 2006). The MLC is computationally simple; which offers advantages for operational mapping. The DT classifier, however, allows the integration of other geospatial data which may assist with the separation of some crops during future testing. Table 11. Classification accuracies derived from MLC, DT and NNC classifiers. Sensors ID Images Used Method Overall Kappa Optical Radar Optical & Radar 1 TM 21 August 2 SPOT 15 July 3 All optical (5 images) 4 All ASAR (6 images) 5 All RADARSAT (12 images) 6 All ASAR & RADARSAT (18 images) 7 TM 21 August, ASAR 17 August 8 All optical, all radar (total 23 images) MLC DT NNC MLC DT NNC MLC DT NNC MLC DT NNC MLC DT NNC MLC DT NNC MLC DT NNC MLC DT NNC Impact of Training/Testing Site Selection on Classification Accuracy It is known that classification performance is largely dependent on the quality of the training data. The training data should be representative, have sufficient number of pixels, and provide a good spatial distribution. To test the adequacy of the training and testing data used in this study, additional classifications were run by swapping the training and testing data (i.e. the original testing data were used for training and the original training data were used for testing). The results are given in Table 12.

11 Table 12. Comparisons of classification accuracies (MLC) derived with and without swapping the training and testing data, summer ID Data used Train/Test Cereal Corn Pasture Potato Soybean Overall Kappa original All TM (3 images) switched average original All ASAR (6 images) switched average original switched ASAR 17 August + TM 21 August All ASAR + All TM (total 9 images) average original switched average In general, for the image combinations tested there were changes in the overall classification accuracies (ranging from 0.5% to 4.6%) and Kappa coefficients (ranging from 0.01 to 0.06). For individual classes, small variations appear in crop types with large acreage, such as corn and soybean. For less abundant crops, such as cereal and potato, the accuracy differences are more pronounced, where the highest variation in the potato class reached 19.6% (#1). This was likely a reflection of the limited number of training pixels available and the spatial distribution of these training samples. The impact of training and testing data on classification accuracy will be further explored in future work. Comparison between 2004 and 2005 Results In general, classification over the two growing seasons demonstrated some similar trends and patterns. Variation does exist, however, between the two years in terms of level of accuracies. Table 13 shows classification results of selected image combinations from both years. The accuracies shown are calculated by taking the average of the accuracies derived from classifications before and after swapping the training and testing sites. Table 13. Comparison of classification results (MLC) from 2004 and 2005 growing seasons. ID Data Used Year Cereal Corn Pasture/ Potato Soybean Overall Kappa 1 Landsat 15 July, SPOT 10 June, 22 Aug Landsat 2 June, 21 Aug., 6 Sept SPOT 22 Aug, ASAR 22 Aug Landsat 21 Aug, ASAR 17 Aug Due to limitation of data availability, no direct comparison between 2004 and 2005 could be made. For both data input combinations selected, the 2004 classifications produced higher accuracies. Two factors might have contributed to this: the selection of training and testing data, and the differences in the way cereal crops were treated between the two seasons. For the 2004 season, training/testing sites were selected manually with the training/testing pixels far away from the field boundaries. For 2005 the sites were selected automatically with a smaller buffer from the edge of the field. This resulted in more spectrally uniform areas within each field being selected in 2004, perhaps reducing the variability in the training data statistics. Another factor could be that in the 2004 classification, for the only cereal crop classified was wheat, and barley and oats were excluded from classification due to their small coverage. For 2005, all three cereal crops were included in the classification and then merged as one cereal class. This might have affected the overall accuracies by further increasing the variability in the training data statistics. Additional testing will be conducted to further explore the robustness of these methods. EO data have been acquired over the same study site for the 2006 growing season, which will provide a great opportunity for further comparison.

12 SUMMARY Based on the results of this study, the following conclusions have been reached: Data acquired later in the growing season (late August to early September) are critical for crop separation. Early season data are less useful. This observation holds true for both optical and radar data. Multi-temporal optical data are ideal for crop classification. Alternating polarization (ASAR) radar imagery provides better classification results when compared to single polarization (RADARSAT) imagery. By combining ASAR and RADARSAT (VV, VH, and HH), the classification accuracies are further improved. Radar can support crop classification when optical data are not available. The integration of a single radar image and a single optical image provide an alternative when multitemporal cloud-free optical images are not available. However the timing of data acquisition is critical. Overall, MLC and DT provide comparable results. Since NNC does not give better performance and is computationally less efficient, it is not recommended for operational land use mapping. The selection of adequate and representative training sites is critical for successful crop classification. ACKNOWLEDGEMENTS Funding for this research was provided by the Canadian Space Agency under GRIP (Government Related Initiatives Program). Sincere thank goes to Patrick Assouad, Leander Campbell, Rene Chenier, Bahram Daneshfar, Thierry Fisette, Pierre Yves Gasser, Eric Gauthier, Greg Gibbons, and Anna Pacheco for their contribution through the course of this study. REFERENCES Allen, R.; Hanuschak, G. and Mike Craig, M. (2005). History of Remote Sensing for Crop Acreage in USDA's National Agricultural Statistics Service, 2002, (URL: Campbell, J.B. (1997). Introduction to Remote Sensing (2 nd Edition). The Guilford Press, New York, USA, 622p. Champagne, C., Shang, J., and McNairn, H. (2005). Exploiting spectral variation from crop phenology for agricultural land-use classification. In Proceedings of SPIE Optics and Photonics 2005, Volume 5884, San Diego, CA, USA, July 29 August 4, (ditial proceedings). Champagne, C., H. McNairn and J. Shang (2006). An Object-Oriented Approach to Land Use Mapping in the Canadian Prairies. Eleventh Biennial USDA Forest Service Remote Sensing Applications Conference. April 24-28, Salt Lake City, Utah, USA. McNairn, H., Shang, J., and Champagne, C. (2005). Report on Land Use Mapping for the Eastern Ontario Pilot Site 2004 Growing Season, IN , Agriculture and Agri-Food Canada GRIP Projects annual report to the Canadian Space Agency on Government Related Initiatives Program (GRIP), 33 pages. McNairn, H; Ellis, J; van der Sanden, J J; Hirose, T, and Brown, R J (2002). Providing crop information using RADARSAT-1 and satellite optical imagery. International Journal of Remote Sensing, Vol. 23, 5, pp Richter, R. (2004). Atmospheric/topographic correction for satellite imagery: Atcor 2/3 users guide version 6.0., DLR German Aerospace Centre: Wessling, Germany, 71 pages. Schowengerdt, R.A. (1997). Remote Sensing: Models and Methods for Image Processing (2 nd Edition). Academic Press, San Diego, California, USA, 467 pages.