AN EVALUATION OF RESOURCESAT-1 LISS-III VERSUS AWIFS IMAGERY FOR IDENTIFYING CROPLANDS INTRODUCTION AND BACKGROUND

Transcription

1 AN EVALUATION OF RESOURCESAT-1 VERSUS AWIFS IMAGERY FOR IDENTIFYING CROPLANDS David M. Johnson, Geographer National Agricultural Statistics Service United States Department of Agriculture 3251 Old Lee Highway, suite 305 Fairfax, VA ABSTRACT The National Agricultural Statistics Service has been utilizing Resourcesat-1 (IRS-P6) AWiFS imagery for the past four growing seasons to derive crop specific classifications over intensive agricultural regions of the US. AWiFS, primarily because of its large 737 km swath width, has proven an ideal data source for such an application despite its relatively coarse 56 m ground sample resolution. Resourcesat-1 carries a second sensor,, which collects data in tandem and utilizes an identical 4-band multi-spectral reflectance detector. However, s camera s focal length provides a finer 23.5 m pixel resolution at the cost of a narrower swath width of 141 km. The purpose of this study was to quantitatively assess the impact that the different spatial resolutions have on land cover classification, especially targeted to agricultural cover types. To test the differences the USDA acquired clear sky, same path, same date imagery in parallel with that of AWiFS collected over agricultural areas of the central US. Imagery dates are from the mid-summer peak of most crops phenologic cycle. Ground truth information derived from the Farm Service Agency reporting data was used for training and validation of the classifications. Quantitative assessment results are presented along with subjective findings and show that classifications derived from data outperform the equivalent from AWiFS by several percent. The study further attempts to answer the scale question of what the ideal pixel size is for classifying agricultural cover types. INTRODUCTION AND BACKGROUND Since the early 1990s, the United States Department of Agriculture (USDA) / National Agricultural Statistics Service (NASS) has developed summer-based land cover classifications, with an emphasis on documenting row crop agriculture, over certain intensive growing regions of the United States (US) (Craig, 2001). The detailed crop specific classifications are termed by NASS the Cropland Data Layer (CDL). They have been typically undertaken on a statewide basis, particularly over those in the Midwest US, also commonly called the Corn Belt, and Mississippi River Alluvial Plain, also referred to as the Delta. The main impetus of the work has been to derive planted crop acreage estimates to improve or validate those collected from NASS traditional annual probability survey programs. Landsat-5 Thematic Mapper (TM) and Landsat-7 Enhanced Thematic Mapper Plus (ETM+) imagery have traditionally been used as the CDL input data source, but within the last couple of years a switch to Resourcesat-1 56 m resolution Advanced Wide Field Sensor (AWiFS) data was made by the agency. Regardless, TM, ETM+ and AWiFS imagery are all well suited for monitoring agriculture. NASS has leaned toward utilizing AWiFS operationally for cropland mapping for several reasons. They include the large 737 km swath width (often providing near statewide coverage in a single pass), inclusion of red, near-infrared (NIR), and shortwave-infrared (SWIR) spectral bands (all helpful for discriminating crop types), and the spatial resolution of 56 m (a tolerable size for mapping homogenous cover types such as agriculture fields which are usually large in area). Additionally, the satellite is operational in nature, healthy, and delivery of the imagery by the vendor is relatively fast. Agriculture and Agri-food Canada (AAFC) has also researched and tested the use of AWiFS for mapping croplands (Champagne et al., 2007) and found it to be practical. The sensor, also onboard Resourcesat-1, has not been utilized by NASS to date primarily because of cost and limited geographic coverage. However, does have a finer spatial resolution of about 23.5 m and utilizes an identical type of sensor to AWiFS which is appealing for cross validation of the coarser data.

2 Comparisons of AWiFS imagery to TM have been undertaken (Johnson, accepted; Boryan and Craig, 2005) and show TM to be slightly better for identifying crop types over the US because of a combination of TM s finer spatial resolution and increased number of spectral bands. Ultimately, TM is disadvantaged though because its revisit rate is about three times less than that of AWiFS, an important consideration for monitoring dynamic landscapes like agriculture. Additionally, high revisit rates are especially important in frequently cloudy areas like those surrounding the Great Lakes. No known applied use comparisons of versus other sensors have been published. Reflectance calibration comparisons (Chander and Scaramuzza, 2006) for all three sensors are available though. AWiFS and rely on identical charge- Table 1. AWiFS and sensor specifications. coupled device (CCD) reflectance sensors and system differences (Table 1) are mainly found in lens focal length (affecting ground sample size), digital quantization of the data (10-bit versus 7-bit), and the fact that two cameras (each with an approximately 370km wide field of view) are used to realize the entire AWiFS swath (which is 737 km) while (with a 141km swath) uses a more traditional single camera configuration (Himanshu et al., 2006; NRSA, 2003). The two systems ride in tandem on the Resourcesat-1 (IRS- P6) satellite platform which has a circular polar sun synchronous orbit above the earth at an altitude of 817 km with a 101 minute orbital period, not that unlike Landsat. However, the orbital path is repeated only once every 24 days, versus Landsat s 16, thus it takes that time period to reimage the same area. Since AWiFS has a much larger swath, resulting in large overlaps between adjacent scenes, the effective revisit period is a maximum of 5 days. Resourcesat-1 s sensors gather data at the very same instant over the nadir overlap area so that collection times are identical. Instantaneous Geometric Field of View Spectral bands AWiFS 56m (nadir) 70m (field edge) B2: B3: B4: B5: km each head 737 km (combined) Of note, the two AWiFS cameras overlap by approximately 7 km at nadir but since the radiometry between the two AWiFS cameras are identical, data from one side over the other is not preferred. Visual subset examples of AWiFS and data collected from the central North Dakota region during the peak of the growing season of 2006 are shown in Figure 1 (with a histogram stretch applied equally to each s lookup table). As one might expect given the sensor specifications, coincident AWiFS and data look spectrally very similar but spatially provides more detail. While larger cropland fields (shown by homogenous chunks of land) and water bodies (darkest blue areas) are distinguishable from within the AWiFS imagery example, the finer spatial resolution of helps clearly show more detailed features such as roadways (distinguished by linear features), small water bodies (small dark spots), and wetlands (darker areas usually near water). Even subtle variations and anomalies occurring within and across agricultural fields (e.g. varying degrees of the brighter green) are evident within the from which AWiFS cannot resolve. Histograms (not shown) between the sensors are very similar for all four bands. Swath 23.5 m B2: B3: B4: B5: km Integration time 9.96 msec 3.32 msec Quantization 10 bits Number of gains 1 7 bits (B5 band has 10-bit quantization, selected 7 bits out of 10 bits are be transmitted by the data handling system) 4 for B2, B3 and B4. For B5 dynamic range obtained by sliding 7 bits out of 10 bits

3 Figure 1. AWiFS (left) and (right) imagery examples from North Dakota (red, green, blue colors equal reflectance bands 5 (SWIR), 4(NIR), and 3(red), respectively). A second raw data example is given in Figure 2. The imagery for it were captured over southwestern Wisconsin and also during the peak of the summer. Water (darkest blues) and woodland (large tracts of medium green) areas are easy to identify with either camera but finer scaled features (like those that are man-made) are very difficult to separate from within the AWiFS imagery. Wisconsin has many areas where agricultural fields (depicted here by a range of green hues for vegetated fields and reddish-purple for barren ones) are smaller or less rectangular (especially from strip style contour farming) than that typically found throughout the rest of Midwest US. Thus, narrow or complex field plots are difficult to identify with coarser scaled imagery like AWiFS. Urban infrastructure, such as that along the left side of the subset images (reddish-purple to dark blues) shows up much more clearly with. Within the AWiFS scenes urban areas can easily be confused with barren agricultural lands (reddish-purple areas). A sense of the elevation and cropland management contours are present with that does not exist for AWiFS. Again though, spectrally the scenes look very similar in cast and contrast overall. Figure 2. AWiFS (left) and (right) imagery examples from Wisconsin (red, green, blue colors equal reflectance bands 5 (SWIR), 4(NIR), and 3(red), respectively).

4 METHODS AND RESULTS SUMMARY To test the usefulness of AWiFS versus for performing land cover classification, and particularly focused on crop cover types, independent classifications of the coincident Resourcesat-1imagery were performed. The study sites were chosen primarily based on the acquisition of cloud free imagery over highly intensive cultivated areas. These were also areas that NASS is highly familiar with because of past classification efforts within the CDL program. The first study site was over North Dakota (Figure 3) with data collected on August 22, The second site was over Wisconsin (Figure 4) and captured on July 31, Three image tiles were obtained for each study site (for a total of six) along with the in tandem collected overlapping AWiFS imagery (which turned out to be four AWiFS scene quads per site, thus eight in total). The imagery subset examples depicted in the Figures 1 and 2 came from the overall study sites and were meant to represent typical ground cover conditions found in each. Figure 3. Intersection area of (central image) and AWiFS (full swath image) over North Dakota, 8/22/06. The AWiFS data were provided as 8-bit from the vendor after being linearly rescaled from 10-bit. The data were provided as the native 7-bit and left as is. Preprocessing steps for each included first mosaicking and reprojecting, to a common Albers Conic Equal Area projection from the native Lambert Conic Conformal, same sensor path data together to form a single scene. Next the AWiFS data were clipped to the maximum extent of the scene so that only the identical areas of overlap were being analyzed and compared. Leica Geosystems ERDAS Imagine was used for all the image handling steps.

5 Figure 4. Intersection of (central image) and AWiFS (full swath image) over Wisconsin, 7/31/06. Ground truth information was obtained from the USDA / Farm Service Agency (FSA). Vector polygons from the FSA Common Land Unit (CLU) program were combined, utilizing ESRI ArcGIS software, with farmer reported land use form 578 data through a common ID linkage to build a database of land units where the cover type was known during the 2006 growing season. Because the CLU polygons often have more than one cover type within a unique polygon, only those records that matched one-to-one were used. Furthermore, the 578 reported acreage use was compared against that of the polygon area and if not within ten percent the polygon any associated information discarded. Figure 5 shows an example of the resulting polygons labeled by cover type. A unique numeric was assigned to each FSA cover type in order to build an output raster file consisting of digital numbers. Next, the polygons were then randomly divided into two groups, one for training and one for validation. The training set of polygons were buffered negatively (inward) by 56 m (equivalent to an AWiFS pixel) to help assure no edge pixels (which could be spectrally mixed) could come into play negatively in training of the classifier. Finally, both datasets were rasterized to match the same projection and grid cell sizes of the raw imagery data. Figure 5. Example USDA/FSA CLU-based ground truth polygons.

6 With the ground truth in place, a supervised classification was applied in the same manner to each of the image pairs utilizing the intersection of the training set against the raw imagery to obtain the samples. Rulequest See5.0 was used to derive the classification tree using it boosted algorithm and the NLCD Mapping tool (available at was used to interface the See5.0 application with ERDAS Imagine. The See5.0 derived decision trees were then applied back to the raw data and the classification created. Non-agricultural areas, for which there is no ground truth information to provide training and validation by the FSA dataset, where simply burned-in after the fact as depicted by the 2001 National Land Cover Dataset (NLCD). The classifications were then compared against the raster imagery derived from the validation set of polygons and error matrices created (Congalton and Green, 1999). Because there is no true ground truth for non-agricultural areas, the assessments only pertain to cropland cover types. Figure 6 depicts classification outputs from a small subset (albeit a different area from that shown in Figure 1) of the North Dakota area classification. On a macro level both classifications identify fields about equally. Both classifications exhibit a certain percentage of pixel noise but overall patterns are about the same. Field boundaries are more defined in the based classification. The overall for Table 2. Accuracies for top three cropland types found in North Dakota study area. AWiFS producers AWIFS user s producer s users Spring wheat 67.9% 62.4% 72.6% 64.4% Soybeans 64.2% 51.9% 67.0% 53.6% Sunflowers 32.0% 36.3% 32.4% 38.4% the AWiFS classification was 50.1% while the classification was slightly higher at 52.4%. Overall Kappa values were and 0.415, respectively. The top three row crops by acreage in the study area are spring wheat, soybeans, and sunflowers and their accuracies are presented in Table 2. Spring wheat and soybeans gave respectable results while sunflowers struggled. Conditional Kappas for the top three row crops are not shown but trended similarly for each case. It should be noted that these numbers are much lower than a typical NASS classification where normally several scenes collected throughout the growing season are used in additional to ancillary data such as elevation. Figure 6. AWiFS (left) and Classification output example from North Dakota region. Dominant crops include sunflowers (yellow), soybeans (brown), and spring wheat (tan). Non-agriculture is black. Table 3. Accuracies for top three cropland types found in Wisconsin study area AWiFS producer s AWIFS user s producer s users Corn 71.7% 68.7% 75.8% 75.2% Soybeans 69.1% 59.5% 77.2% 62.6% Alfalfa 23.6% 35.1% 33.8% 43.5% In the same manner as for North Dakota, tandem classifications were also run on the Wisconsin datasets. Overall accuracies were 50.4% for the AWiFS data and 55.6% for the. While AWiFS accuracies were about the same for both the North Dakota and Wisconsin analyses, the showed greater improvement in the Wisconsin example. Kappa

7 values for the Wisconsin AWiFS data was and for the. Again, an improvement from the differences found in North Dakota. Primary crops (Table 3) were corn, soybeans, and alfalfa. The later performed very poorly. Visually the Wisconsin example showed more differences between the two (Figure 7). Field delineation is evident in the data but within the AWiFS subset in all cases but for the largest fields the data appears very noisy. The classifier had a hard time interpreting much of the AWiFS based classification when field sizes were very small. Figure 7. AWiFS (left) and Classification output example from Wisconsin region. Dominant crops include corn (yellow), soybeans (orange), and alfalfa (purple). Non-agriculture is black. DISCUSSION AND CONCLUSIONS Classifications from outperformed those over identical areas from AWiFS. The classification improvements were 2.3 percentage points from the North Dakota test area and 5.5 percentage points over Wisconsin. Respectively, this translated to 4.6% (2.3/50.1) and 10.9% (5.5/50.4) gains going from AWiFS to data. Wisconsin tends to have a more complex landscape and thus likely why saw larger improvements there. It is believed the two study sites here are representative of much of US agriculture and thus the 5 to 10% should be general rule to be expected elsewhere. Accuracy of non-agriculture classes were not mapped here due to lack of ground truth but it is believed those cover types with finer spatial detail (such as urban features) would have the most to gain from. In an ideal world NASS would prefer the finer pixel resolution that provides but because it has a much narrower swath and 26 day versus five day revisit rate is still impractical for capturing multiple cloud free scenes over the same area. Cost is also an issue because per scene the two are currently the same price to the USDA but it takes over five times more scenes to cover the same area. It was hoped this study would better confirm what the optimal pixel size is for mapping agriculture. This study shows it is closer to the 23.5 m than the AWiFS 56 m but it is not clear if utilizing yet smaller pixels would improve mapping outcomes further. It appears clear that though if NASS is to create CDLs in regions with fields smaller than those typical found in the Midwest, is very appealing, if not necessary.

8 REFERENCES Boryan, C., and M. Craig, Multiresolution Landsat TM and AWiFS sensor assessment for crop area estimation in Nebraska, Proceedings from Pecora 16, Sioux Falls, South Dakota, American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland, unpaginated CD-ROM. Champagne, C., H. McNairn, J. Shang, D. M. Johnson, Evaluation of Resourcesat-1 AWiFS data for producing an agricultural crop inventory for Canada, ASPRS 2007 Proceeding of the ASPRS 2007 Fall Specialty Conference, Ottawa, Ontario, Canada. American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland, unpaginated CD-ROM. Chander, G., and P. L. Scaramuzza, Cross-calibration of the Landsat-7 ETM+ and Landsat-5 TM with the ResourceSat-1 (IRS-P6) AWiFS and sensors, Proc. SPIE Int. Soc. Opt. Eng. 6407, 64070E. Congalton, R.G., and K. Green, Assessing the Accuracy of Remotely Sensed Data: Principles and Practices, CRC Press, Boca Raton, Florida, 137 p. Craig, M.E., A resource sharing approach to crop identification and estimation, ASPRS 2001 Proceedings of the 2001 Annual Conference, St. Louis, Missouri. American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland, unpaginated CD-ROM. Himanshu, D., D. Chirag, P. Sandip, S.S. Sarkar, P. Himanshu, S.R. Joshi, A. Mishra, M. Detroja, AWiFS Camera for Resourcesat, Proceedings from SPIE Int. Soc. Opt. Eng., Vol Johnson, D.M. (in press). A comparison of coincident Landsat-5 TM and Resourcesat-1 AWiFS imagery for classifying croplands, Photogrammetric Engineering and Remote Sensing, accepted, revised. Mueller, R., Categorized mosaicked imagery from the National Agricultural Statistics Service crop acreage estimation program, Proceedings of the ASPR 2000 Annual Conference, Washington, D.C., American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland, unpaginated CD-ROM. National Remote Sensing Agency, IRS-P6, Data Users Manual, Department of Space, Government of India. Balanagar, Hyderabad, 141 p.