Preliminary results of landcover analysis of Calanda area using ERS-1/2 SAR and Landsat-TM data

Abstract

Temporal evolution of land cover types was studied by means of remotely sensed data. Landsat Thematic Mapper and ERS-1 SAR data sets were available. Comparing automatic TM image classifications from different dates, we achieved detailed statistics of land cover changes. Classification schemes of ERS images were tested, and also in combination with TM data. TM classifications are more accurate than that obtained with SAR and allows temporal change monitoring of 9 classes. Although ERS-1 data had given worse results, interesting information were found.
Keywords: land cover evolution, Landsat Thematic Mapper, ERS-1 SAR, image classifications.

Introduction

Teruel is one of the poorer provinces of Spain. Desertization and downing economy forced to a population loss. The Calanda Desert area is one the most affected zones. Besides, an interesting geographical phenomenon can be seen: in 1981 ENDESA, an Electrical Company, started an open pit exploitation near Andorra village and in 1986 it was extended. After 11 years from the beginning of activities, 176 Has were restored for agriculture and natural vegetation. Remotely sensed data make possible a digital analysis of land cover and thus economical activities.

This paper presents the results of land cover changes using TM and the preliminary results of ERS classifications.

Test area presentation

The study area is located at inner North East of Spain:

The climate is semi arid Mediterranean of cold winters, with annual rain rate of 500 mm, frequent and dominant West winds.

The soil has grown upon a geological substratum of limestone, sand and clay rocks, with fine sand texture, basic (pH = 7-8) and poor in organic materials.

These climatic and soil characteristics, in junction with roughed and gully terrain, cut by steep riverbanks of variable aspects affected by erosion phenomena determine a poor vegetation, esclerophylous high-medium shrubs, like Juniperus comunis and Quercus rotundifolia, and low shrubs, like Rosmarinus officinalis and Thymus vulgaris. There are also Pinus pinaster, and natural pulse herbaceous vegetation. They show that the soil is degraded and low productive.

Some olive and almond trees plantations are present within the valleys, but the main culture lands are dedicated to dry cereal farming (ENDESA, 1994).

Methodology

Following materials were available for this work:

- ERS-1&2 frames from 1993 to 1996

- LANDSAT-5 Thematic Mapper images, from 1984 and 1995

- CORINE digital land cover map and visually interpreted sheets

- Digital Elevation Model (DEM)

Speckle filtering and geocoding of ERS data.

All ERS data were read, subset and transposed in SAR data reading and handling software developed at INDRA Espacio.

Two speckle filters also developed at INDRA Espacio were tested.

a. Geometric filter: 1 iteration was enough for decreasing noise but preserving edges and linear features quite acceptable. It created homogenous zones necessary for further procedures.

b. Sigma filter: noise reduction level was poorer than the geometric filter, and there were too many isolated pixels left even using 5 by 5 window size iterations.

Thus geometric filter was chosen for this task.

Filtered data were grouped in descending or ascending stacks, and then geocoded against TM and DEM composed images. We used a shaded DEM illuminated like a SAR image combined with TM bands. TM bands were combined upon this shaded DEM, for retrieving radiometric contents.

Common point searching procedure had to be done with certain guidelines:

­ Looking for zones not affected by high relief; they could be entry of gorges, top of shallow hills or the bottom of small valleys, and river junctions; railways and roads are excellent, if available.

­ Once the points similarity is sure, it had to be found pixel-size level spatial and spectral pattern coincidence, i.e., the elements form and tone -even though the SAR and TM data have no correlation, it is frequent to find similarities between SAR and near-infrared channel (TM4), due to their sensitivity to surface water content. A shallow hill could have the same shape in both imagery, and its top divides quite clearly both illuminated and shaded slopes.

For ascending pass, shading required the following parameters: Sun azimuth = 255º and Sun elevation angle = 67º. Sixty five (65) control points have been initially found, with a total RMS error of 164,12 meters (6,56 pixels). After deleting points with higher errors, 20 points were left, with total RMS error of 22.95 m (0,92 pixel).

For descending pass, shading required the following parameters: Sun azimuth = 95º and Sun elevation angle = 67º. As well as ascending pass, TM bands were combined upon this shaded DEM, for retrieving optical and infrared radiometric contents. We started with 52 common points, with an initial RMS error of 181 m. After rejecting non-accurate points, 15 points were left. Total RMS error: 15.63 m. Then final ERS geocoded descending pass layer stack were generated.

Land cover temporal evolution analysis.

This analysis were carried out comparing TM classifications of 1984 and 1995. Supervised method had been used. The first classification we carried on was 1995 TM. This will be used as a guide for SAR classification. CORINE legend had to be modified and generalized to better fit to our study area and classifying conditions. Therefore, this legend was established:

Training fields were delimited for retrieving spectral signatures. Maximum Likelihood algorithm was used, and mode filter was applied, for avoiding "salt and pepper" effect. An accuracy assessment was performed based on 242 stratified-random verification points, achieving a high accuracy (more than 80 %). A tabular class inventory was generated, including categories extent.

1984 TM image was also classified. In 1984 differences between croplands and natural vegetation were sharper than now. The overall accuracy level for this one was 78,64%.

Comparing with 95 data, we have retrieved a simple chart of changes, as can be seen in figure 3.

Major changing categories (more than 500 Ha) are #2 Dry Crops, #3 Irrigated Crops, #4 Pasturelands, #5 Shrublands and #6 Pines. #1 Artificial Surfaces had a significant variation, due to natural and culture lands recovering activities from open pit mining zones, in South-West of study area. Croplands, in general, have decreased, and natural vegetation of esclerophylous Mediterranean shrubs and pine plantations have increased. These phenomena show that agriculture activities are going down, meanwhile natural vegetation are recovering their original space, specially in hilly areas. The point is finding out whether the decrease of agricultural categories are caused by economic abandonment of activities.

Classification of ERS data.

At first sight, TM training sites superposed upon SAR images looked like confusing. Specially, artificial surfaces and relief slopes gave very similar response. There was no clear spectral or spatial patterns that allowed to distinguish different cover types, except water surfaces, dry lagoons, railway-looking linear features and villages. It was specially clear the V-shaped lignite processing factory and mining zone in Southwest of the area. We had to reshape training sites to match to ERS data. Finally, only descending pass data were tested.

We included two synthetic bands derived from SAR images, according to Nezry et al. (1995) studies: a) Multitemporal mean image and b) Maximum variation image: absolute maximum difference between dates. It would be possible to improve SAR layer separabilities by means of generating statistical bands that summarize some variables. We also introduced an artifact, adding +20 DN succession for each images, starting with the second date. Signatures means chart showed better separation between categories than former ERS attempts.

As result of separability analysis (without using a priori probability) 3 images were excluded, and following combination were used as input: 23 June 93, 10 November 93, 18 December 94 and Maximum Variation. There were some overclassified classes, like #7 and #3. Majority filters were applied two times, first ignoring class #7 (5x5 pixels window), and second time ignoring classes #3 and #6 (3x3 pixels window).

Then classification was made and accuracy assessment was applied, with 255 stratified random points, and reached to an overall accuracy of 45.42%. This is a very low result, in spite of error filtering, but it could be considered normal, taking in account that we have used intensity images.

A second classification attempt was made, using as inputs: 23 June 93, 10 November 93, 18 December 94, Maximum Variation, TM band #4 (nIR) and slope information derived from DEM.

Also this time Maximum Likelihood algorithm was used, and the accuracy assessment a subtle improvement, although still low: 57.42%.

Conclusions

Plates 1 and 2 show 1995 and 1984 TM classifications respectively. Plate 3 shows both descending and ascending ERS image stacks color compositions. Plate 4 shows ERS descending pass classification.

Temporal evolution of land cover types were carried out successfully using TM classifications. More changing categories were Dry Farming, Shrublands, Pastures and Coniferous. Further studies could aim to correlate these results to possible environmental impacts of lignite mining activities, specifically that affected to dry farming. Either could be caused by actual pollution problems or another economical reasons, like Common Agricultural Policy or rural population loss.

SAR classification using descending pass and Maximum Variation images retrieved poor preliminary results. Although some unexpected classes appeared, by means of avoiding a priori probabilities in the Maximum Likelihood algorithm, such as Coniferous; and Irrigated Crops were quite well discriminated. Recognition of Coniferous could be caused not for its signature, but there exists high correlation between coniferous and steep slope areas. Water and Dry Lagoons were accurately assigned, meanwhile Dry Crops and Shrublands were significantly discriminated (over 60% of users accuracy).

Shallow relief of most part of study area allowed an accurate geocoding of SAR data. This is an important point for combining other data sources and making ground-truth analysis.

Further works will take in account other variables, like topographic data, TM bands merging, and SAR images could be enriched using sigma-naught and coherence images. Also ascending and descending pass combination variables will be tested.

References

ENDESA, 1994, Minería a cielo abierto y medio ambiente en Andorra, Technical Notes. Zaragoza.

Nezry, E., Rémondière, S., Solaas, G. AA., Genovese, G., 1995, "Mapping of next season's crops during the winter using ERS SAR". Earth Observation Quarterly, 50 - December, pp. 1-5.

Curlander, J.C.; McDonough, R.N., 1991, "Synthetic Aperture Radar- Systems and Signal Processing". Wiley InterScience, 647 pages.

Leberl, F.W., 1989, "Radargrammetric Image Processing".Artech House, 595 pages.

Keywords: ESA European Space Agency - Agence spatiale europeenne, observation de la terre, earth observation, satellite remote sensing, teledetection, geophysique, altimetrie, radar, chimique atmospherique, geophysics, altimetry, radar, atmospheric chemistry