Detecting artificial areas inside reference parcels. A technique to assist the evaluation of non-eligibility in agriculture

Transcription

1 1 Detecting artificial areas inside reference parcels. A technique to assist the evaluation of non-eligibility in agriculture R. de Kok, C.Wirnhardt EC Joint Research Centre, IES Motivation Wall-to-wall seasonal coverage has become possible. Does it offer a basis to think about additional analysis or alternatives? Visual interpretation > 0.2 e6 polygons is not a solution. Now, wall-to-wall is at 5 meters but what if it becomes 50 cm? 2 source: 1

2 3 Topics in sequence; Automatic analysis needs one master rule set using transferable feature attributes. Contrast is such a feature attribute, applicable for; adjacent, large and very large neighborhoods. What kind of polygon population will be screened.(farmers block, eligible land, declared parcel). Single mosaic classification is insufficient, multi-stage might be necessary Cost per square KM² is a guideline. Main focus remains on anomalies in agriculture 4 NDVI-- Vegetation Index Based on physical properties; Absorption in Red, Reflection in Infrared Contrast-xxx -- Build up Index (cf. Pesaresi et.al. 2008) Could be based on; Shadow casting, an adjacent feature (on local neighborhood)? Self-repeatability, a proximity feature? ( Pesaresi, M.,Gerhardinger, A. Kayitakire F. 2008; A robust built-up area presence index by anisotropic rotation-invariant textural measure, IEEE Journal. of sel. top. in app. EO and RS, Vol 1. No. 3 Sept 2008) 2

3 5 CONTRAST, Complexity reduction and information preservation 6 CONTRAST, Can be formalized From : de Kok, R. and Taşdemir, K., Analysis of high-resolution remote sensing imagery with textures derived from single pixel objects. SPIE Remote Sensing Conference 8181, Earth Resources and Environmental Remote Sensing/GIS Applications, September, Prague. 3

4 7 dist=1 d=10 OBIA contrast allows experiments on large and very large neighborhoods From a (Binary) Value to a Value-Range 100g g d=50 d=

5 9 d =1 10 d =25 5

6 11 d =100 the Sinkhole Effect 12 RapidEye Red-Band 6

7 17th GeoCAP Tallinn MARS unit 17th GeoCAP Tallinn MARS unit d =25 7

8 e6 Polygons 35 out of 255 Tiles 80 hours processing 16 8

9

10 17th GeoCAP Tallinn MARS unit 17th GeoCAP Tallinn MARS unit Switching from Parcels to 250 Meter blocks. It does change statistics but not remote sensing results. 10

11 21 Detecting spatial clustering is an option 22 11

12 23 24 Source; DMC_II 12

13 25 1 Year Ago 26 13

14

15

16 31 Additional data Multi-stage classification becomes a necessity in order to remove commission failures, especially in textural vegetation detection. The spectral features in a RapidEye single mosaic causes trouble in classifying the tiled-imagery using a single master-process, due to large seasonal differences in the tiles. This influence is limited for artificial surfaces and they are therefore the first candidate to test the production time and protocol conditions. As an alternative to the development of more advanced classification techniques, the choice for extra (cheap) data allows to diminish this problem. Adding additional 5 to 15 meter Red and Infrared information to the mosaic of RapidEye certainly solves a lot of classification failures. The choice is the budget. What is a reasonable price per KM²? Conclusions: 32 o Extension of the search radius allows for a zonal definition. o Shadow is important to explain contrast for adjacent neighborhood but larger neighborhood contrast is more complex. o The extended context (self similar) defines the response to contrast over larger and very large neighborhoods, producing for example the sinkhole effect o The strategy aims at reducing the omission failures. This favors commission failures which require additional data. As long as additional cheap data is not available, the search for a better classifier remains. o OBIA contrast for single pixel objects has a modest effect on overall classification but is promising for special classes (artificial surfaces). o OBIA single pixel pre-processing also allow to exclude pixels from further segmentation and keep the single pixel information available at the lowest level (density, relative area, number of..etc.) 16