Segmentation and thematic classification of color orthophotos over non-compressed and JPEG 2000 compressed images

Abstract

Lossy compression is now increasingly used due to the enormous amount of images gathered by airborne and satellite sensors. Nevertheless, the implications of these compression procedures have been scarcely assessed. Segmentation before digital image classification is also a technique increasingly used in GEOBIA (GEOgraphic Object-Based Image Analysis). This paper presents an object-oriented application for image analysis using color orthophotos (RGB bands) and a Quickbird image (RGB and a near infrared band). We use different compression levels in order to study the effects of the data loss on the segmentation-based classification results. A set of 4 color orthophotos with 1m spatial resolution and a 4-band Quickbird satellite image with 0.7m spatial resolution each covering an area of about 1200×1200m2 (144ha) was chosen for the experiment. Those scenes were compressed at 8 compression ratios (between 5:1 and 1000:1) using the JPEG 2000 standard.There were 7 thematic categories: dense vegetation, herbaceous, bare lands, road and asphalt areas, building areas, swimming pools and rivers (if necessary). The best category classification was obtained using a hierarchical classification algorithm over the second segmentation level. The same segmentation and classification methods were applied in order to establish a semi-automatic technique for all 40 images.To estimate the overall accuracy, a confusion matrix was calculated using a photointerpreted ground-truth map (fully covering 25% of each orthophoto). The mean accuracy over non-compressed images was 66% for the orthophotos and 72% for the Quickbird image. It is interesting to obtain this medium overall accuracy to be able to properly assess the compression effects (if the initial overall accuracy is very high, the possible positive effects of compression would not be noticeable). The first and second compression levels (up to 10:1) obtain results similar to the reference ones. Differences in the third to fifth levels (20:1 to 100:1) were moderate to large (accuracies 61–58% for orthophotos and 67–65% for Quickbird), while more compressed images obtained the worst results (accuracies lower than 55%). As a comparison, the usual independent test areas (covering a small percentage of the classified area) were also used. The results show that this classification evaluation approach must be used with caution because it may underestimate the classification errors.