Digital Classification vs. Visual Interpretation: a case study in humid tropical forests of the Peruvian Amazon Carlos Javier Puig1, Glenn Hyman2 and Sandra Bolaños3 1, 2, 3

1

International Center for Tropical Agriculture (CIAT) Km 17, carretera Cali-Palmira. AA 6713, Cali, Colombia

[email protected]; 2 [email protected]; [email protected]

Abstract – Remote sensing specialists classify land use and land cover from satellite imagery by visual interpretation or by digital classification. The latter method is widely accepted among the scientific community because of its statistical validation and automatic processing. However, precision and accuracy are difficult to achieve in tropical environments where landscape heterogeneity is common and field work is difficult. During the last decades, visual interpretation became less popular compared to automatic processing. But now software and hardware are able to manage images and vector information, permitting interactive map editing and more possibilities for visual interpretation. CIAT has been working in the Peruvian Amazon region monitoring tropical forest and land use change near to the city of Pucallpa. Several classifications were made using visual interpretation and digital image processing in the same area. We compared and described both methodologies in a sample area of the Peruvian Amazon region. The normal fragmentation of tropical deforested areas and the diversity of uses and covers affect the selection of good training areas for digital classification. Visual interpretation is imprecise for recognizing and drawing land cover polygons. Both methodologies required similar processing times. Both were affected by the presence of clouds and their shadows, heterogeneity in the distribution of land use and land cover over the study area, minimal details identified from the classifications, and number of classes. Interpreter experience played an important role in processing time for both classifications. Statistically, digital classification is more accurate if a confidence interval of less than 60% is considered; above this value, both classification methodologies were comparable. Kappa statistics were similar. Visual classification is the preferred method for interpreting land use and land cover in low and mediumresolution satellite images, its application being limited

to high spatial resolution imagery because of the increase of details to recognize. INTRODUCTION

Remote sensing technology, software and data management have continually improved since the middle of the last century. Traditional visual classification of aerial photographs was adapted to the first satellite images. Advances in technology and mathematical algorithms permitted digital image processing and automated classification. Supervised and unsupervised classifications are the two principal methodologies for making land cover maps. Both methods consider the brightness value of pixels and the generation of groups of pixels with similar spectral response, using specific algorithms. Advances in computer, GIS and remote sensing technology offer new possibilities for managing, editing and generating raster and vector data, facilitating the visual interpretation methods. In this case, image data is displayed on a monitor. The interpreter uses a mouse to digitize vectors on top of the displayed imagery. Although automated classification is widely recognized for its statistical validation, visual interpretation has more potential users. Many organizations lack the resources for investing in expensive software and remote sensing training. They could benefit by using easy-to-learn visual interpretation methods with low-cost software. Some previous comparative studies showed that some land cover types in tropical regions have reflectance characteristics easily distinguished on an image. Areas with heterogeneous land covers require the analyst to consider other aspects such as texture, spatial pattern and fragmentation of the landscape for understanding an image (King, 2001). Complex areas poorly

1

classified by automatic processing could be masked using visual techniques in order to improve the classification. Our study compares and discusses the advantages and disadvantages of digital classification and visual interpretation for making land cover maps in the lowland tropics. We made both quantitative and qualitative analyses to understand the differences in the two methodologies. MATERIALS AND METHODS

This comparative study was carried out for our study area in the Central Peruvian Amazon, near the city of Pucallpa (Fig. 1). The region suffered significant deforestation during the last two decades due to tropical wood extraction, shifting cultivation, cattle ranching, and plantation agriculture. (Hyman et al, 2002; JRC, 1998).

identify deforestation hot spots in the global tropical belt. Two Landsat TM satellite image scenes were used; the first corresponds to path 006, row 066 from July 18th, 1998, and the second also from the same position, but from October 17th, 1996. Historical rain data indicates that June to October corresponds to the dry season in the study area (IIAP, 1996). The study site covered a small area of the whole image. The first image was interpreted using the data from the TREES project. A hierarchical key that involved a wide possibility of classes was used. Scale of 1:100,000 on the screen was the maximum detail used to digitize the image. We classified the second image by digital processing using the K-means algorithm in an unsupervised classification. After several tests, the misclassification areas were edited using masks and recoded to reduce the confusion between classes. Small areas were eliminated with a modal 3x3 kernel filter. Both classifications had five (5) final classes corresponding to forest, non-forest, water bodies, advanced forest re-growth and clouds with their corresponding shadow. The visual interpretation from the TREES project was reclassified to these classes. We acquired 108 GPS points from the study site, most of them located in the non-forested areas. All these points were used to test the classification accuracy of each method. We acquired these sample GPS points during a survey campaign in April and October of 1998.

Fig. 1. Study site localization

To develop the comparison between digital classification and visual interpretation methods, we used satellite imagery from two studies carried out near Pucallpa, Peru. Digital classifications were taken from continuing studies on deforestation and land use by researchers at the International Center for Tropical Agriculture (CIAT). We used maps made by visual interpretation from the Tropical Ecosystem Environment Observations by Satellite (TREES) project, an initiative led by the Joint Research Centre (JRC) of the European Commission (REF). This study involved visual interpretation to

An accuracy matrix for each methodology was built to compare the GPS verification data with each classification. Our knowledge of the tropical forest dynamics of the study site, historical survey and census data and image interpretation were used to identify the land use below the corresponding GPS point at 1996 date. We evaluated differences in the two classification methods using “difference between proportions” test and kappa indices RESULTS and DISCUSSION

Two classifications were made over the same scene using two satellite images. We identified five land

2

cover classes (forest, non-forest, water, forest regrowth, clouds and shadows). Digital processing by unsupervised classification was used to iteratively combine 40 classes into the final five-class land cover map. We spent about the same amount of time to classify the imagery using the two methods. The length of time for visual classification depends heavily on interpreter experience, image quality, details to recognize and efficiency of hardware and software. Digital classification consumed more of our time than expected because of the iterative reclassification process and our filtering of the image. Narrowing down the number of classes was especially difficult in areas with mountains, diverse slope orientations, differing sun illumination, and high cloud and shadow density. Visual interpretation required that the analyst knows aspects of the study area in addition to the spectral response of the image. Our classification improved because of our knowledge of the relationship between the different land cover classes (context), texture and historical information of the study area (King, 2001). This experience helped us define classes that were more representative of the real terrain conditions (Singh and Czaplewski, 1994). In this tropical region, land use and land cover patterns tend to be heterogeneous, especially deforested landscapes that have several variations of the tree canopy density, regrowth, and land use. Conditions like this hinder the identification of useful training areas for digital supervised classifications. When the interpreter used visual classification sometimes the tendency was to generalize, especially when study area was fragmentedor composed of a mixture of land use cover classes. Small areas with grass and isolated tree groups normally were drawn inside a big polygon of pasture or intervened area without taking account of small patches of forest (De Grandi et al, 1998). Digital classification, on the other hand, recognized the two main classes of pasture and forest, but drawing several polygons (“Fig. 2”) instead of only a few in the visual interpretation. Statistically, both methods yielded similar precision measures when difference of proportions tests were carried out. Tests withconfidence measures of 60% showed no significant difference between the two methodologies. Below this value, both expressed differences with visual interpretation having greater accuracy. The Kappa test showed the same tendency

for both methodologies. Digital processing gave a Kappa statistic of 0.61 (61% precision), while visual interpretation gave a value of 0.61 (61.4% precision). The low accuracy of both methodologies could be influenced by the distribution of GPS control points. They had not been collected using statistical sampling methods. We collected GPS points from places in the field with easy access, concentrated in “non-forest” areas. Some classes such as “forest re-growth” had few GCP points compared to that class in the imagery. Additionally, the GCP points were collected using a low precision GPS receiver, a possible cause of misallocation of some points and potential confusion between some classes.

Fig. 2. Ways to interpret some characteristics by VC and DC methods.

Due to image filtering, small polygons of non-forest or forest re-growth inside of a larger forest area were incorporated into forest class, increasing its overall area. The same situation occurs for non-forest areas. When statistical tests are applied some confusion occurs due to small polygons that are difficult to validate with GPS points taken in the field. Therefore, future experiments should avoid taking verification points over small areas (e.g. 3x3 pixel polygons). These small areas are suppressed by the filtering process. Several studies showd that forest survival depends on how large the intervened areas are. If they are too large only small patches of forest remain, increasing the likelihood of species and biodiversity loss (Murcia, 1996; Bierregaard Jr., 1996). Digital processing considers these patches as forest. Visual interpretation cannot interpret them because they are too small to be considered as forest. Large forest polygons exclude non-forest patches with visual interpretation methods.

3

The polygons drawn using visual interpretation do not follow raster pixel boundaries. Verification analyses should avoid using test points near polygon borders since these areas have greater uncertainties. One important advantage of visual interpretation and vector storage of land cover data is that it requires less disk space. However when the interpretation involves large areas or huge polygons, the requirement of float memory is higher and the interpretation becomes slower. We had difficulties working in a networked hardware and software environment. Visual interpretation was much easier with processing and data storage on a local computer. While digital processing can create large numbers of classes, visual interpretation can create theseas well. The European CORINE system and TREES project demonstrated the utility of visual interpretation (Joint Research Centre of the European Commission 2001). When visual interpretation is the chosen classification method, only one person or a team trained under the same conditions should make the classification. This will ensure a homogeneous interpretation over the entire image. For success in the result, the interpreter must have appropriate knowledge of the study area. Sometimes when there is not a complete and appropriate set of image bands to analyze, visual interpretation is recommended to overcome lack of spectral differences in the imagery. CONCLUSION

Both methodologies gave us similar precision and processing time for the study area. Digital image classification gave better spatial detail of land use and land cover, although the classes are not always easy to adapt to a classification scheme. Digital classification required more time for editing and processing to reduce errors. Visual interpretation was done by analysts with no formal training in the digital image processing. This is an avantage if extensive remote sensing expertise is not available for a project. One solution to the drawbacks to each method is to use some kind of combination of automatic processing and visual interpretation. The analyst can do an unsupervised classification at the outset, and then correct it with visual interpretation and pixel editing. This combination allows you to mask out clouds,

shadow and water, then vectorize the remaining parts of the image visually. Another advantage of visual classification is that it can be done with simple GIS software in case the analyst does not have digital image processing software. We cannot recommend one methodology over another. Our results show that no substantial difference was found between the two methods. But if technicians want to analyze a satellite image using visual interpretation, they can utilize its many advantages and develop their studies with the same confidence as they have with adigital classification method. Visual interpretation was shown to have similar quality compared to digital classification for analyzing medium-resolution satellite data. The increase in spatial resolution for the new generation of satellites could be an important limitation for future studies due the increase in details that have to be identified, requiring more processing time. REFERENCES Bierregaard Jr., R.O. and V.H. Dale. Island in an Ever Changing Sea. The Ecological and Socio-economic Dynamics of Amazonian Rainforest Fragments. In: Schelhas, J. and R. Greenberg (eds). Forest Paches in Tropical Landscapes. Washington, USA. Pp. 3-18. 1996. De Grandi, GF; Mayaux, P; Rosenqvist, A; Rauste, Y; Saatchi, S; Simard, M and M. Leysen. Flooded Forest Mapping Regional Scale in the Central Africa Congo River Basin. First Thematic Results Derived by ERS1 and JERS1 Radar Mosaics. Retrieval of Bio- and Geo-Physical Parameters from SAR Data for Land Applications Workshop. ESA-ESTEC, Noordwijk, Netherland. Oct. 98. Hyman, G.; C.J.Puig and S. Bolaños. Multisource Remote Sensing and GIS for Exploring Deforestation Patterns and Processes in the Central Peruvian Amazon. International Symposium on Remote Sensing of Environment. In press. Buenos Aires. April, 2002. IIAP (Instituto de Investigación de la Amazonía Peruana). “Deforestación en el Area de Influencia de la Carretera Federico Basadre – Pucallpa”. Dec, 1996). JRC (Join Research Centre). Identification of Deforestation Hot Spot Areas in the Humid Tropics. TREES (TRopical Ecosystem Environment observations by Satellites).. Publications Series B, Research Report No 4. 99p. 1998.

Joint Research Centre of the European Commission. 2001. TREES World Wide Web page at the JRC. http://www.gvm.sai.jrc.it/Forest/defaultForest.htm King, B. Land cover mapping principles: a return to interpretation fundamentals. First annual meeting of the Remote Sensing and Photogrammetry Society. London, UK. Pp 467-484. Sep, 2001.

4

Murcia,C. Forest Management and Pollination of Neotropical Plants. In: Schelhas, J. and R. Greenberg (eds). Forest Paches in Tropical Landscapes. Washington, USA. Pp. 187-204. 1996. Singh. KD and RL Czaplewski. Report: Analyses of Alternative Sample Survey Designs. Food and Agricultural Organization of the United Nations. Forest Resources Assessment 1990 Project. Roma, Italy. Feb, 1991.

5