| 477 | |
| 478 | = PROBA-V PERFORMANCE ASSESSMENT = |
| 479 | |
| 480 | = ABSTRACT = |
| 481 | |
| 482 | The Belgian Federal Science Policy Office (BELSPO) has initiated a Preparatory Evalua-tion/Validation Programme for |
| 483 | the products of the new PROBA-V satellite to be launched in 2012. The satellite will allow daily monitoring of terrestrial |
| 484 | vegetation cover through remote sensing, and will cover the data provision gap between the closure of the SPOT/VEGETATION |
| 485 | Programme and the launch of the SENTINEL-3 mission. The aim of this study is to evaluate the improvements that PROBA-V will |
| 486 | bring along for forest monitoring in the Atlantic Biogeographical Region of Europe, and lies within the objectives of the |
| 487 | FM@PROBA-V project. A representative site in Northern Por-tugal is selected for this reason. VEGETATION, LANDSAT-ÔÌ5, and |
| 488 | MODIS data along with the JRC Forest Cover Map are used to train the classifiers, simulate PROBA-V data, apply the classi-fiers |
| 489 | at 250 m, 1/3 of a km, and 1 km pixels, and validate the results, while quantifying the accura-cies. Maximum Likelihood (ML), |
| 490 | Artificial Neural Networks (ANN), and Support Vector Machine (SVM) methods were tested. From the confusion matrices the best |
| 491 | result is obtained by MODIS 2 bands with ANN classifier. Further analysis on the base of those confusion matrices will be |
| 492 | applied to define the best classifier taking into account all the parameters of the matrices. The best per-forming classifier |
| 493 | will then be recommended to examine its robustness against sudden disastrous events, like fire, in the same area, performing |
| 494 | change detection between sequential dates (before and after the event). The performance of the data and classifiers are |
| 495 | demonstrated, and the pre-liminary results are discussed.The Belgian Federal Science Policy Office (BELSPO) has initiated a |
| 496 | Preparatory Evalua-tion/Validation Programme for the products of the new PROBA-V satellite to be launched in 2012. The satellite |
| 497 | will allow daily monitoring of terrestrial vegetation cover through remote sensing, and will cover the data provision gap between |
| 498 | the closure of the SPOT/VEGETATION Programme and the launch of the SENTINEL-3 mission. The aim of this study is to evaluate the |
| 499 | improvements that PROBA-V will bring along for forest monitoring in the Atlantic Biogeographical Region of Europe, and lies within |
| 500 | the objectives of the FM@PROBA-V project. A representative site in Northern Por-tugal is selected for this reason. VEGETATION, |
| 501 | LANDSAT-ÔÌ5, and MODIS data along with the JRC Forest Cover Map are used to train the classifiers, simulate PROBA-V data, apply |
| 502 | the classi-fiers at 250 m, 1/3 of a km, and 1 km pixels, and validate the results, while quantifying the accura-cies. Maximum |
| 503 | Likelihood (ML), Artificial Neural Networks (ANN), and Support Vector Machine (SVM) methods were tested. From the confusion |
| 504 | matrices the best result is obtained by MODIS 2 bands with ANN classifier. Further analysis on the base of those confusion |
| 505 | matrices will be applied to define the best classifier taking into account all the parameters of the matrices. The best |
| 506 | per-forming classifier will then be recommended to examine its robustness against sudden disastrous events, like fire, in the same area, |
| 507 | performing change detection between sequential dates (before and after the event). The performance of the data and classifiers are |
| 508 | demonstrated, and the pre-liminary results are discussed. |
| 509 | = INTRODUCTION = |
| 510 | |
| 511 | |
| 512 | World forests cover roughly 31% of land area and in the last decade it has been reduced at a rate of 13 million hectares |
| 513 | per year (1). According to the global forest resources assessment 2010 of the Food and Agriculture Organization, significant |
| 514 | progress has been made towards reversing the overall trend of forest area loss and a positive trend over time has been shown |
| 515 | in some countries and regions such as in Europe where the forest area continued to expand. |
| 516 | |
| 517 | Considerable efforts are needed to improve or at least maintain this positive trend. The forest monitoring gives crucial data |
| 518 | to decision makers helping them allocate appropriate financial re-sources for effective forest conservation and management plans. |
| 519 | One of these crucial data is For-est Area, which has been selected as one of the 60 indicators for monitoring progress towards |
| 520 | the Millennium Development Goals, the 2010 Biodiversity Target and the Global Objectives on For-ests. Forest Area and its changes |
| 521 | are relatively easy to measure, especially nowadays with the technological progress made in remote sensing fields. |
| 522 | |
| 523 | The available high temporal resolution images allow operational and near real-time applications at global, continental and regional |
| 524 | scales for forest area mapping and monitoring. The VEGETATION instruments, which were developed with the objective to provide data |
| 525 | specifically for vegetation canopy monitoring, are on board two SPOT5 satellites and will be available until 2013. The instru-ments |
| 526 | of this programme (VEGETATION 1 and VEGETATION 2) have monitored and mapped the worldwide vegetation on a daily basis for more than |
| 527 | 12 years now. They provide essential infor-mation on terrestrial vegetation cover for a large community of users. Therefore, to ensure |
| 528 | the continuity of the data after 2012, the SENTINEL mission, which will be launched at the earliest at the end of 2013 (2), will provide |
| 529 | data which will fulfil the needs of the current VEGETATION data users. To fill the data gap between VEGETATION-2 and SENTINEL-3, the |
| 530 | Belgian Federal Sci-ence Policy Office (BELSPO) has decided to build a satellite mission called PROBA-V which is expected to meet all |
| 531 | the specifications of the vegetation user community (2). PROBA-V will have an increased spatial resolution in comparison to the |
| 532 | VEGETATION instrument, an enhancement expected to provide new opportunities for forest monitoring. |
| 533 | |
| 534 | The aim of this study is to evaluate forest cover classification methods on the European Atlantic biogeographical region by applying the |
| 535 | Maximum Likelihood (ML), Artificial Neural Network (ANN) and Support Vector Machine (SVM) classification algorithms on simulated PROBA-V, |
| 536 | VEGETATION and MODIS data, in order to determine the optimal classification methodology for the region. This work is carried out within |
| 537 | the framework of the FM@PROBA-V project1 towards evaluation and quantification of the improvements of the products for forest monitoring |
| 538 | offered by PROBA-V, in relation to its predecessor, VEGETATION. The preliminary results of the study are presented and discussed in this paper. |
| 539 | == Study area == |
| 540 | |
| 541 | |
| 542 | The area of interest is the southern part of the Atlantic biogeographical region of Europe (Figure 1) which is located |
| 543 | in the north of Portugal. The region is one of the rare mountainous areas in this region, and mountains in the area can |
| 544 | reach an altitude of 1500 m. Forest cover is sparse and agriculture is not very intensive in the region |
| 545 | |
| 546 | = Methods = |
| 547 | |
| 548 | == Dataset and preprocessing == |
| 549 | |
| 550 | |
| 551 | The data used in this study included the images used for the evaluation process of the forest cover mapping in the European Atlantic |
| 552 | biogeographical region and the data used to create the refer-ence data for the classifications. The reference data originated from the |
| 553 | classification of a Landsat TM image, acquired on 11 July 2009, using training samples and reference points identified from the JRC |
| 554 | Forest Cover Map 2006 and Google Earth images. A small subset (20 „e 20 km2) of that scene was used to simulate a part of a PROBA-V |
| 555 | scene, a process undertaken by VITO NV, Bel-gium. The datasets used to evaluate classification algorithms included the simulated |
| 556 | PROBA-V image (Red, NIR, approximated SWIR), a 10-day synthesis VEGETATION (VGT) image at 1000 m spatial resolution with 3 bands (Red, NIR, SWIR), |
| 557 | and an 8-day synthesis MODIS image at 250 m spatial resolution (again Red, NIR, resampled/ upgraded SWIR). The simulated PROBA-V |
| 558 | subset was relatively small (20 „e 20 km2) and the results were expected to be of questionable statistical significance. Hence the |
| 559 | MODIS image was additionally evaluated, as its similar charac-teristics with PROBA-V data (Table 1) would allow better evaluation of |
| 560 | the potential of the upcom-ing sensor, and its comparison with VGT data. Since only the red and NIR channels are available at 250 m |
| 561 | resolution, the SWIR 500 m channel was resampled to 250 m and two MODIS datasets were evaluated, with and without the SWIR data. All |
| 562 | images, including the JRC forest cover map where registered to the Landsat TM image in UTM WGS84, Zone 29N projection, WGS84 datum. |
| 563 | The extent of MODIS and VGT images was reduced to a subset matching the extent of Landsat TM, which was topographically normalized with the ASTER GDEM. |
| 564 | |
| 565 | |
| 566 | {| border="1" |
| 567 | |+ Image data characteristics |
| 568 | ! !! VEGETATION II !! Modis !! Proba V |
| 569 | |- |
| 570 | ! Blue |
| 571 | | 0.43 – 0.47 || 0.45 - 0.47 || 0.44 – 0.48 |
| 572 | |- |
| 573 | ! Red |
| 574 | | 0.61 – 0.68 || 0.62 – 0.67 || 0.62 – 0.69 |
| 575 | |- |
| 576 | ! NIR |
| 577 | |0.78 – 0.89 || 0.84 - 087 || 0.79 – 0.90 || 800 x 800 x 1000 mm |
| 578 | |- |
| 579 | ! SWIR |
| 580 | | 1.58 – 1.75 || 1.62 –1.65 || 1.56 – 1.65 |
| 581 | |- |
| 582 | ! Spatial Resolution |
| 583 | | 1.15 km || 250m (R, NIR), 500m (SWIR) || 300m (VNIR), 600m (SWIR) |
| 584 | |} |