Hyperspectral Imagery for Environmental Mapping and Monitoring

Editor's Notes

• #3: Currently available techniques for non destructive herbage measurements have a limited accuracy and in the absence of fast and automatic means to monitor grassland, the quality of grassland management strongly depends on visual judgments (systematic ground observation campaigns ) and laboratory analyses (to determine grass quality e.g. proteins or cellulose content). Unfortunately, this approach needs a long time to be realised and can be expensive considering the important intra- or inter-parcel variability typical for this type of agricultural production. With the recently airborne imaging spectroscopy development, a new method for non-destructive grassland characterization can be explored.
• #5: With the CAP reform, the European Agriculture should become more competitive and at the same time environmental friendly. AEM are now compulsory and consolidates in many rural development plans in EU. In the Walloon Region, AEM are the subject to specific regulations, and differents constraints (Temporal, Spatial, Technical) have been defined. This study is focused on the temporal constraints and evaluates the opportunity of using airborne imaging spectroscopy to calculate different indices to trace the recent history of grasslands and headlands in terms of managements.
• #6: The 2003 campaign is: On the one hand a continuum and a possible validation of the first flight campaign using CASI-2 (VIS/NIR) and SASI (SWIR) sensors performed in August 2002 on an other Belgian site (Lorraine test site). This previous study had determined that the good spectral resolution of these sensors allows to estimate the quantity (e.g. wet matter, biomass, grass height) and the quality (e.g. protein, VEM, DVE) of grass canopy and so to establish regional inventories on grass production/quality potentials. These first results were very promising and this new campaign were an opportunity to validate these results in different conditions (different places, different moments) On the other hand, this study is innovative and tries to enlighten how imaging spectroscopy are capable to facilitate monitoring and assessment of policies and AEM.
• #9: The study area of this project is the Municipality of Attert, located in the southest part of Belgium (with center coordinates). Attert is a representative grassland region where environmental aspects are important as it is located in a Natrura 2000 area. In the year 2002, three quaters of the declared parcels (about 600) are grasslands and the total acreage of grasslands represent two thirds of the total agricultural area. 33 parcels were followed : Ground cover was scored visually from the middle of May to the airborne flight In addition to these observations, ground measurements were realized simultaneously with the airborne flight.
• #10: A subset of 17 parcels were selected, and 4 sampling units were defined per parcel and georeferenced with a dGPS. Ground measurements were performed simultaneously with the hyperspectral airborne flight. They included a measure of: - grass height, - grass biomass, - floristic composition and - management or cropping techniques used (pasture, meadows, date of cutting, etc.). Two types of samples were considered, depending on the cutting technique used: - a harvest of the whole canopy with a mower and - a harvest of the upper part of the canopy with scissors. Additionnal ground measurements were performed with field spectroradiometers.
• #11: For each dates management practices were observed that allow to identify an Observed Management Class (MCo) for each parcel. In total 3 parcels with AEM (MCo=NP NF) were present on the study area with no cut and no grazing authorized before 30/ 06. 12 parcels (MCo=P) were grazing land.
• #14: The flight has been scheduled during the last two weeks of June and the beginning of July. This period coincides with the cutting and grazing periods and the control of the agri-environmental constraints. Priority was given to the use of spectral bands from the CASI sensor. However, the project required high spatial resolution therefore data were acquired with a minimum of 2.4m by 2.4m spatial resolution along the nadir line. This resolution can indeed be considered as adequate to spectrally identify homogeneity and/or heterogeneity per pixel and per parcel. A total of 96 spectral bands with approximately 6 nm spectral sampling interval in the region from 400 to 900 nm and the ATM thermal radiation emitted by the earth surface in one spectral TIR band were recorded. Level IIb = radiometrically and geometrically corrected. The project has checked the signal and geometric quality of the images We compare field spectrometer measures with spectral response coming from images and field measurements and GCP were used to ensure the sampling unit were located as accurately as possible on the imagery. To obtain representative spectral response for each sampling unit, the CASI-SASI pixel responses were averaged within a 3x3 pixel subset, centrered over the sampling unit location.
• #15: Level IIb = radiometrically and geometrically corrected. The project has checked the radiometic and geometric quality of the hyperspectral images We compare field spectrometer measures with spectral response coming from images and Field maps and GCP were used to ensure the sampling unit were located as accurately as possible on the imagery. To obtain representative spectral response for each sampling unit, the CASI-SASI pixel responses were averaged within a 3x3 pixel subset, centrered over the sampling unit location.
• #16: Samples from each elementary sampling unit were also analysed to determine chemical characteristics of grass. They concern, dry matter content, proteins content, cellulose content, ashes content, sugar content and energy values (VEM, VEVI, DVE and OEB). NIR Spectrometry laboratory-based measurements (on fresh samples havested with scissors) were realised in the same wavelengths as airborne sensors.
• #17: To determine best hyperspectral wavebands for the study of grasslands over the spectral range of 400-1300 nm, and To assess grasslands classification accuracies achievable using various combinations of hyperspectral narrow wavebands
• #18: The methodology is illustrated by this figure. We have considered two approaches - With spectra indices (typical multispectral analysis) - With narraw-waveband du to the spectral resolution This two approaches were realised on pretraited data with a moving average of reflectance and the first derivative calcul
• #19: CASI: Most of the biochimical characteristic are closely linked with narrowbands located in the green-edge domain On the other hands, physical characteristics as grass height & wet matter are more related with red-ege and NIR domain The structure of the chlorophyll red-edge was best represented by calculating the first derivative of the reflectance spectra with respect to wavelength SASI: - Only reflectance are significative except for the derivative computed in the range of 1200-1300 nm
• #20: In addition to the standard CASI-SASI channels a number of channel ratios and normalised channel difference indices were developed. It was decided to limit the investigation to published vegetation indices In addition a « red-edge » transformation was calculated from the bands defining the begining and the end of the red-edge of the grass vegetation response.
• #23: Combination of reflectance, spectral indices and first derivative results
• #26: Scatterplots of measured grassland characteristics versus calculated vegetation indices - Best results were found using the red-edge step (REST) to estimate height and green weight of grassland canopy - the water balance index (WBI2) for dry matter content.
• #28: Different textural indices were calculated depending on the type of approach considered. In a global approach, all the pixels of a parcel are analysed simultaneously while in a local one the analysis is made with a moving windows of 3x3 pixels. In the first case, the texture parameter is simply the variance of all the pixels reflectances for each narrow band and in the second case, this parameter is calculated from moving windows determining the variance of the 3x3 pixels windows which are then aggregated over all the parcel to produce a local variance (intra window variability) which is used in complement of the total or global variance.
• #29: As the number of bands available for analysis is important with hyperspectral data, many of them are highly correlated and provide redundant information. Band selection refers here to the use of a subset of the most important wavelengths for identification of grassland management practices. In a first step, we used a discriminant analysis to identify regions of interest in the reflectance spectra and to choose the relevant vegetation indices. For the global texture parameter, two regions of interest in the reflectance spectra are clearly identified. The first one is located below 500 nm and the second one above 725 nm. The response obtained for the local texture is quite different and presents a minimum value of the probability around 500 nm and above 750 nm but the observed level is not significant (P<0.05). Based on these results, two wave bands from the global texture analysis have been selected (450 nm and 725 nm) together with previously defined indices to proceed in a second step to a discriminant analysis with all the selected dependent variables.
• #30: Even with hyperspectral images characterised by a poor quality, these results are very good and promising
• #32: It is satisfactory for a first screening and identification of the parcels which have to be controlled in the field, minimising by this way the cost of the field campaign.
• #33: Red-edge step and WBI2 where also successfully used to discriminate between grassland management types and different canopy structures. Chemical characteristic of grass was not clearly linked to vegetation indices, except for energy values VEM and VEVI which seem well correlated with NDVI (R²=0.60 ) In most of cases results from scissors cutting samples which represent the upper part of the canopy were best correlated with vegetation indices.
• #34: In general the CASI sensor gives better results than the SASI sensor. Except for 3 cases: Total dry matter production (called biomass) The Dry matter content The Wet matter content The combination of the 2 sensors doesn’t give a better prediction The prediction level est good in general except for cellulose and total dry matter content
• #35: The relative RMSE is lower for the scissors samples so the prediction is better with the upper part of the canopy For the most parameters, the Relative RMSE is lower than 10 % and show a godd quality of prediction Except for sugar and one energy value (OEB)
• #36: The results of this study showed the potentialities of hyperspectral imagery for the characterization of grasslands. The quality of relations between physico-chemical parameters and various spectral components (eg. reflectance curve, first derivative, spectral indices) allows to estimate the quality of grass canopy and so to establish regional inventories on grass production potentialities, which constitute a potential decision tool for local agencies.
• #37: They allow also to establish a discrimination between the various types of meadows (pasture, mowed meadows, etc.). And to ….
• #38: If we consider the parcel limited by the blue lines, we can clearly see thatthe farmer has devided it in 2 parts The lower part with actual pasture The upper one for a next pasture which would occur later in the season Finally, this study also showed the interest of data collected with the CASI sensor compared with the SASI sensor. Qualitative differences observed in results can be explained by the geometrical quality of images received by the project and the SASI sensor sensibility regard to background noises generated by environment (eg. soil, atmosphere).