Ableitung von Vegetations- und Bodenanteilen für Dauergrünland aus dichten Sentinel-2 Zeitreihen
1Geographisches Institut, Humboldt-Universität zu Berlin, Unter den Linden 6, 10099 Berlin, Deutschland; 2Integrative Research Institute of Transformations of Human-Environment Systems (IRI THESys), Humboldt-Universität zu Berlin, Unter den Linden 6, 10099 Berlin, Germany
In Deutschland und Europa stellt Dauergrünland, als einer der artenreichsten Habitattypen, wichtige Ökosystemleistungen im Bereich der Klimaregulation, des Wasser- und Bodenschutzes und der ökonomischen Wertschätzung zur Verfügung. Gleichzeitig ist die Biodiversität besonders auf extensiv genutzten Flächen von Nutzungsaufgabe oder -intensivierung bedroht. Zeitreihen aus Satellitendaten enthalten vegetationsspezifische Informationen, welche sowohl Aufschluss über den ökologischen Zustand (z.B., Vitalität, Trockenstress) als auch die vorherrschende Bewirtschaftung (z.B. Mahd, Düngung, Beweidung) von Grünland geben können. In der Vergangenheit wurden optische Satellitendaten (z.B. Landsat und MODIS) bereits genutzt, um Vegetations- und Bodenanteile in natürlichen Grasländern der Erde zu quantifizieren und Degradation zu detektieren. Aufgrund der geringen zeitlichen bzw. räumlichen Auflösung sind diese Daten für das Monitoring von Weidegrünland oder Mähwiesen nur bedingt geeignet. Da Grünlandvegetation in Deutschland starke raumzeitliche Dynamiken innerhalb einer Vegetationsperiode aufweist, eignen sich dagegen Sentinel-2 Zeitreihen mit hoher räumlicher (10 m) und zeitlicher (5 Tage) Auflösung für die Charakterisierung von Grünlandnutzung.
In dieser Studie werden Methoden der spektralen Entmischungsanalyse, welche bisher vorrangig in natürlichen Grasländern angewandt wurden, auf bewirtschaftetes Grünland in Deutschland übertragen. Raumzeitlich hoch aufgelöste spektrale Sentinel-2 Zeitreihen ermöglichen es, diese Observationen in Zeitreihen von Anteilen grüner Vegetation, trockener Vegetation und offenen Bodens zu transformieren und somit thematisch interpretierbare Größen abzuleiten. Zeitreihen dieser Anteile lassen potentiell sowohl räumlich explizite Schlüsse auf verschiedene Bewirtschaftungsintensitäten als auch auf sich verändernde ökologische Zustände als Folge von Auswirkungen klimatischer Einflüsse (z.B. Dürreperioden) zu.
Die Vorprozessierung der Sentinel-2 Daten erfolgte mit Methoden des „Framework for Operational Radiometric Correction for Environmental monitoring“ (FORCE; Frantz, 2019). Um zeitlich gleichabständige Observationen zu erhalten, wurden Reflektanz-Zeitreihen mithilfe von Radial Basis Convolution Filters (Schwieder et al., 2016) interpoliert. Für jede Observation wurden darauf basierend Anteile grüner und trockener Vegetation sowie offenen Bodens mittels eines regressionsbasierten, spektralen Entmischungsverfahrens geschätzt (SynthMix; Okujeni et al., 2017). Dieses basiert auf Support Vector Regressionsmodellen, welche mittels synthetisch gemischten Trainingsdaten aus einer Spektralbibliothek parametrisiert wurden. Zur Validierung der Ergebnisse wurden hochaufgelöste, multispektrale Aufnahmen von Drohnen genutzt, welche für die Saison 2019 für Testflächen in Brandenburg und Niedersachsen zur Verfügung stehen.
Unsere Ergebnisse zeigen, dass die Vegetationsdynamik von Grünland durch Zeitreihen der drei Oberflächenkomponenten realistisch abgebildet wird. Zeitliche Verläufe von grüner Vegetation und eine kombinierte Betrachtung von trockener Vegetation und Boden lassen sowohl Rückschlüsse auf Bewirtschaftungsereignisse (Mahd, Beweidung) als auch auf Dürreperioden zu. Anteile grüner Vegetation lassen sich mit hoher Genauigkeit ableiten. Die Differenzierung zwischen trockener Vegetation und offenem Boden ist aufgrund der spektralen Ähnlichkeit allerdings mit größeren Unsicherheiten behaftet. Die Ergebnisse zeigen dahingehend, dass spektrale Derivate (z.B. Tasseled Cap Transformation) und spektral-temporale Metriken wichtige Informationen für eine verbesserte Schätzung von Anteilen trockener Vegetation und offenen Bodens enthalten.
Unsere Studie zeigt daher gleichermaßen das Potenzial aber auch die Herausforderungen von dichten Sentinel-2 Zeitreihen für ein operationelles Grünland-Monitoring in Deutschland und Europa. Insbesondere wird deutlich, dass erst durch die präzise Kenntnis von Landnutzungsmanagement einerseits und damit einhergehender spektral-temporaler Eigenschaften in Sentinel-2 Zeitreihen andererseits die präzise Charakterisierung von Grünland möglich wird.
Frantz, D., 2019. FORCE—Landsat + Sentinel-2 Analysis Ready Data and Beyond. Remote Sens. 11, 1124. https://doi.org/10.3390/rs11091124
Okujeni, A., van der Linden, S., Suess, S., Hostert, P., 2017. Ensemble Learning From Synthetically Mixed Training Data for Quantifying Urban Land Cover With Support Vector Regression. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 10, 1640–1650. https://doi.org/10.1109/JSTARS.2016.2634859
Schwieder, M., Leitão, P.J., da Cunha Bustamante, M.M., Ferreira, L.G., Rabe, A., Hostert, P., 2016. Mapping Brazilian savanna vegetation gradients with Landsat time series. Int. J. Appl. Earth Obs. Geoinformation 52, 361–370. https://doi.org/10.1016/j.jag.2016.06.019
Estimation of grasslands traits with multispectral UAS data – a comparison of different sensors and models
KIT, Campus Alpin, Deutschland
Assessment of cereals plant density using UAV multispectral imagery for high-throughput field phenotyping
Institute of Bio- and Geosciences, IBG-2Plant Sciences, Forschungszentrum Jülich GmbH, Jülich, Germany
The assessment of plant traits in agriculture is becoming increasingly important. Modern image processing methods in combination with new sensors and autonomous small aircrafts will revolutionize breeding and agricultural crop production in the coming years. The evaluation of the germination rate of cereals with an umanned aerial vehicle (UAV) is one part of this development. The seed germination has already been determined by UAVs or field bicycles, which assign a single leaf to a particular plant. This identification, however, requires a high spatial resolution, which considerably limited the practicability of the methodology. In this study, a new methodology was applied to determine the germination success of cereals, based on the calculation of fractional cover values from UAV multispectral images. It has been assumed that in early leaf development phases a higher value of break coverage indicates a larger number of plants. Based on this hypothesis, an empirical regression model with reference measurements from the field could be trained to determine the amount of germinated plants with UAVs. In this procedure it is not necessary to assign a single leaf to a particular plant. This allowed the determination of the germination success with a clearly lower spatial resolution which therefore leads to higher practicability of the developed methodology.
In the following, the factors that influence the accuracy of this empirical regression model are determined (i to v) with a performance validation of the developed methodology (vi, vii) to quantify the number of plants with the UAV technology. The primary aims of this study are to answer the following questions: (i) What is the ideal time-point for data collection related to leaf development stage; (ii) Is an RGB camera sufficient or should a multispectral camera with five spectral bands be used; (iii) What spatial resolution is required for high accuracy facilitate as well high-throughput field phenotyping; (iv) How do these above mentioned factors (i to iii) interact with each other; (v) What influence do plant characteristics have on the methodology, such as leaf arrangement, species and genotype; (vi) What is the practicability and robustness of the methodology when transferring the trained empirical regression model to another experiment; and (vii) How many reference measurements in the field are necessary to calibrate the empirical regression model user-based.
The procedure was applied to three different experiments. The first two experimental fields consist of small breeder plots (1.4 × 3 m) planted with three summer barley genotypes and three summer wheat genotypes. The genotypes were sown in four densities (150, 250, 350 and 450 plants/m²) with eight repetitions, resulting in a sample size of 96 per species. Summer wheat and summer barley differ with regard to the plant-characteristics, particularly the leaf-arrangement. The leaf position of barley was erectophil (vertical leafs predominating), while wheat had a planophile (horizontal leafs predominating) leaf position.
The experimental field 3 was a plot arrangement with 42 winter wheat genotypes (1.4 × 3m) and was used as a proof of concept (POC) to test the robustness and usability of the developed methodology. It was sown with a density of 440 seeds m−2 and three repetitions per genotype (n =126).
For the reference measurement, all individual plants within one square meter were counted at all experimental sites in the field. The edges of these square meters were labeled with sticks to mark the area of interest for UAV evaluation. The timing of the reference measurement was set to BBCH development stage 11.
Structure from motion (SfM) algorithms were used to process the individual UAV images into an orthomosaic in Agisoft Photoscan software. The vegetation index ExGR for RGB images and the NDVI for multispectral images were then calculated from the processed orthomosaic. The determined vegetation indices finally allowed the fractional cover assessment by the application of a threshold that divided the pixels into two classes, foreground (plant pixels) and background (soil pixels). The threshold determination was performed automatically by maximizing or minimizing the variance between classes based on the Otsu method [XYZ]. The highlighted areas of one square meter (Figure 1) were considered to create shape files for each experimental site and used to calculate the average fractional cover.
The data set of experimental field 1 and 2, each with a sample size of 96, was first sorted according to the independent variable (measured plant quantity) for the development of the empirical regression model. In the second step, a three-fold cross validation was done. The data set was divided into three subsets (A, B, C), each with a sample size of 32. Each subset was used once to validate the empirical regression model, while the other subsets build on the rest of the 64 data points (A / BC, B / AC, C/ AB) were used to train the empirical regression. Experimental fields 1 and 2 were overflown on four dates at two different flight altitudes (FA) each sensor (Table 1).
For the third experimental site the trained empirical regression model from experimental site 2 (Wheat, BBCH 12, NDVI, 20m FA) was used to evaluate the robustness and practicability of the developed methodology. In addition, the data set with a sample size of 126 was randomly divided 252 times into a calibration and validation data set. The total number of calibration varied between 1/2, 1/3, 1/4, 1/6, 1/8, 1/11, 1/15 and 1/25 of the data set. This made it possible to evaluate the necessary number of reference measurement in the field to enable a user-calibration of the empirical regression model adapted to the data set environment.
The performance of all models to identify the main influencing factors for predicting the germination success from calculated fractional cover values were summarized in Table 1. In the early stage of the development with only one or two unfolded leaves (BBCH 11, 12), spatial resolution had a high influence on ExGR model performance with regression error metrics above 50 plants/m² (Table 1). The spatial resolution had less influence on model performance using the NDVI for germination assessment. Still, the lower FA of 10m achieved the highest R² of 0.93 and the lowest MAE of 24 plants/m² for barley at BBCH stage 12 (Table 1).
Table 1. Average values of the trained empirical regression models for barley and wheat related to the BBCH stage, VI and flight height.
This best model performance for the vegetation index ExGR was observed at BBCH stage 13. Likewise for the multispectral sensor and high FA (NDVI, 20m) the BBCH stage 13 with unfolded third leaf was the optimal development stage for barley, which led to an R² of 0.90 and MAE below 27 plants/m².
The model performance for both vegetation indices declined with growing plant size and increased number of leafs (BBCH Stage 14, 15). Still the R² of approximately 0.85 was good, but with regression error metrics mainly over 40 plants/m² the prediction accuracy declined (Table 1).
In order to evaluate the robustness and practicability of the developed methodology, two different proceedings were applied in the recent study.
First, the empirical regression model from site 2 (Wheat, BBCH 12, NDVI, 20m FA) was transferred to predict the germination rate at a different experimental arrangement. The site was ideal for testing because with 42 different genotypes it included a large number of varieties in plant characteristics. Transferring the trained empirical model yielded to a very stable prediction of UAV estimated plant quantity with a MAE below 21 plants/m² and a R² of 0.83.
Table 2. Empirical regression models of experimental site 3 related to the total number of calibration.
Second, the size of the calibration subset was continuously decreased to evaluate the necessary number of field measurements to enable a user-based calibration of the empirical regression model. The calibration subset of 63 at the beginning of the evaluation led to the highest accuracy with a MAE below 22 plants/m² and a R² of 0.83. The reduction of the calibration subset to a sample size of 42 led to a nearly identical model performance (Table 2). In fact, although the calibration sample size was continuously decreasing to only eight reference measurements in the field, a good model performance with error metrics values below 25 plants/m² could be observed. The slope and intercept of the validation regression lines were however continuously raising with decreasing calibration sample size (Table 2). The accuracy of the developed regression model declined when the calibration sample size was below five reference measurements with an increased slope and intercept value for the validation dataset (Table 2).
The developed approach in this study facilitate high-throughput field phenotyping of cereal germination for high sowing densities up to 450 plants/m² for precision farming, crop models and breeding. The results stressed out the accuracy, transferability, practicability and robustness of the methodology. In contrast to machine learning approaches, expert guidance in feature extraction is not required. With a simple implementation the operator can choose if the trained model is applied or a used-based approach with a low number of field measurements is conducted.
Potential of Planet's Dove Constellation for Crop Type Classification using the Multi Data Approach (MDA)
AG GIS & Fernerkundung, Geographisches Institut, Universität zu Köln,
Current and future agricultural production faces the challenge to provide more food for an increasing world population with increasing nutrition demands per person under the effects of climate change. The proposed strategy to cope with those challenges is the sustainable intensification of agriculture. Its Implementation involves precise knowledge of the type and extent of the crops being grown. Satellite Remote Sensing is particularly suitable to deliver such information cost-efficiently in the form of crop type maps.
However, agricultural areas are very dynamic and their appearance in remote sensing images is determined by the crop-specific phenology and field management practices. Only at times when that look is different from the looks of other crops, can the fields crop type be identified. Hence, for faultless crop identification, single images are inadequate and frequent observations, e.g. a high temporal resolution, are necessary. Most current optical satellite remote sensors either lack the needed high temporal resolution, which is also significantly reduced by frequent cloud coverage or have an inadequate spatial resolution, which precludes mapping smaller field's crop type. The also available microwave sensors demand more sophisticated preprocessing and sometimes fail to differentiate between all of the crop types.
To make use of the advantages of miscellaneous satellite systems, and combine satellite observations with existing geodata, the multi-data approach (MDA) was developed as a framework for enhanced Land Use / Land Cover mapping (LULC). Slowly changing land cover classes, such as urban areas, are extracted from the external geodata, as those classes are usually contained in the external data with high precision. The satellite observations solely deliver the dynamic classes, such as, in this case, the annually changing crop type, which is most often missing in official geodata.
However, combining different sensors of different temporal, spatial and spectral resolutions leads to varying quality of the crop type maps. Furthermore, a very high spatial resolution is needed to distinguish all fields, which is expensive and usually not available over larger areas. Innovation to overcome such issues provides the Planet Dove constellation, which consists of approximately 140 microsatellites, each equipped with a multispectral sensor (Red, Green, Blue, NIR) providing a very high spatial resolution between 2.7 and 4.9 m, depending on the orbit and acquisition angle. The temporal resolution of the constellation is so high that at least one image of every point on earth is provided daily. However, to decrease costs, Planet uses lower grade electronics for building the Dove Satellites, potentially leading to worse measurements. Hence, the overall question is, are the spatially and temporally high-resolution images of the Dove constellation beneficial for crop type mapping and integration into the MDA framework?
Therefore, the current study presents a case study of MDA based crop classification for an agricultural region in western Germany using Planet's Dove imagery. As the area of interest (AOI) an area of about 235 km2 approximately 35 km west of cologne, near the open-pit mine of Hambach, was chosen. The region is characterized by fertile soils and intense agricultural production. A field survey to collect ground truth was carried out at the beginning of July 2019 and more than a hundred fields of eight different crop classes (Grassland, Rapeseed, Maize, Sugar Beet, Potato, Winter Wheat, Winter Barley, Pea) were mapped. As Planet imagery is quickly available after the acquisition, an up-to-date satellite image was available in the field on a mobile device and could be used as assistance during the mapping campaign.
Previous studies identified six relevant periods, hereafter called acquisition windows (AW), during the growing season, where cloud-free satellite images should be obtained to successfully discriminate the crops of this region. Therefore, cloud-free satellite acquisitions covering the whole AOI were searched using the Planet Explorer and ordered as analytic-ready with a pixel spacing of 3 m (orthorectified, multi-spectral, 16-bit).
After mosaicking the single date scenes, they were stacked and used for a supervised pixel-based classification approach, using half of the mapped fields as training and the other half as validation. The Random Forest classifier was chosen as it is a standard classification method in the field of crop type classification. The Random Forest settings were set to 150 trees maximum, and a maximum tree depth of 30. A majority filter with a 3x3 moving window was used as a post-classification enhancement. Through Error-Matrix generation with the independent validation fields, the Overall Accuracy (OA) and F-Scores of the eight crop classes could be generated.
One of the fundamental ideas of the MDA is the fusion with external geodata. Before the classification, the fusion enables not having to consider all non-crop classes, which results in less confusion with those classes and hence, in higher classification accuracy. A post-classification fusion enables showing all relevant non-crop classes alongside the mapped crops. However, for both fusions a high spatial accuracy of the satellite data is mandatory. In this study, field blocks were derived from the German Cadastral Land Register ALKIS to mask the agricultural area. Spatially highly accurate aerial images were used to evaluate the absolute spatial accuracy of the classifications.
The relative spatial accuracy of the six scenes to each other was evaluated to be very high through visual inspection. Absolute spatial accuracy was at some places one to two 3 m pixels off location compared to the accurate aerial images. A pre- and post-classification fusion with external datasets such as ALKIS seems therefore feasible and meaningful. Thereby, enhanced LULC maps, showing all relevant LULC classes alongside the annually changing crop type, can be generated. Notably, this result is reached without further manually refining the spatial accuracy of the input data.
Considering the temporal resolution of the system, it can be concluded, that it is high enough to deliver at least one cloud-free image for the chosen AOI, for all the six previously defined AWs, at least for the study season 2019. The used acquisition dates were May 15, June 2, July 4, July 23, August 26, and October 9.
The crop classification has a very high spatial final resolution of 3 m, looks appearing, and error analysis showed a high OA of 91,6 %. The only F-scores lower than 90 % were obtained for Maize (78%) and Grassland (81%). All other classes were validated with F-scores beyond 90 % (Pea and Rapeseed 91 %, Potato and Winter Wheat 93 %, Winter Barley and Sugar Beet 96 %).
Hence, it can be concluded that the very high spatial resolution, and precision, in combination with the ultra-high temporal resolution of Planet's Dove constellation were demonstrated to be beneficial for MDA-based crop type mapping in the present case study. Future studies could combine the data with images from other satellite systems or apply the approach of the current study to larger regions.