@article{DhillonDahmsKuebertFlocketal.2023, author = {Dhillon, Maninder Singh and Dahms, Thorsten and Kuebert-Flock, Carina and Rummler, Thomas and Arnault, Joel and Steffan-Dewenter, Ingolf and Ullmann, Tobias}, title = {Integrating random forest and crop modeling improves the crop yield prediction of winter wheat and oil seed rape}, series = {Frontiers in Remote Sensing}, volume = {3}, journal = {Frontiers in Remote Sensing}, issn = {2673-6187}, doi = {10.3389/frsen.2022.1010978}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-301462}, year = {2023}, abstract = {The fast and accurate yield estimates with the increasing availability and variety of global satellite products and the rapid development of new algorithms remain a goal for precision agriculture and food security. However, the consistency and reliability of suitable methodologies that provide accurate crop yield outcomes still need to be explored. The study investigates the coupling of crop modeling and machine learning (ML) to improve the yield prediction of winter wheat (WW) and oil seed rape (OSR) and provides examples for the Free State of Bavaria (70,550 km2), Germany, in 2019. The main objectives are to find whether a coupling approach [Light Use Efficiency (LUE) + Random Forest (RF)] would result in better and more accurate yield predictions compared to results provided with other models not using the LUE. Four different RF models [RF1 (input: Normalized Difference Vegetation Index (NDVI)), RF2 (input: climate variables), RF3 (input: NDVI + climate variables), RF4 (input: LUE generated biomass + climate variables)], and one semi-empiric LUE model were designed with different input requirements to find the best predictors of crop monitoring. The results indicate that the individual use of the NDVI (in RF1) and the climate variables (in RF2) could not be the most accurate, reliable, and precise solution for crop monitoring; however, their combined use (in RF3) resulted in higher accuracies. Notably, the study suggested the coupling of the LUE model variables to the RF4 model can reduce the relative root mean square error (RRMSE) from -8\% (WW) and -1.6\% (OSR) and increase the R 2 by 14.3\% (for both WW and OSR), compared to results just relying on LUE. Moreover, the research compares models yield outputs by inputting three different spatial inputs: Sentinel-2(S)-MOD13Q1 (10 m), Landsat (L)-MOD13Q1 (30 m), and MOD13Q1 (MODIS) (250 m). The S-MOD13Q1 data has relatively improved the performance of models with higher mean R 2 [0.80 (WW), 0.69 (OSR)], and lower RRMSE (\%) (9.18, 10.21) compared to L-MOD13Q1 (30 m) and MOD13Q1 (250 m). Satellite-based crop biomass, solar radiation, and temperature are found to be the most influential variables in the yield prediction of both crops.}, language = {en} } @article{KanmegneTamgaLatifiUllmannetal.2023, author = {Kanmegne Tamga, Dan and Latifi, Hooman and Ullmann, Tobias and Baumhauer, Roland and Thiel, Michael and Bayala, Jules}, title = {Modelling the spatial distribution of the classification error of remote sensing data in cocoa agroforestry systems}, series = {Agroforestry Systems}, volume = {97}, journal = {Agroforestry Systems}, number = {1}, issn = {0167-4366}, doi = {10.1007/s10457-022-00791-2}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-324139}, pages = {109-119}, year = {2023}, abstract = {Cocoa growing is one of the main activities in humid West Africa, which is mainly grown in pure stands. It is the main driver of deforestation and encroachment in protected areas. Cocoa agroforestry systems which have been promoted to mitigate deforestation, needs to be accurately delineated to support a valid monitoring system. Therefore, the aim of this research is to model the spatial distribution of uncertainties in the classification cocoa agroforestry. The study was carried out in C{\^o}te d'Ivoire, close to the Ta{\"i} National Park. The analysis followed three steps (i) image classification based on texture parameters and vegetation indices from Sentinel-1 and -2 data respectively, to train a random forest algorithm. A classified map with the associated probability maps was generated. (ii) Shannon entropy was calculated from the probability maps, to get the error maps at different thresholds (0.2, 0.3, 0.4 and 0.5). Then, (iii) the generated error maps were analysed using a Geographically Weighted Regression model to check for spatial autocorrelation. From the results, a producer accuracy (0.88) and a user's accuracy (0.91) were obtained. A small threshold value overestimates the classification error, while a larger threshold will underestimate it. The optimal value was found to be between 0.3 and 0.4. There was no evidence of spatial autocorrelation except for a smaller threshold (0.2). The approach differentiated cocoa from other landcover and detected encroachment in forest. Even though some information was lost in the process, the method is effective for mapping cocoa plantations in C{\^o}te d'Ivoire.}, language = {en} } @article{PhilippDietzUllmannetal.2023, author = {Philipp, Marius and Dietz, Andreas and Ullmann, Tobias and Kuenzer, Claudia}, title = {A circum-Arctic monitoring framework for quantifying annual erosion rates of permafrost coasts}, series = {Remote Sensing}, volume = {15}, journal = {Remote Sensing}, number = {3}, issn = {2072-4292}, doi = {10.3390/rs15030818}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-304447}, year = {2023}, abstract = {This study demonstrates a circum-Arctic monitoring framework for quantifying annual change of permafrost-affected coasts at a spatial resolution of 10 m. Frequent cloud coverage and challenging lighting conditions, including polar night, limit the usability of optical data in Arctic regions. For this reason, Synthetic Aperture RADAR (SAR) data in the form of annual median and standard deviation (sd) Sentinel-1 (S1) backscatter images covering the months June-September for the years 2017-2021 were computed. Annual composites for the year 2020 were hereby utilized as input for the generation of a high-quality coastline product via a Deep Learning (DL) workflow, covering 161,600 km of the Arctic coastline. The previously computed annual S1 composites for the years 2017 and 2021 were employed as input data for the Change Vector Analysis (CVA)-based coastal change investigation. The generated DL coastline product served hereby as a reference. Maximum erosion rates of up to 67 m per year could be observed based on 400 m coastline segments. Overall highest average annual erosion can be reported for the United States (Alaska) with 0.75 m per year, followed by Russia with 0.62 m per year. Out of all seas covered in this study, the Beaufort Sea featured the overall strongest average annual coastal erosion of 1.12 m. Several quality layers are provided for both the DL coastline product and the CVA-based coastal change analysis to assess the applicability and accuracy of the output products. The predicted coastal change rates show good agreement with findings published in previous literature. The proposed methods and data may act as a valuable tool for future analysis of permafrost loss and carbon emissions in Arctic coastal environments.}, language = {en} } @article{DhillonKuebertFlockDahmsetal.2023, author = {Dhillon, Maninder Singh and K{\"u}bert-Flock, Carina and Dahms, Thorsten and Rummler, Thomas and Arnault, Joel and Steffan-Dewenter, Ingolf and Ullmann, Tobias}, title = {Evaluation of MODIS, Landsat 8 and Sentinel-2 data for accurate crop yield predictions: a case study using STARFM NDVI in Bavaria, Germany}, series = {Remote Sensing}, volume = {15}, journal = {Remote Sensing}, number = {7}, issn = {2072-4292}, doi = {10.3390/rs15071830}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-311132}, year = {2023}, abstract = {The increasing availability and variety of global satellite products and the rapid development of new algorithms has provided great potential to generate a new level of data with different spatial, temporal, and spectral resolutions. However, the ability of these synthetic spatiotemporal datasets to accurately map and monitor our planet on a field or regional scale remains underexplored. This study aimed to support future research efforts in estimating crop yields by identifying the optimal spatial (10 m, 30 m, or 250 m) and temporal (8 or 16 days) resolutions on a regional scale. The current study explored and discussed the suitability of four different synthetic (Landsat (L)-MOD13Q1 (30 m, 8 and 16 days) and Sentinel-2 (S)-MOD13Q1 (10 m, 8 and 16 days)) and two real (MOD13Q1 (250 m, 8 and 16 days)) NDVI products combined separately to two widely used crop growth models (CGMs) (World Food Studies (WOFOST), and the semi-empiric Light Use Efficiency approach (LUE)) for winter wheat (WW) and oil seed rape (OSR) yield forecasts in Bavaria (70,550 km\(^2\)) for the year 2019. For WW and OSR, the synthetic products' high spatial and temporal resolution resulted in higher yield accuracies using LUE and WOFOST. The observations of high temporal resolution (8-day) products of both S-MOD13Q1 and L-MOD13Q1 played a significant role in accurately measuring the yield of WW and OSR. For example, L- and S-MOD13Q1 resulted in an R\(^2\) = 0.82 and 0.85, RMSE = 5.46 and 5.01 dt/ha for WW, R\(^2\) = 0.89 and 0.82, and RMSE = 2.23 and 2.11 dt/ha for OSR using the LUE model, respectively. Similarly, for the 8- and 16-day products, the simple LUE model (R\(^2\) = 0.77 and relative RMSE (RRMSE) = 8.17\%) required fewer input parameters to simulate crop yield and was highly accurate, reliable, and more precise than the complex WOFOST model (R\(^2\) = 0.66 and RRMSE = 11.35\%) with higher input parameters. Conclusively, both S-MOD13Q1 and L-MOD13Q1, in combination with LUE, were more prominent for predicting crop yields on a regional scale than the 16-day products; however, L-MOD13Q1 was advantageous for generating and exploring the long-term yield time series due to the availability of Landsat data since 1982, with a maximum resolution of 30 m. In addition, this study recommended the further use of its findings for implementing and validating the long-term crop yield time series in different regions of the world.}, language = {en} } @article{DhillonDahmsKuebertFlocketal.2023, author = {Dhillon, Maninder Singh and Dahms, Thorsten and K{\"u}bert-Flock, Carina and Liepa, Adomas and Rummler, Thomas and Arnault, Joel and Steffan-Dewenter, Ingolf and Ullmann, Tobias}, title = {Impact of STARFM on crop yield predictions: fusing MODIS with Landsat 5, 7, and 8 NDVIs in Bavaria Germany}, series = {Remote Sensing}, volume = {15}, journal = {Remote Sensing}, number = {6}, issn = {2072-4292}, doi = {10.3390/rs15061651}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-311092}, year = {2023}, abstract = {Rapid and accurate yield estimates at both field and regional levels remain the goal of sustainable agriculture and food security. Hereby, the identification of consistent and reliable methodologies providing accurate yield predictions is one of the hot topics in agricultural research. This study investigated the relationship of spatiotemporal fusion modelling using STRAFM on crop yield prediction for winter wheat (WW) and oil-seed rape (OSR) using a semi-empirical light use efficiency (LUE) model for the Free State of Bavaria (70,550 km\(^2\)), Germany, from 2001 to 2019. A synthetic normalised difference vegetation index (NDVI) time series was generated and validated by fusing the high spatial resolution (30 m, 16 days) Landsat 5 Thematic Mapper (TM) (2001 to 2012), Landsat 7 Enhanced Thematic Mapper Plus (ETM+) (2012), and Landsat 8 Operational Land Imager (OLI) (2013 to 2019) with the coarse resolution of MOD13Q1 (250 m, 16 days) from 2001 to 2019. Except for some temporal periods (i.e., 2001, 2002, and 2012), the study obtained an R\(^2\) of more than 0.65 and a RMSE of less than 0.11, which proves that the Landsat 8 OLI fused products are of higher accuracy than the Landsat 5 TM products. Moreover, the accuracies of the NDVI fusion data have been found to correlate with the total number of available Landsat scenes every year (N), with a correlation coefficient (R) of +0.83 (between R\(^2\) of yearly synthetic NDVIs and N) and -0.84 (between RMSEs and N). For crop yield prediction, the synthetic NDVI time series and climate elements (such as minimum temperature, maximum temperature, relative humidity, evaporation, transpiration, and solar radiation) are inputted to the LUE model, resulting in an average R\(^2\) of 0.75 (WW) and 0.73 (OSR), and RMSEs of 4.33 dt/ha and 2.19 dt/ha. The yield prediction results prove the consistency and stability of the LUE model for yield estimation. Using the LUE model, accurate crop yield predictions were obtained for WW (R\(^2\) = 0.88) and OSR (R\(^2\) = 0.74). Lastly, the study observed a high positive correlation of R = 0.81 and R = 0.77 between the yearly R\(^2\) of synthetic accuracy and modelled yield accuracy for WW and OSR, respectively.}, language = {en} } @article{ReinermannAsamGessneretal.2023, author = {Reinermann, Sophie and Asam, Sarah and Gessner, Ursula and Ullmann, Tobias and Kuenzer, Claudia}, title = {Multi-annual grassland mowing dynamics in Germany}, series = {Frontiers in Environmental Science}, volume = {11}, journal = {Frontiers in Environmental Science}, issn = {2296-665X}, doi = {10.3389/fenvs.2023.1040551}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-320700}, year = {2023}, abstract = {Introduction: Grasslands cover one third of the agricultural area in Germany and are mainly used for fodder production. However, grasslands fulfill many other ecosystem functions, like carbon storage, water filtration and the provision of habitats. In Germany, grasslands are mown and/or grazed multiple times during the year. The type and timing of management activities and the use intensity vary strongly, however co-determine grassland functions. Large-scale spatial information on grassland activities and use intensity in Germany is limited and not openly provided. In addition, the cause for patterns of varying mowing intensity are usually not known on a spatial scale as data on the incentives of farmers behind grassland management decisions is not available. Methods: We applied an algorithm based on a thresholding approach utilizing Sentinel-2 time series to detect grassland mowing events to investigate mowing dynamics in Germany in 2018-2021. The detected mowing events were validated with an independent dataset based on the examination of public webcam images. We analyzed spatial and temporal patterns of the mowing dynamics and relationships to climatic, topographic, soil or socio-political conditions. Results: We found that most intensively used grasslands can be found in southern/south-eastern Germany, followed by areas in northern Germany. This pattern stays the same among the investigated years, but we found variations on smaller scales. The mowing event detection shows higher accuracies in 2019 and 2020 (F1 = 0.64 and 0.63) compared to 2018 and 2021 (F1 = 0.52 and 0.50). We found a significant but weak (R2 of 0-0.13) relationship for a spatial correlation of mowing frequency and climate as well as topographic variables for the grassland areas in Germany. Further results indicate a clear value range of topographic and climatic conditions, characteristic for intensive grassland use. Extensive grassland use takes place everywhere in Germany and on the entire spectrum of topographic and climatic conditions in Germany. Natura 2000 grasslands are used less intensive but this pattern is not consistent among all sites. Discussion: Our findings on mowing dynamics and relationships to abiotic and socio-political conditions in Germany reveal important aspects of grassland management, including incentives of farmers.}, language = {en} }