CropSupport application facilitates citizen engagement in crowdsourcing of agricultural field level data. The tool was developed within a European H2020 project – LandSense, which aims to demonstrate opportunities for innovative strategies for citizen involvement in collection of high-quality in-situ information on land use and land cover (LULC). This poster presents the workflow of the CropSupport application and overviews its first testing and validation activities in an operational environment.
The CropSupport tool was developed in 2017 as a web and mobile based application. It has been primarily (but not exclusively) designed for farmer community to collect data related to crop type and farm management activities – such as seeding, tillage and harvesting date. Besides being a tool for in-situ data collection and sharing, the CropSupport application offers several added value services to its users – such as NDVI maps, parcel-based weather forecast, and a farm activity diary. This was achieved by conflating crowdsourced inputs with Earth Observation data (i.e. agriculture related remote sensing indices from Sentinel 2 imagery). Moreover, the application offers a management service for measuring and monitoring of the users’ contributions via the CropSupport web-based Administrator Panel. The management service allows monitoring a particular user’s contribution/activity through a dashboard, such as number of parcels, images, and data related to farm activities (e.g. dates of seeding, fertilization, and irrigation.). The dashboard also provides specific information which helps in assessing the quality of a user data contribution - number of images connected with the parcels, number of images without parcels and parcels without activities. Besides monitoring of user performance, the administrator panel allows for checking validity and reliability of collected data. In particular, the service enables an administrator to compare the content of image with the user’s claim. In this way, misinterpretation of certain land use can be minimized, as inappropriate images will be deleted, and a generic message will be sent to the user informing and explaining to him/her why the image was deleted.
The CropSupport application had been tested and validated in operational environment between November 2017 and September 2018, when a group of 60 high school and university students were engaged into an agricultural crowdsourcing campaign. The preliminary results indicate that CropSupport application shows a promising potential for scalability and replicability to other geographical areas.

Area A - Board 1

LandSense is a European H2020 project that explores opportunities for innovative citizen engagement in collection of high-quality in-situ information on land use and land cover (LULC). By focusing on urban, agricultural and forest land, LandSense employs different strategies to involve citizens in the LULC data collection through three demonstration cases. The process of the data collection for each of the cases is facilitated by use of portable devices and specially developed web applications. As such, LandSense represents a showcase project that promotes concept of citizen science as it encourages members of the general public to become a part of a collaborative project with professional scientists.
This poster aims to overview and discuss the LandSense’s agricultural crowdsourcing campaign in order to present a novel approach for a field level data collection. The campaign was implemented in Serbia between November 2017 and September 2018, where over 60 students from two agricultural high schools and a group of master students from the University of Novi Sad were involved in collection of crop data and information related to farm management activities. They used CropSupport application to collect geotagged images of crops, to delineate the crop parcels, and to record information on farm management activities (e.g. seeding, tillage and harvesting date).
Our approach to motivate the students to be active data collectors was to blend citizen science with gamification. Therefore, the student teams were formed, and they competed among each other to provide as much crowdsourced LULC data as possible. Also, the engaging information package was created, where the goals and values of the pilot activities were clearly identified and communicated. In addition, the student feedbacks related to the campaign activities were collected during the several workshops, and then synthesized and analysed in order to guide iterative improvement of the CropSupport application and organisation of the upcoming crowdsourcing campaign, scheduled for 2019. As a reward for their activities, all students received participation certificates, T-shirts and USB memory sticks.
The campaign is regarded as successful as there had been collected over 1800 images and delineated crop parcels of agricultural fields as well as over 300 datasets of specific farm activities. However, the quality of collected data has to be carefully assessed as the participants were involved for the first time in crowdsourcing activities, while, as well, the CropSupport application had been tested for the first time in a production environment.
The most common barriers encountered during the campaign were related to poor performance of some of the students’ smartphones (e.g. GPS device out of function, lack of memory space, etc.). On the other hand, majority of the engaged master students faced the problem related to lack of agricultural knowledge. Namely, in contrast to the agricultural high-school students that were able to easily differentiate crops, the master students had difficulties to do the same. In order to overcome this barrier, a detailed presentation about the most common crop types in the region along with printed manual were prepared and disseminated to students. Finally, consistent engagement of the students to collect data through the entirety of the campaign was necessary as the students were not used to participate in similar activities before.

Area A - Board 2

Early spring each year, the bright yellow flowers of rapeseed appear across the European landscape. Rapeseed yield potential is strongly dependent on conditions during flowering as this influences pod formation and the number of seeds. Better information on flowering time can thus improve yield forecasts. The striking yellow also provides an excellent opportunity to combine information from different sources to pin down the temporal development and spatial pattern of rapeseed flowering. Data is combined from Sentinel-2 observations, crowd-sourced and third-party collected street-level imagery, field observations, and a dedicated Facebook/Twitter #YellowFlowersEU social media campaign. Integrating such novel data sources, the high-resolution satellite observations with short revisit cycles, and increasingly open authoritative data, can transform crop monitoring. Issues related to timeliness, coherency, and complementarity of these different data sources is investigated.

The data was collected during spring 2018, during a total of 2 months. In any given location, the flowering time of rapeseed can last from about 3 to 4 weeks. Over 1000 Tweets were sent using #YellowFlowersEU. The Tweets harvested using the #YellowFlowersEU tag were complemented by opportunistically harvesting Tweets mentioning ‘koolzaad’ (Dutch for rapeseed), “colza” (French), “raps” (German) and “oil seed rape” (OSR). This resulted in >60 observations of fields with flowering rapeseed in 12 countries, dated from DOY 100 to 150 - in respectively the south of France and Lithuania. In most cases, geo-location was approximated on the basis of meta-information. A total of six dedicated one-day surveys were carried out along road transects in southern and northern Germany. High quality close-up photographs and detailed observations on flowering stage were made at more than 200 fixed parcel locations. For the complete transects, >200.000 street-level images were collected at 1 frame per second with side-looking cameras attached to the roof of a car covering >2000 parcels with rapeseed. We describe post-processing routines to convert the survey results into a spatial database which can then be easily integrated with other data sets for intercomparison.

Our primary use of the images taken during survey is to match them with Sentinel-1 and Sentinel-2 observations, and novel crop mapping initiatives like onesoil.ai. Due to exceptionally good weather conditions, a large selection of suitable Sentinel-2 imagery is available along the surveying route during the rapeseed flowering period. Rapeseed is one of the easiest separable crop types in Sentinel-1 time series as well. We discuss possibilities to use the combined data for regional rapeseed area estimates and phenology stage.

Area A - Board 3

The completion of the operational ramp-up of the high resolution Copernicus Sentinel-1 (S1) and -2 (S2) since July 2017 has led to a consistent, continuous and rich streams of high quality, near-instantaneous Earth Observation information. The unprecedented temporal density of both the dual polarization C-band microwave backscattering signal derived from S1 and the multi-spectral visual, near- and shortwave infrared S2 sensors provide unmatched radiometric and spatial detail on landscape dynamics, in particular in fast changing agricultural terrain. The efficient uptake of these data streams, however, is challenging. This challenge is not only from a data processing perspective, but primarily due to a lack of matching high quality field observations that support the validation of the signals. Therefore, this may limit the effective uptake of dense EO time series in inversion models and quantitative methods underlying advanced applications in the agricultural user domain. In effect, the upscaling of EO data streams is not matched by a similar upscaling in efficient surveying techniques that capture the dynamics of the soil-canopy interface underlying the signal variation of the EO sensors.

An effort to address this shortcoming is the use of guided street-level imagery. The use of, opportunistic, street-level imagery is well-known for geo-visualization approaches (e.g. Google Street View, Mapillary). Deriving quantitative information from such imagery, e.g. for characterization of crops, is less developed. The idea to “guide” street-level imagery acquisition by integrating existing geospatial data (e.g. for routing) and EO data (e.g. to visit pre-marked anomalies) is relatively new [1], but demonstrating a factor 7-8 in increased data collection efficiency compared to conventional surveys. Based on the latter ideas, a multi-date survey was designed to derive crop type and phenology stage. The field campaign was carried out in Flevoland (NL) on 8 dates in the growing season (March-October) in 2018, along a fixed 200 km route. This route was determined on the basis of the agricultural parcel distribution in 2017 to ensure statistical meaningful sample density for the main crop types in the area. The route predominantly covers inner roads to maintain a maximum driving speed of 50km/h and ensure clear imagery.

Street-level imagery with 2 side-looking SONY HDR-AS300R cameras was collected at 1 frame-per-second along the entire route. Still camera photography was collected at 110 stops at which crop type and phenology information (adapted BBCH scale) were collected for 10 crop types, resulting in more than 210 sample points. For most survey dates, the latest S2 composites were used on a tablet with the PDFmaps application, to guide to fixed locations and register crop type and phenology stage including special observations (e.g. on cultivation measures, crop treatment).

The street-level imagery includes over 500,000 pictures. We have developed a set of routines to rapidly process the data sets into consistent geospatial sets, handle data redundancy, for cross-analysis against parcel data sets, selection of time series of S1 and S2, geolocate still camera picture, etc. We will also demonstrate the use of the data set for inspection of very high resolution imagery from the Planet and TripleSat sensors that is made available under the Netherlands Space Office “satellite data portal” initiative.

The data set’s prime use will be for the training and testing of computer vision techniques to automatically identify crop type and determine phenology stage [2]. We discuss further improvements in collection strategies and extended use in relevant agricultural use contexts (e.g. CAP checks by monitoring, crop production statistical sampling, crop growth anomaly detection).

[1] d’Andrimont, R.; Lemoine, G.; van der Velde, M. Targeted Grassland Monitoring at Parcel Level Using Sentinels, Street-Level Images and Field Observations. Remote Sens. 2018, 10, 1300.

[2] d’Andrimont, R., Lemoine, G., Fasbender, D., Iordanov, M., de Vries, M., van der Voet, P. and van der Velde, M., Monitoring crop phenology with street-level imagery using computer vision: a study case in the Netherlands, Abstract submitted to the Living Planet Symposium, 2019

Area A - Board 4

Recent European Union (EU) legislation for the Common Agricultural Policy (CAP) introduces the concept of checks by monitoring, which would require the development of machine learning approaches for country-wide, full season use of the EU's Copernicus Sentinel time series at agricultural parcel level. The excellent consistency of dense Sentinel-1 time series over the EU make these a prime candidate to generate gap-free, consistent feature vectors for the full set of parcel declarations for use in machine learning.
Our machine learning set up combines the use of Google Earth Engine, for which we extract weekly averaged dual polarized (VV, VH) backscattering signatures, that constitute the feature vectors for machine learning runs. For the latter, we use tensorflow with convolutional neural network models. The successful use of this technique to identify outliers has been demonstrated with open-access parcel data sets in the Netherlands, Belgium and Denmark for the 2017 growing season [1], with overall outlier rates between 4-6% or, hence, an overall accuracy of the machine learning results of 94-96%, for those crops that cover approximately 95% of the declared parcel area.
We build on this by extending machine learning via transfer learning for (1) classifying the Netherlands 2018 parcel set, with trained models derived from the 2017 season and (2) geographically constraining machine learning models to one region and applying the model in another region. Both approaches are compared to the machine learning methods in which the full parcel set of the same year is used for both the training and testing.
The multi-annual approach is particularly useful, as it allows further fine-tuning by removing confirmed outliers from the reference year results to create an improved trained model. The thus “cleaned” model can than be applied to obtain relative early season results for crops that are phenologically distinct at that stage. The transfer learning results also provide key indicators on the general applicability of the methods and the related needs for co-temporal, locally trained data sets. We discuss the relevant factors contributing to differences in the method results. This is particularly interesting for 2018, which has been an anomalous growing season in North-Western Europe due to extensive drought, which is reflected in the time series for specific crop types (and, therefore, contributes to overall confusion in the transfer learning results).

Machine learning results are exclusively based on the use of Sentinel-1. An important aspect of the CAP checks by monitoring approach, however, is to ensure high plausibility of the (in)correctness of the declared parcel use, which is captured in the “marker concept”. Marker refers to a particular observable quality or behavior of a given land phenomenon present on the ground. Marker instances are derived from specific sensor signals (e.g. Sentinel-2, Sentinel-1 coherence, other sensor data) and can be of a spatial or temporal nature (or a combination thereof), but do not need to be continuous. The design and instantiation of the marker properties and associated parameters is based on the expected signal behavior linked to the relevant phenomena that confirm the presence or absence of the declared parcel use (the scenario). We illustrate this concept by integrating marker analysis for further categorization of the outliers identified at the machine learning stage. We will discuss practical use issues and how the approach extends beyond crop type delineation to the determination of crop management practices.


[1] Lemoine, G., Devos, W., Milenov, P. and d’Andrimont, R., 2019, Machine Learning for Crop Type Identification Using County-wide, Consistent Sentinel-1 Time Series, Abstract submitted to the “Big Data from Space” Conference, Munich, 19-21 February 2019

Area A - Board 5

New approaches to collect in-situ data are needed to complement the high spatial (10 m) and temporal (5 d) resolution of Copernicus Sentinel satellite observations. Making sense of Sentinel observations requires high quality and timely in-situ data for training and validation. Classical ground truth collection is expensive, lacks scale, fails to exploit opportunities for automation, and is prone to sampling error. Here we evaluate the potential contribution of opportunistically exploiting crowdsourced street-level imagery to collect massive high-quality in-situ data in the context of crop monitoring. This study assesses this potential by answering two questions: (1) what is the spatial availability of these images across the European Union (EU), and (2) can these images be transformed to useful data? To answer the first question, we evaluated the EU availability of street-level images on Mapillary—the largest open-access platform for such images—against the Land Use and land Cover Area frame Survey (LUCAS) 2018, a systematic surveyed sampling of 337,031 points. For 37.78% of the LUCAS points a crowdsourced image is available within a 2 km buffer, with a mean distance of 816.11 m. We estimate that 9.44% of the EU territory has a crowdsourced image within 300 m from a LUCAS point, illustrating the huge potential of crowdsourcing as a complementary sampling tool. After artificial and built up (63.14%), and inland water (43.67%) land cover classes, arable land has the highest availability at 40.78%. To answer the second question, we focus on identifying crops at parcel level using all 13.6 million Mapillary images collected in the Netherlands. Only 1.9% of the contributors generated 75.15% of the images. A procedure was developed to select and harvest the pictures potentially best suited to identify crops using the geometries of 785,710 Dutch parcels and the pictures’ meta-data such as camera orientation and focal length. Availability of crowdsourced imagery looking at parcels was assessed for eight different crop groups with the 2017 parcel level declarations. Parcel revisits during the growing season allowed to track crop growth. Examples illustrate the capacity to recognize crops and their phenological development on crowdsourced street-level imagery. Consecutive images taken during the same capture track allow selecting the image with the best unobstructed view. In the future, dedicated crop capture tasks can improve image quality and expand coverage in rural areas.

Area A - Board 7

New satellite sensors and cloud computing now provide us with the data and processing power required to produce highly accurate systems to monitor our planet's surface. As an example, today's forest monitoring systems can provide global forest changes with 30 m resolution. Novel statistical learning approaches using satellite image time series have shown very high accuracy in crop mapping. These algorithms require an enormous amount of training data, including consistent spatial and temporal referencing, i.e., the location and the period of the crop occurrence. Ideally, these training datasets are produced using in-situ ground truth collections, which are usually costly and challenging to scale up to large areas. In this work, we present a new method to collect crop reference data using Google Street View images. The approach is scalable and provides consistent reference data for training algorithms for satellite time series classification. We have developed a web platform that provides an interactive environment where the user can navigate through Street View images to identify and collect crop reference samples. As additional information, we provide crop calendars with the most probable crops for a given period of the year and a list of crop types consistent with the indicative crop classification for the agricultural census from the Food and Agriculture Organization (FAO). Combining geolocation and the dates of the labelled crop reference collected using our tool, we then derive the phenological cycles of the crop types from satellite image time series, which feed into the statistical learning algorithms. To evaluate the potential of our tool, we collected a crop reference data set for Brazil in the agricultural frontiers of the Amazon and Cerrado biomes. We tested the separability of the samples using clustering algorithms on MODIS vegetation index time series (MOD13Q1). We also evaluated statistical learning algorithms using our sample collection for training and validation and an independent in-situ data set to test the classification accuracy. Our results show that the use of Street View images has potential to scale up crop reference data collection with a low cost, contributing to overcoming the lack of training data for statistical learning algorithms and to improving the quality of crop maps.

Area A - Board 8

Sustainable agricultural intensification requires the provision of adequate and timely information to farmers on the ground in order to provide them with the necessary tools for them to make the best decisions and optimize crop production. Optimization of field-crop production leads to several positive outcomes for the farmers and the environment such as reduction of costs and proper use of fertilizers, avoiding contamination of aquifers and increasing yields. Large amounts of information and research oriented to crop performance optimization exist and are available to farmers over the internet. Some other sources of information that are also useful for improving cropping results, such as benchmarking historical data, potential of the land for increased production and real-time crop status indicators tend to be restricted or proprietary. To contribute to the solution of access to information, especially for small farmers, a joint cooperation between the International Maize and Wheat Improvement Center (CIMMYT) in Mexico and the International Institute for Applied Systems Analysis (IIASA) was formed to produce the AgroTutor mobile application. AgroTutor is a freely available app, aimed to provide targeted and timely information for farmers across Mexico in order to promote sustainable intensification, supporting CIMMYT’s existing agricultural extension network. Once farmers download the app, they can register their plots and the crops grown in them. The app allows them to geo-locate the plot and delineate it over very high resolution satellite imagery. Once some basic information, such as planting date and cultivar characteristics are entered, the app can provide specific recommendations, such as anonymous historical costs, income, yield and profits from farms around the selected plot, which is possible due to the vast amount of information that CIMMYT has been collecting over the years. Additionally, historical potential yield from the area is also displayed, based on the Environmental Policy Integrated Climate (EPIC) model, which is run at IIASA. Historical, forecasted and current weather information is also displayed in the app, with emphasis on agriculturally relevant data such as evapotranspiration and growing degree days. Additionally, using weather information, planting date and cultivar characteristics, the calendar can display windows of opportunity, i.e. best suggested times to execute agricultural activities, e.g. fertilization. Furthermore, farmers can register all activities performed in the parcel, with those dates also appearing in the calendar, allowing farmers to keep track of their plots. The information entered can also feed crop models, crowdsourcing, in this way, usually hard to obtain in-situ data. Currently, the app is being continually developed and improved, with additional features being added. This includes a price forecasting module, where prices for commodities such as maize and wheat are forecasted for the following 12 months. These forecasts are based on the selection amongst an array of economic models that have been run at IIASA and partner institutions in Vienna. Also, a high performance data driven recommendations module will be included with recommendations such as which varieties perform best for the area where the parcel is located. The aim of AgroTutor is to become a decision support tool allowing farmers, especially smallholders, to have access to usually restricted, advanced crop benchmarking information, where the information is presented in a simple, understandable way. Farmers can then make the best decisions for their crops and, through improved yields and less financial and environmental costs, achieve improved livelihoods.

Area A - Board 10

In order to modernize and simplify the Common Agricultural Policy (CAP), the next CAP reform has introduced the concept of CAP monitoring to simplify the management of subsidies and the control of farmers’ parcel declarations for area support measures. The use of high resolution Sentinels satellite images combined with the availability of high-quality Land Parcel Identification System (LPIS) is driving the move towards at-parcel agricultural monitoring. This monitoring system to be used will include the monitoring of different agricultural works carried out during the year, such as ploughing and harvesting, in addition of the crop type identification.
In situ information is important in the study of agricultural crops by remote sensing for calibration and for validation of different models or algorithms, such as classification or change detection algorithms. The quick and efficient acquisition of data from the fields, recordable in numeric databases and big enough to have consistent training and validation samples, is an issue to tackle in order to develop operational and real time methods of agricultural monitoring.
In the present study, an on-board dash camera embedded in a car is used to collect field information with a high temporal frequency in order to monitor field works, especially the moment of the destruction of intercrops. The intercrops are important and are strictly regulated by the CAP because of the negative impact on the environment of bare soils during long periods, such as erosion and nitrate percolation. The detection of intercrop destruction dates is one of the useful information used to check the respect of legal obligations in terms of compulsory period to maintain these crops.
The test area is located in Walloonia, Southern part of Belgium, covering about 700 square kilometres. A pre-defined itinerary was systematically done about twice a week during one month and a half in order to identify when field works happen. For each travel, an overall video and geo-tagged photos corresponding to fields where works were carried out were recorded. Firstly, this in situ dataset is used to calibrate and validate a system based to detect intercrop destruction dates. In the second phase of study the proposed method is extended to identification of different field works as harvest and grass mowing at the field level.
Both NDVI evolution derived from Sentinel-2 optical images and SAR backscattering and coherence evolution derived from Sentinel-1 images at the parcel level are used in order to detect agricultural field works. After calibration and validation over the considered test area, the results were extended to all Wallonia (16.5 thousand square kilometres).

Area A - Board 11

During the past years, suitable in-situ field level information gained an increasing importance for satellite earth observation frameworks. Particularly classification results, e.g. using machine learning algorithms, strongly depend on the availability of high quality ground truthing information. However, besides the identified shortcoming in representative significance, scale, and/or thematic adequacy (Foody, 2002; Olofsson et al. 2013), in-situ field data collections are often only insufficiently planned, if at all (SIGMA, 2017). Conversely, there have been recently major investigations to address these shortcomings.
Adequate sampling design tools have been developed within the FP-7 project SIGMA (Haub, Kleinewillinghöfer & Gilliams, 2017). Main focus was to overcome uncertainties in planning and design of field data collection for EO stakeholders. Besides the sampling, a better fit between the observation dimension of field samples and the remote sensing products was addressed. In comparing the data models of the LUCAS point sampling scheme by EUROSTAT and the current Copernicus high resolution layers, the European Environmental Agency elaborated in 2015, how to ensure high synergies between Copernicus and LUCAS (Buck et al, 2015). This had been kept up by Eurostat, together with the JRC with a specific focus to increase the suitability of the LUCAS observation dimension. An extra “COPERNICUS” module within the LUCAS 2018 area frame survey has been introduced (Eurostat, 2018).
The paper will describe the above investigations, how these have been implemented in recent field campaigns by means of dedicated GeoIT solutions, such as integrated plug-ins or mobile data collection apps, and it will assess the impact of these findings for major recent EO mapping activities by the authors, such as for instance:
• the Biosphere reserve monitoring by the people in Benin and Togo (2017),
• agricultural monitoring projects, such as the grassland monitoring in Germany (2017/2018),
• Crop productions estimation systems in Malawi (Haub et al., 2016).

References:

Buck, O., Haub, C., Woditsch, S., Lindemann, D., Kleinewillinghöfer, L., Hazeu, G., Kosztra, B., Kleeschulte, S., Arnold, S., Hölzl, M. (2015) Task 1.9 - Analysis of the LUCAS nomenclature and proposal for adaptation of the nomenclature in view of its use by the Copernicus land monitoring services (Service contract report No. 3436/B2015/R0-GIO/EEA.56166). European Environment Agency (EEA), Copenhagen.

Eurostat (2018) How LUCAS data can be used. https://ec.europa.eu/eurostat/statistics-explained/index.php?title=LUCAS_-_Land_use_and_land_cover_survey [Online: accessed on 11.11.2018].

Haub, C., Giovacchini, A., Namaona, A., Mwanaleza, E.; Remotti, D., Kleinewillinghöfer, L. (2016): Malawian Point Frame Area Survey – How GMFS Services Evolve Towards Operationability. Oral presentation in the session “Agriculture - Global” of the European Space Agency (ESA) Living Planet Symposium 2016, 09.-13.05.2016, Prague, Czech Republic. http://lps16.esa.int/page_session25.php#2589p [Online: accessed on 11.11.2018].

Haub, C., Kleinewillinghöfer, L. & Gilliams, S. (2017) New understanding of thematic Land Cover validation on basis of the SIGMA validation protocol and QGIS Plug-in. Oral presentation in the session “Approaches and Practices for Land Products Validation” of the European Space Agency (ESA) Land Product Validation and Evolution 2018, 27.02-01.03.2018, Frascati, Italy. https://earth.esa.int/web/sppa/meetings-workshops/lpve-2018/programme [Online: accessed on 11.11.2018].

SIGMA (2017) Regional Training Workshop on Geospatial Technologies for Agricultural Statistics & Monitoring. Nairobi 21.-24.03.2017 http://www.geoglam-sigma.info/Pages/news.aspx?CustomID=66 [Online: accessed on 11.11.2018].

Foody, G.M., 2002. Status of land cover classification accuracy assessment. Remote sensing of environment 80, 185–201.

Olofsson, P., Foody, G.M., Stehman, S.V., Woodcock, C.E. (2013) ‘Making better use of accuracy data in land change studies: Estimating accuracy and area and quantifying uncertainty using stratified estimation’. Remote Sensing of Environment 129, 122–131. doi:10.1016/j.rse.2012.10.031

Area A - Board 13

Withing the framework of ESA Express Procurement entitled "Land Products Validation and Characterisation in support to Proba-V, S-2 and S-3 missions" we developed system based on in-situ and remote sensing data for operational crop yield modelling. Field measurements were conducted at agriculture study site that belongs to Joint Experiment for Crop Assessment and Monitoring (JECAM).

We used multi-scale approach, modelling crop yields from local crop fields to Wielkopolska Province (Nuts-2 administrative division) applying Sentinel-2 and Proba-V satellite images (jpg on workflow attached). The area of the cultivated arable land in the province is around 1 779 thousand hectares. The annual crops are as follows: 4 431 thousand tons of cereals (15 % of all-Poland crops), 1 106 thousand tons of potatoes (12 %) and 642,000 tons of vegetables (14 %).Wielkopolska is an agricultural region the crop rotation systems are practiced. Crop calendar for winter cereals begins in September and lasts to July, while for spring/summer crops calendar stars in April and ends in October.

In-situ measurements LAI, APAR, hyperspectralresponses 350 – 2500 nm with FieldSpec4, Chlorophyll fluorescence with OS5p+, OPTI-SCIENCES and soil moisture with TRIME-PICO64 systems were performed. The measurements were performed every 2 weeks during the groiwing season for winter wheat, corn, rape and grasllands. The in-situ data are stored in the database.

For modelling crop yields at local scale we performed crop classifiation for 2018 using combination of Sentinel-1 and Sentinel-2 images registered from April to September. Total accuracy of crop classification was achieved 83%. Next for The PROtotype Biomass and Evaporation (PROBE) (Maas et al., 1992, Dabrowska-Zielinska et al. 2001 ) meteorological observations (temperature, precipitation, solar radiance), LAI in-situ and satellite-modelled and evapotranspiration observations were taken into account. The results on winter wheat crop yield modelling indicate very high precision of total biomass forecasting, at 94% at local scale.

The model was re-calibrated incorporating Proba-V 100m and 1000m in order to forecast crop yields at regional scale. The results of crop yield modelling at regional scale for government statistical agency i.e. Central Statistical Office are still undergoing testing.
The aim of the our project is integration of in-situ and remote sensing data with developed crop yield method to establish service for farmers and statistical office offering application and platform enabling on-line accessibility of the final products to the end-users, adjustable to user’s needs at local, regional level.

Area A - Board 14