[Job] [STAGE] Estimating unbiased planted areas for major crops, France (67)

Annonce postée par : Josef Wagner (jwagner(a)unistra.fr) ---------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------- Estimating unbiased planted areas for major commodity crops from remote sensing, without the use of ground truth validation data Keywords: NASA Harvest, Geostatistics, GIS, Agriculture, Planted area estimation, Remote Sensing Application submission deadline: February the 1st, 2024 or until position is filled Contact: Josef Wagner | wagnerj(a)umd.edu | jwagner(a)unistra.fr -------------------------------------------------------------------------------------------------------------------------------------------------------------- 1. Problem statement 1.1 Context Timely information about the expected major commodity crops production can reduce market uncertainties and thus prevent unexpected price spikes and volatility for the importing countries. As a consequence, in-season estimates of crop planted areas are extremely important for providing first seasonal insights on the upcoming agricultural production of a country. Planted area can be estimated from farmer surveys and mandatory reporting or from satellite derived crop type maps. The latter are prone to (i) classification errors characterized by confusion matrices and corresponding commission and omission error, and (ii) mixed pixel effects on class transitions (e.g. a pixel contains both crop and trees). As a result, planted areas cannot be directly computed from the map, by counting pixels belonging to a class and multiplying by pixel area. This estimator, called pixel counting, is biased because of aforementioned edge, omission and commission errors. It is, therefore, essential to provide unbiased estimates of planted areas, before reporting to authorities and stakeholders. For this, well-established statistical procedures that rely on a sample-based approach, where satellite-derived maps are used for stratification purposes, should be followed. Those procedures can be broken down into four steps, based on a crop type map: (i) estimating a required number of samples based on class precision and pixel counted areas; (ii) randomly selecting the required number of points within each map class (this step is called a random stratified sample); (iii) annotating each point either through ground visits or satellite image annotation; (iv) computing unbiased areas and standard errors per map class using a set a pre-defined formulas. Steps 1,2 and 4 are fully automated and do not require more than basic statistical knowledge. Ground campaigns for sample annotation are very time consuming and sometimes impossible due to conflict or budget issues. An alternative approach is to label the set of random points using satellite imagery and phenological knowledge. This approach restricts area computation to only “recognizable-from-satellite” crop types or crop groups. For instance, winter crops can be distinguished from summer crops because they green up in March/April, whereas summer crops green-up in May/June. In turn, rapeseed can be distinguished from winter cereals thanks to its yellow flowering signal. However, it is not possible to distinguish maize from sunflower or from soybean based on visual inspection of satellite imagery. 1.2 Problematic and approach outline In this project we would like to simulate the use case where data would be available for training a crop type mapping algorithm to predict detailed crop types, but no ground data campaign would be feasible. In this specific case we are interested in the following research question: - How to estimate unbiased planted areas for detailed crop types when ground data acquisition campaigns are not feasible? We propose the following approach: (i) Build a detailed crop type map comprising different crops such as maize, sunflower, soybean, wheat, barley and rapeseed. (ii) Estimate unbiased areas when possible (e.g. for crop groups such as summer crops, winter cereals and for crop types such as rapeseed). (iii) Remap crop types, forcing the pixel counts to match the unbiased estimated areas (Optional step). (iv) Break down crop group areas per crop type using the pixel counted proportion of each crop within the crop group. We would like to assess the effects of map accuracy and estimation scale (national, regional, sub-regional) on detailed crop type area estimates. This project aims to result in a scientific paper, with the chosen candidate listed as the second author. 1.3 Proposed method and project milestones In order to get Common Agriculture Policy subsidies, French farmers are forced to report their planted crop types and field boundaries every year. This data is opensourced in vector format with a one-year delay. Although it does not cover every single field in France, it can be used as a close to complete coverage ground truth layer. Instead of mapping crop types for France, we will rasterize the 2023 RPG and introduce a random but controlled amount of error into the map to simulation the effect of mapping inaccuracies. Unbiased areas will be estimated first for “recognizable from satellite” crop types and crop groups. Then area proportions will be used within crop groups to refine area estimates. Crop group and crop type areas will be compared to official French statistics at national, regional and sub-regional level. The project is divided into five milestones: Milestone 1: Literature review The review will be focused on establishing standard accuracies reported for different crop types in literature. This will help guide the range of accuracies to be tested. Milestone 2: Data acquisition RPG and official French cropped area statistics collection. Milestone 3: RPG – official statistic adequation test (i) Compare aggregated RPG vector areas per crop type to official planted area statistics statistics at national, regional and sub-regional level. (ii) Convert the RPG vector layer to raster format and repeat the area comparision exercise at national, regional and sub-regional level. Milestone 4: Artificial error introduction pipeline Artificially introduce predefined error levels into the RPG 2023 crop type map. Milestone 5: Evaluation of the error level and scale on detailed crop type planted area estimates For different error levels, run the unbiased area estimation procedure (pipeline is not yet automated) and estimate unbiased planted areas per crop group and detailed planted areas per crop type based on pixel counted proportions. 1.4 Resources - Oloffson (2014) Unbiased Area Estimation: https://www.sciencedirect.com/science/article/abs/pii/S0034425714000704 2 Internship specifications 2.1 Administrative details This internship opportunity spans a minimum duration of six months, hosted jointly by NASA Harvest and the University of Strasbourg, located at the ICube Laboratory, 300 Boulevard Sebastien Brant, 67400 Illkirch-Graffenstaden, France. Start and end dates are flexible, tailored to the candidate's availability for this full-time position requiring 35 hours per week, offering a salary of 4.35 euros per hour. 2.2 About NASA Harvest As our chosen candidate you will be integrating an international and multicultural team, hosting top tier remote sensing and agriculture scientists. NASA Harvest stands as an esteemed research consortium at the forefront of agricultural monitoring and food security analysis, pioneering advancements in satellite-based observations. Comprised of leading experts and institutions worldwide, NASA Harvest leverages cutting-edge technology to provide critical insights and data-driven solutions for sustainable agriculture and global food supply management. Through its collaborative efforts, NASA Harvest aims to address pressing challenges, enhance decision-making, and foster resilience within the agricultural landscape on a global scale. As a growing team, NASA Harvest is now also implanted in Strasbourg, France where this internship is hosted. 2.3 Your profile We're committed to considering all applications equally, irrespective of race, age, gender, or any other identifying factor. We're looking for a candidate proficient in English communication and writing, a collaborative team player with initiative and scientific curiosity. Essential skills include expertise in geospatial data management, vector and image data processing, Python coding, and a foundational understanding of basic statistics. Proficiency in Google Earth Engine would be valuable. Optional, yet beneficial, skills include knowledge in remote sensing, agronomy. 2.4 Application To apply for this internship, please submit a cover letter and a concise CV to wagnerj(a)umd.edu or jwagner(a)unistra.fr by February 1st, 2024. Please note that while the deadline is February 1st, the internship will remain open until the position is filled. 2.5 Questions We will be happy to answer any question regarding the internship you sent to either wagnerj(a)umd.edu or jwagner(a)unistra.fr ---------------------------------------- L'annonce est située https://georezo.net/forum/viewtopic.php?pid=364993#p364993 Pour vous désabonner connectez-vous sur le forum puis Profil / Abonnement -- Association GeoRezo - le portail géomatique https://georezo.net (Message généré automatiquement. Ne pas répondre à ce message, utiliser les coordonnées contenues dans l'annonce)