Spelling suggestions: "subject:"comprehensiveprobabilistic framework"" "subject:"semiprobabilistic framework""
1 |
A model for crop monitoring and yield prediction fusing remotely sensed data and prior information in a deterministic-probabilistic frameworkLovison-Golob, Lucia 31 January 2024 (has links)
This research focuses on the development of a deterministic-probabilistic framework for agricultural land use and management, specifically for both annual crops, such as wheat, barley and maize, and permanent crops, such as vineyards. The goal is to predict crop greening and peak crop development progressively through the growing season, based on accumulating information as the crop develops and matures, and to provide an accompanying uncertainty statement (credible interval) with each prediction. The integrated area underneath the phenology curve can be associated, although not explicitly in our example, with per-area crop yield. The prediction model relies on remotely sensed data, including science data products from the Landsat and MODIS (Moderate Resolution Imaging Spectroradiometer) spaceborne instruments, field data from agro-meteorological stations, and statistical data from prior years.
The development of the deterministic-probabilistic model focuses on northeastern Italy, a region of small agricultural plots set in a diverse physical landscape, which is typical of many areas of old-world and developing-nation agriculture. The estimation process uses the phenological cycle of the MODIS Enhanced Vegetation Index (EVI), extracted from the satellite imagery at 500 m spatial resolution. Landsat data, at 30-m spatial resolution, are fused with MODIS data, to provide fine-scale information better suited to small-field agriculture.
By applying a piecewise logistic function to model the time trajectory of EVI values, crop development and peak greenness are estimated and characterized based on the main phenological stages determined from the remote imagery trained with ground station observations. The deterministic-probabilistic model is later validated with observations from reference testing stations and statistical crop and yield data obtained independently by administrative districts such as regional and national organizations. A temporal filter of the main phenological stages, here called a crop calendar, plays a critical role. A Bayesian approach to integrate stochastically the parameters related to a certain area provides a way to include the different datasets at the different dimensions and scales and to assess the probability to obtain a vegetation index within a given uncertainty. The model becomes, therefore, a typical generalized linear model problem, deterministically described by a piecewise logistic function, with the parameters describing the peak phenological curve estimated probabilistically, with their own uncertainty. / 2026-01-31T00:00:00Z
|
Page generated in 0.0873 seconds