University of Twente Student Theses

Login

Crop Type Identification And Yield Estimation Using Multi-task Learning CNN And Swin Architectures

Gamil, Mohamed (2024) Crop Type Identification And Yield Estimation Using Multi-task Learning CNN And Swin Architectures.

[img] PDF
3MB
Abstract:Early crop yield estimation (i.e., before the harvest season) is essential for effective commodity market management, ensuring food security and understanding various crop production trends at preliminary stages. Although several crops can be cultivated in a given area, most research focuses on single-crop yield estimation using Earth Observation (EO) data and Artificial Intelligence (AI) models. However, developing these crop-specific models is time-consuming and computationally expensive. A classified image of crop types helps to improve the accuracy of multi-crop yield estimation models because it guides the model to focus on the relevant images’ regions. However, such a layer is not always available because its creation is expensive and time-consuming. Furthermore, the creation of such a crop-type layer requires a substantial amount of labelled ground truth data that is often not available. To estimate multi-crop yields while addressing the unavailability of a classified crop type layer, we developed multi-task learning models for crop type identification and yield estimation. We hypothesized that multi-task learning models can learn both tasks concurrently. Additionally, these models solely require satellite imagery inputs. In this thesis, we estimated the yield of two crops: corn and soybean and the case study covered the top four states in the USA in corn and soybean production namely Indiana, Iowa, Illinois, and Minnesota. Consequently, we developed two base CNN models for multi-crop yield estimation. The first one only used sentinel-2 images with eight bands and the second one used a classified crop layer (known as CDL) as an additional band to demonstrate the role of the CDL in achieving higher accuracies. Regarding the multi-tasks approach, we developed two models: one based on U-net and the other utilizing the Swin transformer-based architecture. The two multi-task learning models showed promising results in multi-crop yield estimation. For instance, both models achieved yield estimation accuracies comparable to the CNN model that relies on CDL as input. Additionally, those models proved their applicability for multi-crop yield estimation while solving the lack of a CDL-like layer in many other countries. However, there is still room for further improvements to increase the models’ accuracies. These improvements relate to date, data acquisition and model architecture aspects. For instance, for date-related enhancements, incorporating temporal data in different years and different times in the mid-season of the crops could be beneficial. Regarding data acquisition, integrating different bands of sentinel-2, different sensors’ data, and additional parameters such as weather data, could improve the models’ generalizability. For model-related improvements, different loss functions and optimization techniques could improve performance. Finally, the primary contribution of this research lies in the development of multi-task learning models for crop type identification and yield estimation. These models proved to be actionable models that can be used directly to estimate multi-crop yields without the need for CDL while still achieving good results. This approach addressed the limitations in developing crop-specific models by reducing the number of parameters to be learned and the computation time and resources required.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Subject:38 earth sciences, 48 agricultural science, 54 computer science
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:https://purl.utwente.nl/essays/101732
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page