University of Twente Student Theses
Deep learning based multi-temporal crop mapping accounting for sample imbalance
Antony Ravi, Margarret Ashmita (2022) Deep learning based multi-temporal crop mapping accounting for sample imbalance.
PDF
4MB |
Abstract: | Accurate mapping of crop types is essential for solving problems of food security, crop inventory and to help farmers in making decisions regarding improving productions and managing agricultural practices. A variety of scientific methods starting from conventional supervised and unsupervised classifiers to machine and deep learning classifiers has been utilized to perform crop type classification from remote sensing images. In reality an agricultural field appears to have a lot of majority crops with a few scattered minority crops in a particular cropping season. Generally, crop classification studies ignore these minority crops or gather them together into a class called other crops and carry out classification. This study is aimed at providing a solution for the problem of class imbalance that occurs in deep learning based classification. Owing to the limited availability of labelled data, a 1D CNN sequential deep learning model is selected. The study area chosen is Jhansi District in the Bundelkhand Region of Uttar Pradesh in India. Sentinel 2 freely accessible optical data with a high spatial resolution of 10 m is chosen for the study. Taking into account the Rabi Season, 23 multi-temporal Sentinel 2 images covering Rabi crop season is considered. The majority crop class identified is wheat and the minority crop classes identified are mustard and chickpea. The Hyper-parameter of the model is optimized through Bayesian Optimization. Two approaches are considered in this study to address the problem of class imbalance. They are algorithm level balancing techniques and data level balancing techniques. Cost sensitive learning is introduced in the 1D CNN model by making changes to the loss function. Three different loss functions such as categorical cross entropy, focal loss and class weighted loss are employed to arrive at one best loss function that can solve imbalance problem in crop classification. Secondly, data level balancing methods such as under sampling, oversampling and hybrid methods (combination of under sampling and oversampling) are applied to the training dataset and the 1D CNN model is trained and tested with these balanced dataset. The results are assessed based on accuracy metrics such as Overall Accuracy, F1 –score, Precision, and Recall. G-Mean score is preferred to assess the accuracy of individual classes as it is proved to be utilized in imbalanced classification problem. It was observed that class weighted loss performed the best out of all loss functions with an overall accuracy of 70.37 % and a G-Mean score of 56.24%. When it comes to data level balancing techniques, under sampling outperformed oversampling as it lead to the generation new incorrect artificial samples. After carefully interpreting the results of both these techniques, and comparing those with sampling techniques algorithm level balancing with class weighted loss was chosen to be the best out of all the different methods incorporated to solve the imbalance in the classification problem. |
Item Type: | Essay (Master) |
Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
Programme: | Geoinformation Science and Earth Observation MSc (75014) |
Link to this item: | https://purl.utwente.nl/essays/106266 |
Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Repository Staff Only: item control page