Welcome back to my research journey at the KMUTT Geospatial Engineering and Innovation Center (GEIC) in Bangkok. This past week (September 14th to 18th) has been an absolute rollercoaster of learning, adapting, and problem-solving. From presenting my initial progress to entirely rebuilding a deep learning model, here is a look at what happens behind the scenes of mapping Above Ground Biomass (AGB) in Thailand.
The Transition and The Language Barrier
Coming from a Network and Security background, shifting my focus to Machine Learning and Deep Learning for remote sensing is a significant leap. Early in the week, I had to take a step back and thoroughly revise deep learning fundamentals to ensure I fully grasped their application in satellite analysis.
I presented my initial workflow for mapping mangrove AGB across several Thai provinces to Ajarn Parivate, Pi Nalin, and the GIS class. While the progress was swift, I encountered a unique, non-technical hurdle while exploring QGIS and clipping shapefiles: the language barrier. Navigating map layers written entirely in Thai was quite the challenge! It pushed me to actively learn the Thai language on the side, which has been incredibly helpful for both my research comprehension and daily life.
In research, you rarely get perfect data. Mid-week, Pi Boy provided the ground truth data, but it only covered macro or average areas rather than precise granular metrics. Instead of waiting for better data, I realized that adaptation was the only way forward.
While deep diving into Recurrent Neural Networks (RNN) and LSTM algorithms to decide on the best architectural approach, I formulated a plan. I spread the average mangrove data into 500 distinct points to manually craft a usable dataset. Once the dataset was fixed, I integrated Sentinel 1 and Sentinel 2 satellite imagery to prepare the model for training. Alongside this, I had to heavily optimize my code to ensure my hardware could handle the training process efficiently.
The 19% Heartbreak and The 70% Breakthrough
By Friday, I was ready to validate my trained model. However, upon reviewing the results with Pi Boy, we realized that the anomalies in my crafted dataset had severely impacted the model's performance. The accuracy sat at a disappointing 19%.
Instead of getting discouraged, we sat down for an in-depth discussion to figure out how to make this limited data truly usable. We eventually found a clear path forward. I immediately went back to the code, executed our revised plan, and re-trained the data. The result? The model’s accuracy jumped to over 70%!
This week was a powerful reminder that in research, things rarely go perfectly on the first try. I learned that having a high-end setup or perfect data isn't everything. True progress comes from code efficiency, the ability to adapt to current conditions, and raw curiosity. I am incredibly excited to continue refining this model and enhancing my AGB mapping workflow in the coming weeks. Stay tuned!
Comments
Post a Comment