CoverType prediction

Predicting forest cover type from cartographic variables only.

Authors: Andrea Somma, Lorenzo Paggetta, Pietro Marelli

The present work showcases different methods to develop a classifier for the Cover Type dataset, in order to achieve an accurate and balanced model for the cover forest type from the cartographic variables in the dataset. The Cover Type dataset contains trees observation from four wilderness areas of the Roosevelt National forest in Colorado. The data is made of cartographic variables only, with no remotely sensed data. It is a rather large dataset, made of 7 forest cover types, more than half a million instances and 54 features, which include data such as elevation, aspect, slope, distance to hydrology, soil type and many others.

🏹 Targets

	Forest Cover Type
1	Spruce/Fir
2	Lodgepole Pine
3	Ponderosa Pine
4	Cottonwood/Willow
5	Aspen
6	Douglas-fir
7	Krummholz

📚 Features

Label Code	Label Type	Data Type
1	Elevation	Integer
2	Aspect	Integer
3	Slope	Integer
4	Horizontal Distance To Hydrology	Integer
5	Vertical Distance To Hydrology	Integer
6	Horizontal Distance To Roadways	Integer
7	Hillshade 9am	Integer
8	Hillshade Noon	Integer
9	Hillshade 3pm	Integer
10	Horizontal Distance To Fire Points	Integer
11-14	Wilderness Area	Binary
15-54	Soil Type	Binary

Data classes

🐎 Performance ML

Model	Accuracy [%]	Parameters	Size [MB]	Training Time
Bagging-based - Rescaled	97	3.9M	24	5 min
DecisionTree-based - Rescaled	92	6k	3	2 min
DecisionTree-based opt - Rescaled	90	3k	0.72	~20 seconds

🐎 Performance NN

Model	Accuracy [%]	Parameters	Size [kB]	Training Time
NN - Rescaled	93.3	233.9k	2850	9 min
NN opt - non-quantized - Rescaled	90.3	10.6k	172	4 min
NN opt - quantized - Rescaled	90	10.6k	19.5	4 min

🏛️ NN Acrhitecture

Convolutional Neural Network (NN)

Convolutional Neural Network Optimized (NN opt)